Content uploaded by Canyu Chen
Author content
All content in this area was uploaded by Canyu Chen on Jul 02, 2024
Content may be subject to copyright.
Evaluating the Social Impact of Generative AI Systems
in Systems and Society
Irene Solaiman1∗Zeerak Talat2∗William Agnew3Lama Ahmad4
Dylan Baker5Su Lin Blodgett6Canyu Chen7Hal Daumé III8
Jesse Dodge9Isabella Duan10 Ellie Evans11 Felix Friedrich12,13
Avijit Ghosh1Usman Gohar14 Sara Hooker15 Yacine Jernite1
Ria Kalluri16 Alberto Lusoli17 Alina Leidinger18 Michelle Lin19,20
Xiuzhu Lin11 Sasha Luccioni1Jennifer Mickel20 Margaret Mitchell1
Jessica Newman21 Anaelia Ovalle22 Marie-Therese Png23 Shubham Singh24
Andrew Strait25 Lukas Struppek12,26 Arjun Subramonian22
1Hugging Face, 2Mohamed Bin Zayed University of Artificial Intelligence, 3Carnegie Mellon
University,
4
OpenAI,
5
DAIR,
6
Microsoft Research,
7
Illinois Institute of Technology,
8
University of
Maryland, 9Allen Institute for AI, 10University of Chicago, 11Independent Researcher, 12TU
Darmstadt, 13hessian.AI, 14 Iowa State University, 15Cohere for AI, 16Stanford University, 17Simon
Fraser University,
18
University of Amsterdam,
19
Mila - Quebec AI Institute,
20
University of Texas at
Austin, 21University of California, Berkeley, 22 University of California, Los Angeles, 23Oxford
University, 24University of Illinois Chicago, 25Ada Lovelace Institute, 26 DFKI
Abstract
Generative AI systems across modalities, ranging from text (including code), im-
age, audio, and video, have broad social impacts, but there is no official standard
for means of evaluating those impacts or for which impacts should be evaluated. In
this paper, we present a guide that moves toward a standard approach in evaluat-
ing a base generative AI system for any modality in two overarching categories:
what can be evaluated in a base system independent of context and what can be
evaluated in a societal context. Importantly, this refers to base systems that have no
predetermined application or deployment context, including a model itself, as well
as system components, such as training data. Our framework for a base system
defines seven categories of social impact: bias, stereotypes, and representational
harms; cultural values and sensitive content; disparate performance; privacy and
data protection; financial costs; environmental costs; and data and content mod-
eration labor costs. Suggested methods for evaluation apply to listed generative
modalities and analyses of the limitations of existing evaluations serve as a starting
point for necessary investment in future evaluations. We offer five overarching
categories for what can be evaluated in a broader societal context, each with its
own subcategories: trustworthiness and autonomy; inequality, marginalization,
and violence; concentration of authority; labor and creativity; and ecosystem and
environment. Each subcategory includes recommendations for mitigating harm.
∗
Both authors contributed equally. The following author’s order is alphabetical by last name. Contact
information: irene@huggingface.co and z@zeerak.org. This chapter will appear in Hacker, Engel, Hammer,
Mittelstadt (eds), Oxford Handbook on the Foundations and Regulation of Generative AI. Oxford University
Press.
Preprint. Under review.
arXiv:2306.05949v4 [cs.CY] 28 Jun 2024
Introduction
Understanding an AI system’s social impacts from conception to training to deployment requires
insight into aspects such as training data, the model itself, material infrastructure, and the context
in which the system is developed and deployed. It also requires understanding people, society, and
how societal processes, institutions, and power are changed and shifted by the AI system. Generative
AI systems are machine learning models trained to generate content, often across modalities, and
have been widely adopted for diverse downstream tasks. For generative AI systems, such as language
models, social impact evaluations are increasingly normalized. The Conference and Workshop on
Neural Information Processing Systems (NeurIPS) establishing a Broader Impacts section has shifted
norms for including social impact considerations in AI publications while raising challenges [
18
], but
there exists no broadly applied standard. We propose a framework for social impact evaluations of
generative AI systems across modalities. We address this work to three groups of readers: researchers
and developers; third-party auditors and red-teamers; and policymakers who evaluate and address the
social impact of systems through technical and regulatory means.
We define social impact as the effect of a system on people and society along any timeline with a
focus on active, measurable, harmful impacts. This document is concerned with impacts that have
already been documented or directly follow from current and emerging methods.
2
Since social impact
evaluation covers many overlapping topics, we propose a technical framework of the aspects of a
system that can be evaluated along its lifecycle.
We focus on generative models across five modalities: text (including language and code), image,
video, audio, and multimodal combinations of aforementioned modalities. The given categories
and evaluation methods are based on popularly deployed evaluations in use today and do not
exhaustively cover all methods. Social impact evaluations offered in our framework are key to,
but differ from, harm mitigation and alignment methods; evaluations aim to improve understanding
of social impact and inform appropriate uses in different contexts, which is a critical precursor to
taking action. The goal of understanding systems requires quantitative and qualitative evaluations
and should seek to capture nuances in complex social topics. While evaluations can serve regulation
and risk mitigation needs, they may be reductive and miss nuances that are critical to attaining a
holistic understanding of the impact of AI systems, especially of those at the margins [
179
]. While
the potential for downstream harm depends on deployment context or risk evaluation gaps [
163
,
208
],
system-level evaluations are still beneficial, e.g., to find patterns that are inadmissible in any context.
Moreover, harmful impacts reflected in generative AI systems are rarely limited to the system itself.
Long-term societal inequity, power imbalances, and systemic injustices [
378
] are all reflected in
the training data, influence system development and deployment [
339
], and shape social impact
[
176
,
261
]. While technical evaluations can probe and isolate aspects of social impact in a specific
system, holistic evaluation and mitigation encompasses human and infrastructural social harms.
As we highlight in each section, the existing social impact evaluation landscape requires more
investment. The evaluations that have been developed, especially for aspects of models that may
be tied to their more negative social impacts, can overfit to certain lenses and geographies, such
as evaluating a multilingual system only in the English language. Often, developers and deployers
will rely on evaluations built within the same company (e.g., OPT-175B [
386
] from Meta’s safety
evaluations). While we underscore the need for formal social impact evaluation, evaluations cannot
justify the rights-violating applications of generative AI. Since there is currently no consensus or
governing body to determine what constitutes the social impacts of any AI system nor how to evaluate
them, our work aims to make the social impact evaluation landscape more accessible.
2
Downstream harms for when generative systems are embedded in physical systems such as robotics,
autonomous vehicles, and drone warfare are out of the scope of this paper.
2
Background
The social impact aspects of an AI system are largely dependent on context, from the deployment
sector to the use case. Generative AI systems include, but are not limited to, large language models
(LLMs) (BLOOM [
38
], GPT-3 [
55
], LLaMA [
354
]), text to image models (ImaGen [
300
], DALL-E
[
283
], Stable Diffusion [
296
]), and increasingly multimodal models [
69
] (GPT-4 [
258
], Claude 3
[
17
], Gemini [
343
]), which can combine unimodal harms and amplify them in novel ways [
273
,
344
].
Generative AI systems are sometimes referred to as a type of General-Purpose AI Systems: systems
capable of a wide range of tasks that may be applicable across sectors and use cases. These systems
are mostly examined for generalization properties and societal impact [
46
], but their social impact
evaluations are sparse, could benefit from increased standardization, and do not provide adequate
coverage across risks [
117
]. Although there are more common evaluations for performance and
accuracy (e.g., GLUE) [
367
], many of these are saturated, and a select few can fail to assess important
capabilities [
187
,
282
,
334
]. Social impact as a complex qualitative concept is even more difficult to
evaluate.
At the same time, proposed AI regulations across numerous jurisdictions include or mention evaluating
the impact of an AI system. There are more than 1,000 AI policy initiatives from dozens of countries
around the world [
253
]. However, regulatory bodies that have announced plans and guidelines still
skew toward Western and East Asian governments: European Union [
107
], United States of America
[
291
], Canada [
158
], United Kingdom [
357
], South Korea [
292
], Japan [
347
], and China [
93
]. While
many of these proposed requirements only apply to systems that fall into “high risk” categories as
defined by the proposed EU regulation on generative AI[
107
], generative AI systems are largely still
being scoped.
1 Methodology
First, we convened thirty experts across industry, academia, civil society, and government to contribute
to a two-part workshop series. The first workshop created the underlying framework for defining and
categorizing social impacts that can be evaluated. The second workshop examined the feasibility of
evaluating categories, including past approaches to evaluations and metrics, limitations, and future
directions for improvements. For the first workshop, we asked experts to discuss the possible impacts
of systems for each of the five modalities of generative systems. For the second workshop, we created
meta-categories of impacts and collected existing methods for evaluation within these categories. The
findings from the discussions inform our framework and evaluation method sections. Both workshops
were conducted under modified Chatham House Rules, where contributors could opt into authorship.
Another workshop in the form of a CRAFT session at ACM FAccT 2023 invited 30 more researchers
to build on the framework, particularly examining the landscape of existing evaluations per modality
in each category. Over one year, 30 researchers conducted literature reviews in each impact category,
collected existing evaluations, and shared analyses to distill modality-specific overviews reflected in
this paper.
2 Related Work
Toolkits and repositories for evaluating qualitative aspects of AI systems are broad and constantly
evolving. Many are aimed at public agency procurement and deployment. In 2018, AI Now released
its framework for algorithmic impact assessments focused on public agencies [
289
]. Many public
interest organizations and government initiatives have since published frameworks and assessment
tools, such as the OECD’s Classification Framework for AI risks [
254
] and Canada’s Algorithmic
Impact Assessment Tool [
355
]. The U.S. National Institute of Standards and Technology (NIST)
Artificial Intelligence Risk Management Framework (AI RMF) is also intended to be applicable to all
AI systems, although specific applications to generative AI systems [29] are in progress [245].
Evaluation suites across system characteristics for specific generative system modalities, such as
language, include Holistic Evaluation of Language Models (HELM) [
48
], Dynabench [
187
], ML-
Commons AI Safety Benchmark [
364
], BigBench [
328
], and Language Model Evaluation Harness
[
124
] have been proposed. These evaluation suites incorporate capabilities evaluations as well as
3
evaluations across the categories in this paper and are similarly living resources. Researchers from
Google DeepMind developed a sociotechnical evaluation framework that looks at generative AI
system capability across modalities, including human interaction and systemic impacts as further
elements to evaluate [373].
Technical evaluation suites are often specific to a type of system. Auditing frameworks [
281
] have
also been presented and can be powerful tools. An increasing body of work taxonomizes dangers
[
33
], social impacts [
159
], sociotechnical harms [
313
], and social risks, of all [
116
] or specific
types of generative AI systems like language models [
371
]. Evaluating these risks and impacts is a
complementary ongoing research area.
Categories of Social Impact
We divide impacts into two categories for evaluation: what can be evaluated in a technical system and
its components and what can be evaluated among people and society. While the high-level categories
overlap, this framework highlights opportunities and gaps in existing evaluations between technical
systems and their context of use.
This first section includes evaluations for base systems, which refer to AI systems, including models
and components, that have no predetermined application. The latter section examines systems in
context and includes recommendations for mitigating harmful impacts. Aspects of what can be
evaluated in the can inform categories in People and Society. As shown in Figure 1, each Technical
Base System category connects with at least one People and Society category.
Figure 1: Evaluation Categories and Connections
4
3 Impacts: The Technical Base System
Below we list the aspects possible to evaluate in a generative system. Context-absent evaluations only
provide narrow insights into the described aspects of the type of generative AI system. The depth of
literature and research on evaluations differ by modality, with some modalities having sparse or no
relevant literature, but the themes for evaluations can be applied to most systems.
The following categories are high-level, non-exhaustive, and present a synthesis of the findings across
different modalities. They refer solely to what can be evaluated in a base technical system:
• Bias, Stereotypes, and Representational Harms
• Cultural Values and Sensitive Content
• Disparate Performance
• Environmental Costs and Carbon Emissions
• Privacy and Data Protection
• Financial Costs
• Data and Content Moderation Labor
3.1 Bias, Stereotypes, and Representational Harms
Contributors: Yacine Jernite, Lama Ahmad, Sara Hooker, Irene Solaiman, Zeerak Talat, Margaret Mitchell,
Usman Gohar, Jennifer A Mickel, Dylan Baker, Alina Leidinger, Felix Friedrich, Anaelia Ovalle, Avijit Ghosh,
Hal Daumé III
Generative AI systems can embed and amplify harmful biases. Different types of bias, from system
to human to statistical, interact with each other [
308
] and are intertwined. Evaluations of bias are
often referred to as evaluations of “fairness.”
Understanding biases in the final system requires interrogating the full development chain, from
project statement and data collation and curation to training, adaptation, and deployment choices
[
336
,
339
]. Scaling systems for improved model performance have been shown to encode harmful
biases [
42
]. Training datasets also encode specific worldviews that can exacerbate social biases [
33
].
The overall level of harm is furthermore impacted by modeling choice [
156
]. These can include
choices about many stages of the optimization process [
134
,
192
]; privacy constraints [
26
], widely-
used compression techniques [
4
,
157
,
255
], inferring demographic attributes from proxies [
129
] and
the choice of hardware [
387
], which have all been found to amplify social biases, and thereby harms,
towards people with marginalized identities [
34
]. The geographic location, demographic makeup,
and team structures of researcher and developer organizations can all introduce biases. Moreover, the
ways in which people’s data are measured and aggregated reinforce harmful categorization biases.
3.1.1 What to Evaluate
While the degree of harm depends on many factors, from the type of output to the cultural context of
training and deployment, focus on bias evaluations has centered on protected classes as defined by
the United States [
111
] and United Nations [
356
] guidelines. These frameworks are non-exhaustive,
and harm exists outside and at the intersection of categories. These limitations may be addressed by
adding categories or measuring their intersections.
Popular evaluations focus on harmful associations [
62
] and stereotypes [
206
,
241
,
243
], including
methods for calculating correlations and co-occurrences as well as sentiment [
88
] and toxicity
analyses.
Across modalities, biases can be evaluated using intrinsic and extrinsic methods [
136
,
137
], where
the former seeks to evaluate biases within model weights and the latter evaluates the manifestation of
biases in the outputs of downstream tasks (e.g., captioning). Evaluations can also be specific to a
certain function of a modality, such as question-answering in language [264].
In text, at the dataset level, approaches include embedding similarities [
23
,
62
], topic modeling
[90, 96, 284], entropy-based calculations [7, 217, 314], and measurements based on co-occurrences
5
[
290
] to capture how discrete items such as tokens and words cluster. At the output level and across
languages, biases can be represented differently [
219
], occurring at the word [
62
], sentence [
225
], or
document [74] level.
In image, approaches include image comparisons and utilizing tools such as captioning systems [
78
].
The synthetic nature presents an added complexity when grounding in social categories. An important
aspect when evaluating text-to-image models is the language used to generate images. Several works
have investigated the impact of using different languages [
122
] and scripts [
332
]. Different levels of
evaluation help investigate bias amplification [
121
]. At the dataset level, works have measured hate
[43]. The output level can examine sets of generated outputs [215].
3.1.2 Limitations
Due to the contextual and evolving nature of bias [
119
], evaluation cannot be fully standardized and
static [
168
]. Protected class categorization itself cannot be exhaustive, varies across cultural contexts
[
37
,
280
], and can be inherently harmful [
280
]. Certain protected classes, such as race and gender,
are often more represented in publications and publication venues around biases of (generative)
systems [
137
]. Many evaluations focus on distinct or binary groups [
94
], due to the complexity of
operationalizing intersectionality;
3
and in many cases, assumptions used to simplify for the sake
of mathematical notation and interpretation result in obscuring the very phenomena they seek to
describe [
86
]. In many cases, this simplification itself adversely harms the group in consideration
[50].
Legal considerations around collecting certain protected attributes can lead to selection bias in
annotations.
4
Moreover, geographic and cultural contexts shift the meaning of different categories.
5
Annotators often have different perceptions of concepts like race, can racialize groups differently
[
185
], or are influenced by their own lived experience [
370
] when selecting protected categories
[276].
Evaluations for stereotype detection can raise false positives and can flag relatively neutral associations
based on facts (e.g., population x has a high proportion of lactose-intolerant people). Additional
tooling used to aid in identifying biases, e.g., image captioning, can introduce its own biases. Tools
broadly risk miscategorization and misrepresentation.
3.2 Cultural Values and Sensitive Content
Contributors: Irene Solaiman, Zeerak Talat, Alina Leidinger, Isabella Duan, Margaret Mitchell, Felix Friedrich,
Jennifer Mickel
Cultural values are specific to groups, and sensitive content is normative. Sensitive topics also vary
by culture and can include hate speech [
234
]. What is considered a sensitive topic, such as egregious
violence or adult sexual content, can vary widely by viewpoint. Due to norms differing by culture,
region, and language, there is no standard for what constitutes sensitive content.
Distinct cultural values present a challenge for deploying models into a global sphere, as what may
be appropriate in one culture may be unsafe in others [
340
]. Generative AI systems cannot be neutral
or objective, nor can they encompass truly universal values. There is no “view from nowhere”; in
quantifying anything, a particular frame of reference [302] is imposed [339].
3.2.1 Hate, Toxicity, and Targeted Violence
Beyond hate speech and toxic language, generations may also produce invasive bodily commentary,
rejections of identity [
260
], violent or non-consensual intimate imagery or audio [
80
], and physically
threatening language, i.e., threats to the lives and safety of individuals or groups of people. This can
inflict harm upon viewers who are targeted, help normalize harmful content, contribute to online
radicalization, and aid in the production of harmful content for distribution (e.g., misinformation and
non-consensual imagery).
3See [133, 196, 261, 368] for further discussion on intersectionality and machine learning.
4See for instance [15, 362].
5See [162, 228, 301].
6
3.2.2 What to Evaluate
Cultural values are used as an umbrella term to encompass a variety of topics ranging from social
values to political ideology to humor. Many existing evaluations build on pre-existing scholarship,
such as the World Values Survey [
145
] and Geert Hofstede’s work on cultural values [
153
,
154
].
Increasingly, more evaluations are inductive and participatory, grounded in a specific cultural context
[
95
,
278
]. A non-exhaustive categorical framework and human-reviewed evaluations [
322
] can
capture some aspects of culture, yet others may be missed, and the choice of cultural value grounding
in previous scholarship can affect the perspectives represented in cultural evaluations.
In text, approaches include ethical scenarios [
319
], U.S. political value representation [
171
], and
geopolitical statements [
202
]. Examining hate includes approaches characterizing harmful text [
286
],
toxicity[270], hurtfulness [252], or offensiveness [98].
In image, approaches include common object representation by geography [
126
,
210
], biases in
regional representation of locations, occupations, and other attributes[
242
,
307
], and cross-cultural
offensiveness [
214
]. Security evaluations examine hidden functionalities that can trigger harmful
content generation [333].
3.2.3 Limitations
Cultural values encompass an infinite list of topics that contribute to a cultural viewpoint. Human-led
evaluations [
260
] engaging with hateful and sensitive content can have a high psychological cost.
The types and intensity of sensitive content produced across modalities may vary. For evaluations
that rely on a third-party API, such as the many evaluations that leverage Google Perspective API for
toxicity detection, it is important to use the same version of the tool to avoid reproducibility issues
[
274
]. Toxicity-scoring tools suffer from their own biases, including the over-flagging of identity
terms as toxic [303] and the under-flagging of coded expressions [226].
The majority of existing literature equates nationality with cultural context, blending together poten-
tially culturally diverse regions; culture does not necessarily align along country boundaries [
342
].
This can lead to an inadequate representation of cultural values, which prioritizes the dominant
cultural values of a country or cultural values related to the group in power rather than representing
the differing cultural values people can have within a country [
278
]. These differences can be further
amplified by different languages. Often, cultural stereotypes are tightly bound to the language(s)
close to this culture. Evaluations across languages are important but challenging.
Furthermore, the scholarship and frameworks upon which cultural value evaluations are built may
reflect the regional and cultural values of those who contributed. Without adequate representation of
people with differing cultural values, evaluations can narrow to a subset of cultural values, potentially
missing the values of marginalized communities.
3.3 Disparate Performance
Contributors: Yacine Jernite, Irene Solaiman, Usman Gohar, Jennifer Mickel, Margaret Mitchell, Arjun
Subramonian, Anaelia Ovalle, Avijit Ghosh, Hal Daumé III
Disparate performance refers to AI systems that perform differently for different subpopulations,
leading to unequal outcomes for those groups. A model may perform unequally across subpopulations
for varying reasons, such as dataset skew with fewer examples from some subpopulations and feature
inconsistencies where some features are more predictive or easier to detect for some subpopulations.
Disparate performance can be due to systemic issues in data collection, due to dataset disparities,
and exacerbated by modeling choices. Colloquially, this category is often referred to as a bias but is
distinct from the representational biases discussed in Section 3.1.
Data availability differs due to geographic biases in data collection [
311
], disparate digitization
of content globally, varying levels of internet access for digitizing content, content filters [
99
],
and infrastructure created to support some languages or accents over others, among other reasons.
Overrepresentation and underrepresentation can suffer from a positive feedback loop if generative
models are trained on model-generated or synthetic data[
341
,
381
]. Interventions to mitigate harms
caused by generative AI systems may also introduce and exhibit disparate performance issues [91].
7
3.3.1 What to Evaluate
Across modalities, decisions made about training data, including filtering and reward modeling, will
impact how the model performs for different groups or categories of concepts associated with groups.
Evaluating model outputs across subpopulation languages, accents, and similar topics using the same
evaluation criteria as the highest-performing language or accent can illustrate areas where there
is disparate performance. One way to capture this is non-aggregated (disaggregated) evaluation
results with in-depth breakdowns across subpopulations. Existing common metrics include subgroup
accuracy, calibration, AUC, recall, precision, min-max ratios, worst-case subgroup performance, and
expected effort to improve the model decision from unfavorable to favorable [
133
]. Finally, coverage
metrics can also be used to ensure that a wide representation of subgroups have been identified.
In text, approaches include cross-lingual prompting on standard benchmarks can give insight to
performance of monolingual and multilingual language models [
160
], examining dialects [
48
],
analyzing hallucination disparity [
170
], and conducting multilingual knowledge retrieval evaluations
[
310
]. In data, studies find that retaining duplicate examples in a training dataset biases a model in
favor of generating such phrases [199].
In image, approaches include examining generation quality across concepts [
300
], accuracy of cultural
representation [214], and realism across concepts [330].
3.3.2 Limitations
Similar limitations that lead to disparate system performance contribute to disparate attention to
evaluations for different groups. Performance evaluations for similar tasks in non-English languages
will vary by the amount of resourcing for a given language. More spoken and digitized languages
may have more evaluations than lower-resource languages [
174
]. Another critical limitation is the
exponential number of subgroups and intersectionality, which becomes infeasible [
133
]. On the other
hand, smaller subgroups (including languages, cultures, race, etc.) suffer from data sparsity, which
will lead to uncertainty and less accurate evaluations [
368
]. Additionally, while several scholarly
resources propose hallucination mitigation procedures, limitations exist in measuring disparities with
respect to hallucinated content.
Furthermore, evaluations are bounded by the conceptualizations of performance and disparities them-
selves. Model evaluators should especially interrogate the extent to which notions and measurements
of performance capture the needs of the people affected.
3.4 Environmental Costs and Carbon Emissions
Contributors: Sasha Luccioni, Marie-Therese Png, Irene Solaiman, Usman Gohar, Michelle Lin
The compute power used in training, testing, and deploying generative AI systems, especially
large-scale systems, uses substantial energy resources and emits greenhouse gasses [
331
]. Overall,
information about emissions is scarce, and emission reporting should consider supply chains, manu-
facturing and hardware, and many indirect variables [
144
]. There is no consensus on what constitutes
the total environmental or carbon footprint of AI systems.
3.4.1 What to Evaluate
Existing efforts have pursued two main directions: the creation of tools to evaluate these impacts and
empirical studies of one or several models. The same tool can be used for multiple modalities, as
seen with CodeCarbon [
83
] and Carbontracker [
16
], measuring the carbon footprint for training and
inference carbon in audio generative models [101].
Existing metrics for reporting range from energy, compute, and runtime to carbon emissions. Ad-
ditional metrics include CPU, GPU, and TPU-related information such as hardware information,
FLOPS (Floating Point Operations), package power draw, GPU performance state, and CPU fre-
quency, as well as memory usage. Approaches include a web-based and programmatic approach for
quantifying models’ carbon emissions [
49
,
195
], evaluating power consumption [
83
], an experiment-
impact-tracker for energy and carbon usage reporting research [
149
], conversion based on power
consumed in the U.S. [
331
], and examining environmental impact across compute-related impacts,
8
immediate impacts of applying ML, and system-level impacts [
175
]. A holistic approach proposes a
Life Cycle Assessment (LCA) [36].
3.4.2 Limitations
Uncertainty around what variables to measure and lack of standardization complicates this category,
including marginal costs such as studying relative contribution of added parameters to a model
to their energy consumption and carbon footprint, as well as the proportion of energy used for
pre-training, inference [
218
], and fine-tuning ML models for different tasks and architectures [
380
].
There is also a need for added transparency from equipment manufacturers and data/hosting centers
to aid in accurately estimating GPU footprints and hosting-side impacts. Holistic approaches should
consider the effects of vertical integration and market concentration on the availability and adoption
of energy-efficient technologies [365, 384].
3.5 Privacy and Data Protection
Contributors: Lukas Struppek, Ellie Evans, Yacine Jernite, Irene Solaiman, Canyu Chen, Jessica Newman,
Shubham Singh, Isabella Duan, Arjun Subramonian
Examining the ways in which generative AI systems developers and providers leverage user data
is critical to evaluating its impact. Protecting personal information and personal and group privacy
depends largely on training data, training methods, and security measures. Intellectual property and
privacy concerns arise with generative models generating copyrighted content and highly sensitive
documents or personally identifiable information (PII), such as phone numbers, addresses and private
medical records. Privacy is additionally a matter of contextual integrity; generative models should
ensure that individuals’ data cannot be obtained from contexts in which they do not expect it to appear.
Critical practices for privacy protection throughout the AI lifecycle include data minimization, opt-in
data collection, and ensuring dataset transparency and accountability [
189
]. Some classical notions
of privacy protection, like data sanitization and differential privacy, can be difficult to translate to the
generative paradigm [54].
We suggest that providers should seek active consent and respect the explicit choices of individuals
for collecting, processing, and sharing data with external parties, as sensitive data could be inevitably
leveraged for downstream harm such as security breaches, privacy violations, and adversarial attacks.
Oftentimes, this might require retroactively retraining a generative AI system, in accordance with
policy such as the California Consumer Privacy Act (CCPA) [
61
]. In third-party hosted systems,
deployed language models can leak private input data that is hidden from users in its generations.
Companies often specialize LLMs by prepending system prompts to user inputs; these system prompts
may include proprietary company information or samples for in-context learning that contain PII or
sensitive database records. Consequently, LLMs may reveal information about system prompts in
their generations, violating privacy [277].
3.5.1 What to Evaluate
Evaluations can preserve privacy rights via authorizing access to evaluators. Some evaluations operate
as a proxy for a system’s ability to generate copyrighted or licensed content found within pre-training
data [
48
]. Memorization of training examples remains a critical security and privacy problem [
66
],
where models reveal parts or complete samples during the inference phase. Some hypothesize a
trade-off between fitting the long tails of data distribution and unintentional memorization of outliers
[114].
In text, the main approaches examine memorization, data leakage, and inferring personal attributes.
Research in measuring memorization can examine the maximum amount of discoverable information
given training data or the amount of extractable information [
244
] without training data access. Re-
search has examined unintended memorization when the underlying model reveals out-of-distribution
data [
65
]. Further analysis examines qualities (parameter count, sample repetitions in training data,
and context window) that increase the likelihood of memorization [
67
]. Despite mitigation techniques,
e.g., fine-tuning approaches [
166
] and data deduplication [
199
], discovering memorization is a hard
problem. Direct prompting over time to reveal PII can show varying success as models update [
204
].
Data subjects can use tools such as ProPILE to audit if their PII is likely to be revealed when given
9
enough prompts [
188
]. A classic technique to evaluate data leakage in machine learning models
called Membership Inference Attack (MIA) may not have as high performance for language models
[
103
]. Further work studies the divergence between model and human judgments on inference-time
privacy violations using multi-tiered evaluations based on Contextual Integrity and Theory of Mind
[231], showing the need to understand privacy context and purpose.
In image, approaches focus on training data memorization [
66
]. Methods to estimate severity include
adversarial MIAs and experiments to identify the proportion of images generated at the inference time
[
142
] with high similarity to training data. Research also detects memorized prompts by exploiting
the magnitude of noise prediction based on the text conditioning [375].
In audio, existing fraud detection methods [
299
] may be repurposed to scrutinize how well the
state-of-the-art audio generation models can synthesize a particular individual’s audio and trick the
detectors.
3.5.2 Limitations
Generative AI systems are harder to evaluate without clear documentation, systems of opt-out, and
appropriate technical and process controls to secure user data that can threaten the privacy and
security of individuals. Robust evaluations will often go beyond evaluating artifacts in isolation. The
immense size of training datasets [
169
] makes scrutiny increasingly difficult. Rules for leveraging
end-user data for training purposes are unclear, where user prompts, geolocation data, and similar
data can be used to improve a system. Private information may not be privacy-violating, and there’s a
need for more fundamental solutions to address the privacy problems at the design time rather than
ad-hoc safeguards that are patched to the models [231].
Moreover, generative AI requires contextual, community-centered definitions of privacy, which need
to be developed with participation from marginalized groups [
178
,
259
]. Research examining memo-
rization often relies on ground truth datasets for validity, which may not be accessible. Recognizing
PII content may require access to deeper private information for verification. Incentives for model
performance can sometimes be at odds with privacy; the more accurately an LLM can reproduce its
training dataset, the more likely it is to leak private information. Evaluations for non-generative AI
systems do not all translate well to generative systems; newer, better, and more specific evaluations
are needed to evaluate data leakage and identify privacy harms in the context of generative systems.
3.6 Financial Costs
Contributors: Irene Solaiman, Anaelia Ovalle
The estimated financial costs of training, testing, and deploying generative AI systems can restrict the
groups of people able to afford developing and interacting with these systems. Concretely, sourcing
training data, compute infrastructure for training and testing systems, and labor hours contribute to
the overall financial costs. These metrics are not standardized for any system but can be estimated for
a specific category, such as the cost to train and host a model.
3.6.1 What to Evaluate
Researchers and developers can estimate infrastructure, hardware costs, and hours of labor from
researchers, developers, and crowd workers. Popular existing estimates focus on compute costs using
low-cost or standard pricing per instance-hour [
203
]. Research lowering training costs also shows
tracking compute cost by day as the model trains and scales [
363
]. Frameworks break down the cost
per system component: data cost, compute cost, and technical architecture of the system itself [
248
].
Other variables used to calculate cost include the size of the dataset, model size, and training volume
[312].
In text, approaches examine costs for storing the data used for model training, compute hardware for
training, and hosting/inference. For storage, pricing should be considered for both the dataset and
resulting model, which vary depending on whether storage hardware is in-house or in cloud services
and what model architecture is used. Typical services are a combination of memory and tier-based.
In training, costs vary depending on whether the model is trained with in-house GPUs or is done
10
on per-hour-priced instances. Cost tradeoffs often consider model and dataset size. For hosting and
inference, if using cloud services, cost options include low-latency serving.6
In image and video, costs can vary depending on how an image is trained, and costs can vary by pixel
density and frames used for inference. Cost also depends on the model architecture.
In audio, costs depend on preprocessing; spectrograms are often used to chop up the audio signal into
smaller segments of time for later training and inference.
For all API-accessible models, inference costs largely depend on the service provided. For example,
inference costs are typically assessed by token-usage, with some variation in factors such as initial
prompt length, requested token response length, and model version.
7
However, not all languages cost
the same; what constitutes a token is largely dependent on how the model’s tokenizer was trained.
Exposure to languages that models have not been trained on can result in a larger number of tokens at
inference time, and prove more costly [
5
]. Inference volume considerations require optimizing for
decreased latency and robust delivery to meet demand. Further considerations include hosting the
model on a low-latency platform and monitoring demand.
3.6.2 Limitations
Only accounting for compute cost overlooks the many variables that contribute to a system’s training.
Costs in pre- and post-deployment, depending on how a system is released [
321
], are also difficult
to track as cost variables may not be directly tied to a system alone. Human labor and hidden costs
similarly may be indirect. Finally, it is necessary to keep track of the changes of costs and economy
of components over time.
3.7 Data and Content Moderation Labor
Contributors: Dylan Baker, Yacine Jernite Alberto Lusoli, Irene Solaiman, Jennifer Mickel, Arjun Subramonian,
Zeerak Talat
Human labor is typically conducted via a process called crowd computation, where distributed work-
ers, also called crowdworkers, complete large volumes of individual tasks that contribute to model
development. This can occur in all stages of model development. Before training, crowdworkers can
gather, curate, clean, and label training data. During development, crowdworkers can evaluate an
interim model and provide additional data for future training steps. After deployment, crowdworkers
can evaluate, moderate, or correct a model’s output. Crowdwork is often contracted to third-party
companies, such as Amazon Mechanical Turk [183].
Two key social impacts include working conditions and acknowledgment of the work itself via
documentation. Manual review is often used to limit the harmful outputs of AI systems, including
generative AI systems. Labor protections and pay vary, and crowd workers can be subject to graphic
content. Critical aspects of crowd work are often left poorly documented or undocumented entirely
[128].
3.7.1 What to Evaluate
Researchers and developers should examine whether crowdworking is conducted under established
standards, such as the Criteria for Fairer Microwork [
35
], the guidelines outlined in the Partnership
on AI’s Responsible Sourcing of Data Enrichment Services [
173
], or the Oxford Internet Institute’s
Fairwork Principles [
109
]. Concurrently, researchers and developers should document the role of
crowdwork in all dataset development undertaken during generative AI systems development, e.g.
using frameworks like CrowdWorkSheets [
97
] and sections 3.3 and 3.4 in Datasheets for Datasets
[
127
]. Details such as crowd workers’ demographics, the instructions given to them, andhow they
were assessed and compensated, are foundational for interpreting the output of AI systems shaped
by this labor [
227
]. Transparent reporting [
131
] can aid understanding model output and help audit
labor practices.
6See pricing for Amazon Web Services [12], Databricks [89], and VertexAI [140].
7See OpenAI API call pricing [257].
11
External evaluators can use evaluation metrics designed specifically around crowdwork, such as those
proposed by Fairwork [
109
], to assess quality of working conditions. Relevant labor law interventions
by jurisdiction may also apply. Since many critical crowdworking jobs involve long-term exposure to
traumatic content [
295
], such as child sexual abuse material or graphic depictions of violence [
269
],
the availability of immediate trauma support and long-term professional psychological support to
crowd workers can be documented. Other variables for documenting conditions include regular breaks
and psychological support, and controlling the expected amount of traumatic material annotators are
exposed to in any given session.
3.7.2 Limitations
The lack of regulation and rules around crowdworker protection for AI contributes to minimal
to no documentation or transparency. The lack of information makes crowd work difficult to
evaluate. Incentives to conduct crowd work at a low cost with little transparency contribute to
less literature on evaluating crowd work. Outsourcing labor also creates barriers to evaluation by
further complicating reporting structures, communication, and working conditions. Furthermore, the
precarious employment of crowdworkers can prevent crowdworkers from documenting substandard
working conditions.
4 Impacts: People and Society
Evaluating the impact of AI on people, communities, and societies [
135
] encounter similar challenges
as those arising in sampling, surveying, determining preferences, and working with human subjects
[
14
,
184
,
385
,
1
,
194
] in addition to challenges that stem from the scale at which AI development and
deployment occur. The scale and scope of generative AI technologies necessarily mean they interact
with national and global social systems, including economies, politics, and cultures. Taxonomies of
risks and harms of generative AI systems [
116
], including their impacts on human rights,
8
strongly
overlap with what should be evaluated. However, most taxonomies lack evaluations or examples of
evaluating social impact. We should understand the reason for evaluations, such as helping provide
data that can be critical for mitigating harmful impacts.
Timing will change how we view a system; training data and generated outputs may not reflect the
world in which it is deployed [
337
]. We also acknowledge how perceptions of society, and society
itself, have been influenced by existing AI and social media tools [
235
]. Historical context gives
insight into when social impacts engage with systemic harms [
261
,
369
]. In crafting and conducting
evaluations, we can often encroach on others’ privacy and autonomy due to the need for highly
personal information to evaluate how harms are enacted and distributed across populations [
15
]. Any
evaluations should examine how consent is obtained and its limitations. Similarly, evaluations should
also consider existing and potential future impacts for people included as data subjects [259].
Longer-term effects of systems embedded in society, such as wide-scale economic and labor impacts,
largely require the ideation of generative AI systems’ possible use cases and have fewer available
general evaluations. The following categories heavily depend on how generative AI systems are
deployed, ncluding the direct deployment environment. In the broader ecosystem, methods of
deployment [323] also affect social impact, especially the potential for misuse.
The following categories are high-level, non-exhaustive, and present a synthesis of the findings across
different modalities. They refer solely to what can be evaluated in the interactions of generative AI
systems with people and society:
• Trustworthiness and Autonomy
–Trust in Media and Information
–Overreliance on Outputs
–Personal Privacy and Sense of Self
• Inequality, Marginalization, and Violence
–Community Erasure
8See [9, 24, 275] for further discussion on AI and human rights
12
–
Long-term Amplification and Embedding of Marginalization by Exclusion (and Inclu-
sion)
–Abusive or Violent Content
• Concentration of Authority
–Militarization, Surveillance, and Weaponization
–Imposing Norms and Values
• Labor and Creativity
–Intellectual Property and Ownership
–Economy and Labor Market
• Ecosystem and Environment
–Widening Resource Gaps
–Environmental Impacts
4.1 Trustworthiness and Autonomy
Trustworthiness is complex and includes numerous properties relating to decisions throughout the AI
lifecycle, including potential use cases [
247
]. Generative AI systems’ inherent limitations, including
non-determinism, opacity, hallucinations or confabulations, harmful bias, vulnerability to adversarial
attacks, and a lack of reliability, can contribute to people’s concerns about whether a system could
be trusted in a particular instance. Mechanisms such as disclosure about system design, guardrails,
and other characteristics can improve trust [
56
]. Lessons can be drawn from parallel fields, for
example some have argued that the Zero Trust framework in cybersecurity, which calls for frequent
network verification, could also be applied for generative AI verification [
164
]. Human trust in
systems, institutions, and outputs further evolves as systems become increasingly embedded into
daily life. The increased ease of access to generating content and potential misinformation, and
difficulty distinguishing between human and AI-generated content, poses risks to trust in media and
content authenticity [2].
4.1.1 Trust in Media and Information
Contributors: Jessica Newman, Irene Solaiman, Canyu Chen, Arjun Subramonian
The increasing sophistication of generative AI has expanded the possibilities of misleading content and
disinformation campaigns [
71
] and made it harder for people to trust information [
57
]. AI-generated
misinformation [
139
] can be perpetuated by reinforcement and volume [
267
] when widely distributed
online. Real-world impacts can include loss of trust in mainstream news [
143
]. LLM-generated
misinformation can be harder to detect accurately than human-generated misinformation with the
same semantics, indicating it can have more deceptive styles [
70
]. GPT-3-level systems can also
produce more compelling disinformation [
327
]. Moreover, automated detection systems show flaws,
such as often incorrectly flagging non-native language speakers’ writing as machine-generated [
207
].
Issues highlighted in Section 3.3 can contribute to misleading information among subpopulations.
4.1.1.1 What to Evaluate Surveys can examine trust in AI systems [
152
,
272
] to output factual
information; trust in researchers, developers, and organizations developing and deploying AI [
232
];
mitigation and detection measures [
318
]; and trust in overall media and how it is distributed [
361
].
Quantitative surveys of consumers across countries gauging understanding of, satisfaction with, and
trust in outputs from generative systems, with the survey conducted multiple times over a given
period of time, can give insights to user trust in commercialization [130, 222].
Trust can be measured in the category of information, such as information about democratic and
policy institutions [
265
]. Evaluations and countermeasures of false and misleading information
remain challenging. There is no universal agreement about what constitutes misinformation, and
much of the research on intervention remains siloed [141].
In text, approaches include examining “adversarial factuality” [
335
], human preference votes, and
leaderboards [75] .
13
In image and video, approaches include drawing from deepfake detection methods [
383
] and compar-
ing model architectures’ impact on detection [293].
In audio, surveys include a level of trust in the person whose voice is replicated and/or the institution
or process that person represents [28].
4.1.1.2 Mitigations and Interventions Interventions include encouraging media users to scruti-
nize post accuracy before sharing [
268
], encouraging companies to use crowdsourced fact-checking
[
132
] and digital forensics to detect AI-generated content [
110
]. However, detection loses accuracy
as AI systems become more powerful [
298
]. Research efforts towards watermarking are ongoing
[
190
] yet can be circumvented by users, as are efforts to mitigate the memorization of undesired
concepts learned by diffusion models [123, 150].
Emerging legal and regulatory approaches around the world include the EU AI Act, which requires
labeling AI-generated content. Policymakers and developers can also ban use cases where false
outputs have the greatest risks. In the U.S., the Department of Commerce is developing guidance for
content authentication and watermarking to label AI-generated content, which will be used by Federal
agencies. The U. S. Federal Trade Commission is also working to prohibit the use of generative AI
for impersonation fraud [112].
4.1.2 Overreliance on Outputs
Contributors: Jessica Newman, Irene Solaiman
Overreliance on automation, in general, is a long-studied problem [
262
], and carries over in novel
ways to AI-generated content [
266
]. People are prone to overestimate and overtrust AI, including AI-
generated content, especially when outputs appear authoritative or when people are in time-sensitive
situations [
58
]. This can lead to the spread of biased and inaccurate information [
148
]. Persistent
security vulnerabilities that trick systems into outputting inaccurate information [
315
] exacerbate the
harm of overreliance. LLMs can also exhibit deceptive behavior in certain instances [263].
The study of human-generative AI relationships is nascent but growing, and highlights that the
anthropomorphism [
2
] of these technologies and automation bias [
125
] may contribute to unfounded
trust and reliance [
285
]. Improving the trustworthiness of AI systems is an important ongoing effort
across sectors.9
4.1.2.1 What to Evaluate In text, conversational interfaces for chatbots can elicit trust and
other strong emotions and can be abused to subtly change or manipulate people’s behaviors or even
encourage self-harm [
2
]. LLMs can be evaluated for their refusal to generate responses to questions
with non-existent concepts or false premises [
211
]. Here, we consider text and code separately.
Although both rely on textual representations, the goals, formatting, and semantics of code and text
are very dissimilar and require separate consideration and evaluation.
In code, inaccurate outputs [
82
] can nullify potential benefits and should be evaluated for their
limitations [
73
] and hazards [
186
], in addition to categories listed in the Technical Base System
section.
4.1.2.2 Mitigations and Interventions Protections include vulnerability disclosure, bug bounties,
and AI incident databases. In policy and legislation, components of the EU AI Act may also be helpful,
such as requiring labeling of AI-generated content and prohibiting certain kinds of manipulation
[
107
]. The U.S. Federal Trade Commission also protects consumers from false or exaggerated claims
about AI products [21].10
4.1.3 Personal Privacy and Sense of Self
Contributors: Jessica Newman, Irene Solaiman, Arjun Subramonian
9See [358, 247].
10
While the FTC has a mandate to protect consumers against false or exaggerated claims, they may opt not to
enforce them [180].
14
Privacy is linked with autonomy, referring to one’s ability to act under self-governance, where lack of
privacy can hinder one’s ability to act independently. Privacy can protect both powerful institutions
and vulnerable peoples, and is interpreted and protected differently by culture and social classes
throughout history [
237
]. Legal definitions and protections vary globally [
346
,
182
] and, when
violated, can be distinct from harm [
63
]. Privacy can refer to content shared, seen, or experienced
outside the sphere a person has consented to, or in which they expect it to appear or be inferable
[249]. Publicly-available content may have varying privacy considerations [317].
4.1.3.1 What to Evaluate In addition to system-level Privacy and Data Protection evaluations,
societal impacts [
325
] and harms [
324
] from the loss and violation of privacy are difficult to enumerate
and evaluate, such as loss of opportunity or reputational damage. Violations can shift in power
differentials and to personal expectations of privacy [
224
] and autonomy. The type of private
information violated, such as medical information, can trigger different impacts and responses.
4.1.3.2 Mitigations and Interventions Mitigations first should determine who is responsible for
an individual’s privacy while recognizing limitations of technical or data literacy. Robust protection
requires both individual and collective action [
10
]. Outside of an individualistic framework, certain
rights such as refusal [81] and inclusion also require consideration of individual self-determination.
Technical methods to preserve privacy in a generative AI system [
54
] cannot guarantee full protection.
Upholding privacy regulations requires engagement from multiple affected parties [
279
] and consider-
ing the effectiveness of methods such as opt-outs [
59
] from data collection [
213
]. Improving common
practices and better global regulation for collecting training data can help. Opt-in approaches can
provide better protection [
352
]. Privacy options for users should ease accessibility [
376
]. Meaningful
consent of data and model subjects is key, as per the EU’s General Data Protection Regulation [
106
].
4.2 Inequality, Marginalization, and Violence
Generative AI systems can exacerbate inequality, as argued in Sections 3.1 and 3.3. When deployed
or updated, systems’ impacts on people and groups can, directly and indirectly, harm and exploit
vulnerable and marginalized groups.
4.2.1 Community Erasure
Contributors: Dylan Baker, Yacine Jernite, Irene Solaiman, Zeerak Talat
Biases and safety provisions, such as content moderation, can have unequally distributed costs and
benefits and can lead to community erasure [
146
]. Common methods for harmful content removal can
lower system performances for marginalized communities [
382
] and suppress community-specific
and reclaimed language [
77
]. Automated removal can perform poorly or be harmful to marginalized
populations [303, 338].
Mitigations often combine four methods: data sourcing [
38
]; human moderation of content included
in training data [
87
]; automated moderation of content included in training data [
147
]; and keyword
deny-lists [
300
]. Given that the exclusion of harmful content within datasets stand to create distinct
harms to marginalized communities [
99
,
259
], efforts towards mitigation of generating harmful
content becomes a question of the politics of classification [
50
,
200
,
100
,
350
] and its potential harms.
4.2.1.1 What to Evaluate Evaluating Disparate Performance once systems have undergone safety
provisions can give signal to possible erasure. Accounting for the demographics and composition
of human crowdworkers can also provide information [
304
] about subsequent impacts. Longer-
term impacts of erasure depend on the system’s deployment context, leading to opportunity loss or
reinforced biases and norms.
4.2.1.2 Mitigations and Interventions Better democratic processes for developing and deploying
systems and safety provisions such as content moderation should work with marginalized populations.
This should include more investment in representative crowdworkers and appropriate compensation
and mental health support. Lessons from social media content moderation can apply, such as working
with groups who have been erased and documenting patterns of erasure to improve future approaches
[312].
15
4.2.2 Long-term Amplification and Embedding of Marginalization by Exclusion (and
Inclusion)
Contributors: Yacine Jernite, Irene Solaiman, Zeerak Talat
Inclusion without consent can also harm marginalized groups, including via surveillance and ex-
ploitation. For example, while some research strives for greater inclusion of underrepresented and
Indigenous languages [
167
], Indigenous groups have resisted AI approaches [
84
]. Work conducted
on low resource and Indigenous cultural artifacts, such as language and symbolism, should ensure
meaningful inclusion of the proprietors and community members [306].
4.2.2.1 Disparate Performance in Critical Infrastructure Generative AI use in critical infras-
tructure directly impacting human wellbeing can also be classified as high-risk use cases. This
includes use in financial services, healthcare such as mental health and medical advice, and demo-
cratic processes, such as election or political information. Examples include crisis intervention, as
well as research [
118
] and action [
238
] to use chatbots for eating disorder prevention. Technical
tooling used in human systems and processes that have long-recorded discrimination patterns [
374
]
can instead exacerbate harm [197].
Recent research highlights a disconnect between static fairness objectives and their long-term impacts
[
133
]. For instance, in lending, algorithms can reinforce stigmatization and worsen marginalization
over time [
212
]. Understanding these long-term effects requires considering complex interactions
between automated choices, individual responses, and societal dynamics.
4.2.2.2 What to Evaluate Systems should again undergo Disparate Performance evaluations
once updated for a high-risk task in critical infrastructure and account for overall deployment context.
Evaluating marginalization will depend on context, and should account for marginalization when work
about and by marginalized populations is less visible or uncredited [
377
]. Evaluating marginalization
impacts on individuals, such as through health [25], is ongoing research.
4.2.2.3 Mitigations and Interventions Information about disparate performance can improved
through better evaluation work for underrepresented populations and low-resource languages as well
as crediting and including local researchers [45] from these communities.
Engagement with populations should be done in ways that embody local approaches [
51
]. Policies
should be crafted to better respect rights of refusal [
320
], and which aim to mitigate the common power
disparities between model builders and these communities. Nations that address these discriminatory
patterns through regulations should coordinate with other nations to promote broad and international
protections where possible.
4.2.3 Abusive or Violence Content
Contributors: Irene Solaiman, Zeerak Talat
Generative AI systems can generate outputs that are used for abuse, constitute non-consensual content,
or are threats of violence and harassment [
68
]. Non-consensual sexual representations of people,
include representations of minors and child sexual abuse material (CSAM) [
240
]. Abuse and violence
can disparately affect groups, such as women and girls[
108
,
155
] These harms are additionally
experienced disproportionately on the basis of gender, race, ethnicity, sexual orientation, religion,
and other social categories [80].
4.2.3.1 What to Evaluate Sensitive topics and trauma’s impacts on people are by nature chal-
lenging to evaluate and should be done with care. Consequences of abuse of children and minors can
be long-term or lifelong [
8
]. Impacts and trauma can resurface throughout a person’s life in many
aspects. Evaluations for generative AI impacts can overlap with similar harms such as image-based
sexual abuse [
177
]. As seen in Section 3.2, consent from affected people should be evaluated with
the person themselves.
16
4.2.3.2 Mitigations and Interventions Research to detect, mitigate, and report abusive and
violent content such as CSAM is ongoing [
349
] and tools specific to modalities such as images can
help identify content that is not yet labeled as CSAM [
351
]. Additionally, datasets should be flagged
for containing abusive and violent content, and appropriately not be distributed or used for model
training [
40
,
348
]. Relevant regulation should be updated to address generated content that may
not accurately portray an existing person or their body or self, but lead to real harms. Institutions
such as the Canadian Centre for Child Protection (C3P) and the Internet Watch Foundation (IWF)
are dedicated to evaluating for CSAM with relevant legal context. Furthermore, training evaluation
models on CSAM or having CSAM reference content for evaluation is also often illegal.
4.3 Concentration of Authority
The concentration of power and authority in decisions about and access to generative AI occurs
in numerous interrelated and simultaneous, but not always straightforward ways. Concentrating
authoritative power can exacerbate inequality and lead to exploitation.
Few countries, companies, and organizations currently have the ability to develop advanced AI
systems [
340
], and the costs of training generative systems are high, which has caused a shift towards
market concentration [
193
]. Greater transparency across many different indicators is one way that
researchers hope to counteract some of the negative effects of the concentration of authority for public
accountability [47].
4.3.1 Militarization, Surveillance, and Weaponization
Contributors: Irene Solaiman, Jessica Newman
Concentrating power can occur at various levels, from small groups to national bodies. National level
power includes surveillance, and interest in the militarization of generative AI systems is growing
[
151
]. Use includes generating synthetic data for training AI systems [
359
] and military planning
[
113
]. Military use is not inherently weaponization and risk depends on the use case and government
interest. AI deployed for national security interests require differentiating national security interests
from undue harm [60].
Generative AI systems are also enabling new kinds of cyberattacks, and amplifying the possibilities
of existing cyberattacks. For example, synthetic audio has been used for more compelling fraud
and extortion [
181
]. LLMs are also facilitating disinformation campaigns, influence operations, and
phishing attacks [
138
]. Research shows Russian military intelligence actors, North Korean threat
actors, Iranian threat actors, and Chinese state-affiliated threat actors are using LLMs to enhance their
efforts for reconnaissance, social engineering, and cyber operations [229].
4.3.1.1 What to Evaluate If deployed covertly, under NDA, or without transparency, generative
AI systems used for surveillance or weaponization can be difficult to track or evaluate. Evaluations
can broadly analyze the quantity of where such systems have been deployed, such as the number of
devices sold, or number of system deployments, as a brute force measure. AI developers can also
study and monitor how threat actors use generative AI systems to better evade detection or expand
malicious capabilities. There are a small number of evaluations that test the ability of LLMs to help
carry out cyber attacks [271] or help develop chemical or biological weapons [205].
4.3.1.2 Mitigations and Interventions For release or procurement of technical systems, devel-
opers can restrict surveillance and weaponization as use cases [
191
]. Similarly, academic venues
can adopt codes of ethics for weaponization or violating human rights using AI [
246
]. Government
development of generative AI systems for surveillance and weaponization requires additional pro-
tocols. Governments and militaries can make commitments toward ethical and responsible uses of
AI [
92
] and joint commitments from multiple countries [
360
,
230
] can create accountability among
military powers. Regulatory approaches can draw boundaries for harmful uses by militaries, but will
grapple with tensions on what constitutes national security [
379
], operating within the framework of
International Humanitarian Law [
44
] with respect to autonomous weapons, and moving forward on
international agreements on autonomous weapons [161].
17
For organizations to protect themselves against cyberattacks that make use of generative AI technolo-
gies, standard cybersecurity practices such as multifactor authentication and Zero Trust defenses can
be helpful. AI developers should also continuously monitor their AI systems to understand and block
malicious uses and attacks.
4.3.2 Imposing Norms and Values
Contributors: Dylan Baker, Yacine Jernite, Marie-Therese Png, Irene Solaiman, Zeerak Talat
Global deployment of a model can consolidate power within a single, originating culture, to determine
and propagate acceptability across cultures [
39
,
233
,
353
]. Highest performing characteristics of
generative systems such as language, dominant cultural values, and embedded norms can overrepre-
sent regions outside of where a system is deployed. For example, a language model that is highest
performing in the English language can be deployed in a region with a different dominant language
and incentivize engaging in English, further excluding those who do not have an English language
background. Establishing or reinforcing goodness with certain languages, dialects, accents, imagery,
social norms, and other representations of peoples and cultures can contribute to these norms and
values imposition.
4.3.2.1 What to Evaluate In addition to evaluations and limitations outlined in Section 3.2,
complex, qualitative, and evolving cultural concepts such as beauty and success are viewed differently
in context of an application and cultural region. The impacts of norm and value impositions are
already manifesting [
165
] and require critical foresight as they evolve. Imposition contributes
to homogenization, including the suppression of marginalized identities [
94
], languages, cultural
practices, and epistemologies [221].
4.3.2.2 Mitigations and Interventions Mitigations should be cognizant of preserving irreducible
differences among cultures [
105
] and practicing value-sensitive design [
120
], including by focusing on
system components such as data extraction and use [
85
]. Methods for cultural value alignment [
322
]
can improve and require improving methods and infrastructure for working with underrepresented
groups. Novel alignment techniques by modality can determine preferable principles [
372
] and values
[
27
] for generative AI systems. Prominent AI regulations should account for “copycat” legislation in
other countries.
4.4 Labor and Creativity
AI systems deployed as tools and assistants for human labor, thought, and creativity, should be
evaluated for the ongoing effects generative AI systems have on skills, jobs, and the labor market.
4.4.1 Intellectual Property and Ownership
Contributors: Irene Solaiman, Yacine Jernite
Rights to the training data [
53
,
220
] and replicated or plagiarized work [
52
] are ongoing legal and
policy discussions [
366
], often by specific modality. Impacts of generated content to people and
society, such as reputational damage and economic loss [
172
], will necessarily coexist with impacts
and development of intellectual property law.
4.4.1.1 What to Evaluate Determining whether original content has been used in training data
depends on developer transparency or research on training data extraction [
66
]. Given the large sizes
of training datasets, possible methods of evaluating original content inclusion could be through search
and matching tools [
104
,
326
]. In addition to unclear legal implications, the ambiguity of impacts on
content ownership [102] makes evaluation difficult.
In image, surveys can examine AI authorship of generated imagery and measure user sentiments
towards AI imagery generally [76].
4.4.1.2 Mitigations and Interventions Similar to Personal Privacy and Sense of Self (see Sec-
tion 4.1.3), opt-in and opt-out mechanisms can protect intellectual property but depend on adherence.
18
Regulation and stricter rules from a developer organization about training material will differ by
modality. Ongoing lawsuits set legal precedent [
72
]. Tools are being developed to protect certain
modalities from being used nonconsensually as training data [345].
4.4.2 Economy and Labor Market
Contributors: Alberto Lusoli
Key considerations about the impact of automation and AI on employment center on whether these
technologies will generate new jobs or, in contrast, will lead to a large-scale worker displacement in
the next future [
316
]. Some suggest productive supplementing of repetitive tasks [
13
] while others
warn of displacement and polarization [
3
]. Automation unevenly affects workers, since efficiency
can be measured and prioritized disparately [
19
]. Long-term, research shows how technological
advancements have historically increased earning inequality between education, sex, race, and age
groups [
3
]. Market incentives and value attributed to varying skills and may not accurately reflect
societal needs, and are often based on gendered and racialized preconceptions of the value of labor
[19].
4.4.2.1 What to Evaluate In text, approaches to examine short-term effects of LLMs on produc-
tivity include measuring a selected group of people using an LLM to supplement a given task, where
productivity is measured as earnings per minute. This factors in time for a task and the quality of the
output [251].
Across modalities, substitution of labor for capital might cut costs in the short term [
79
]. For specific
tasks, evaluating the quality of generated output compared to human output can signal the likelihood
of a generative AI system replacing human labor [
309
]. The long-term impact on the global economy
is unclear and depends on industry decisions. Potential evaluation variables include unemployment
rates, salaries for a given skill or task, economic class divisions, and overall cost of services. Positive
externalities
11
could stimulate competition, drive prices down, and have a net-positive effect on
employment [
201
]. A task-polarization model [
22
] shows how AI can potentially widen the gap
between high and low-wage occupations at the expense of the middle tier. Further evaluations can
investigate types of jobs created and sunsetted.
See Section 3.7 for evaluating human labor in the research, development, and deployment process.
4.4.2.2 Mitigations and Interventions Workers affected by AI can be supported via re-skilling
and upskilling opportunities. This will also help reduce the barriers to entry to new jobs. Managing
the transition can be challenging and require policy intervention. Prominent movements and worker
disapproval, such as in the film and entertainment industry [
30
], can set precedence. In addition to
labor protection laws, more inclusive design processes can open technological decisions to democratic
participation and steer innovation in socially desirable directions.
Proposed policy interventions include an “automation tax” [
256
] to compensate for negative external-
ities (i.e., unemployment). In practice, an automation tax could be determined by the extent to which
layoffs can be attributed to automation, or by measuring the intensity of automation, and adjusting
the amount each firm contributes in unemployment insurance payments accordingly. Taxes can gener-
ate revenue to support re-skilling programs and slow the introduction of employment-substituting
technologies, providing governments with more time to prepare for the potential effects of structural
under- and unemployment. Limitations include difficulty in clearly identifying labor-saving from
labor-enhancing technologies, the risk of imposing double taxation on capital investments, and
possibly slowing technological innovation, GDP and wage growth [20].
4.5 Ecosystem and Environment
Impacts at a high-level, from the AI ecosystem to the Earth itself, are necessarily broad but can be
broken down into components for evaluation.
11
We broadly understand externalities as the unanticipated effects of economic activities on the social
environment.
19
4.5.1 Widening Resource Gaps
Contributors: Irene Solaiman
As described in Section 3.6, the high financial and resource costs necessarily excludes groups who do
not have the resources to train, evaluate, or host models. The infrastructure needed to contribute to
generative AI research and development leads to widening gaps which are notable among sectors,
such as between industry and academia [198, 329], or among global powers and countries [11].
4.5.1.1 Access and Benefit Distribution Ability to contribute to and benefit from a system
depends on ability to engage with a system, which in turn depends on the openness of the system, the
system application, and system interfaces. Level of openness and access grapples with tensions of
misuse and risk. Increasing trends toward system closedness [321] is shifting access distribution.
4.5.1.2 Geographic and Regional Activity Concentration In the field of AI as a whole histori-
cally, top AI research institutions from 1990-2014 have concentrated in the U.S. [
250
] More recent
data highlights the U.S., EU, and China as primary hubs [
294
]. Even within the U.S., AI activity
concentrates in urban, coastal areas [239].
4.5.1.3 What to Evaluate Evaluation should first determine AI-specific resources, then track
trends by sector and region. To determine and evaluate level of access, first components of access
should be established. This includes technical details, upstream decisions, auditing access, and
opt-out or opt-in reliability. Specific resources such as compute power [
6
] are popularly tracked by
annual reports on the field of AI [32, 329].
4.5.1.4 Mitigations and Interventions Policymakers can minimize resource gaps by making
high-cost resources, such as compute power, accessible via applications and grants to researchers and
low-resource organizations. Intercultural dialogues [
64
] that meaningfully address power imbalances
and lowering the barrier for underrepresented peoples to contribute can improve harms from resource
gaps. This can include accessible interfaces to interact with and conduct research on generative AI
systems and low- to no-code tooling.
4.5.2 Environmental Impacts
Contributors: Sasha Luccioni, Irene Solaiman, Marie-Therese Png, Michelle Lin
In addition to the Environmental Impacts and carbon emissions from a system itself, evaluating
impact on the Earth can follow popular frameworks and analyses.
4.5.2.1 What to Evaluate Environmental, social, and governance (ESG) frameworks and the
Scope 1, 2, and 3 system can give structure to how developers track carbon emissions [
288
]. Scope 3
emissions, the indirect emissions often outside a developer’s control, should account for a generative
AI system’s lifecycle, including in deployment [
216
]. Long-term effects of AI environmental impacts
on the world and people can range from inequity to quality of life [
287
]. Negative environmental
impacts of AI are unevenly distributed globally, with mining, data center water consumption, and
carbon emissions disproportionately affecting the Global South [
115
,
209
,
236
]. Research to measure
overall impacts of climate change is ongoing [358].
4.5.2.2 Mitigations and Interventions Systemic change can improve energy and carbon effi-
ciency in ML systems, from energy efficient default settings for platforms and tools, to an awareness
of balancing gains with cost; for example, weighing energy costs, both social and monetary, with the
performance gains of a new model before deploying it. Regulatory proposals move toward mixed
strategies for sustainable AI, including sustainability by design and consumption caps. Best practices
for developers and researchers include choosing efficient testing environments, promoting repro-
ducibility, and standardized reporting. An energy efficiency leaderboard can incentivize sustainable
research [149].
Antitrust overlap with environmental costs associated with the AI compute stack underscores a
complex interplay between market concentration, technological advancement, and environmental
20
sustainability. Vertical integration observed in the AI compute market has significant implications for
the environmental footprint of AI systems [
31
] when proprietary hardware and software configurations
may not be optimized for energy efficiency. By mandating interoperability and separating compute
hardware from software, regulatory measures could encourage the adoption of more energy-efficient
technologies across the AI lifecycle. Applying antitrust principles to AI compute markets could
incentivize greener technologies to attract environmentally conscious consumers and comply with
sustainability standards. Antitrust considerations include intervention through merger enforcement
and tackling anti-competitive conduct.
Standards and transparency for carbon emissions reporting and accounting for efficiency can help
better understand evolution and compare the emissions of different approaches and models. While
certain conferences such as NeurIPS are starting to include compute information in submissions as
checklists, reporting and figures can vary widely depending on what factors are included. Accuracy
may trade off with efficiency. Incentivizing and including these metrics when comparing two or more
models (e.g., in benchmarks and leaderboards) can help users make trade-offs that consider both
aspects and choose the model that best corresponds to their use case and criteria.
Legislative approaches emphasize the urgent need for comprehensive studies on the AI environmental
impacts of artificial intelligence [223].
Broader Impacts and Future Work
Understanding an AI system from conception to training to deployment requires insight into training
data, the model itself, and the use case/application into which the system is deployed. It also requires
understanding people, society, and how societal processes, institutions, and power are changed and
shifted by an AI system.
5 Lack of context for base model
Context is critical to robust evaluation; the way in which we define and evaluate harm in any given
application requires an understanding of the target industry, task, end-user, and model architecture.
Communication across model developers, model deployers, and end-users is key to developing a
comprehensive evaluation and risk mitigation strategy. Actors across the ecosystem should collaborate
to craft robust evaluations and invest in the safeguards needed to prevent harm.
6 Context of the Evaluation
Systems can be deployed in contexts where there is not sufficient attention towards evaluating and
moderating performance. This means disparate performance is not caught, as seen with social media
platform moderation outside of the most commonly-written languages and wealthiest countries [
297
].
Moreover, as cultural values change between cultural contexts, both within and outside of any given
language, the particular cultural values that are being evaluated should be made explicit. A byproduct
of such specificity is that it becomes clear where evaluations should be extended while providing a
framework for such extensions.
7 Choosing Evaluations
Further work is needed to compare, select, and document evaluations. The evaluations selected to
determine a model’s performance will impact the values it propagates out during deployment. There
is no universal evaluation by which to determine a model’s performance, and any evaluation metrics
should be used with deployment context in mind [
282
,
305
].Evaluations themselves require further
scrutiny and evaluation, to be able to determine the most appropriate evaluations to run. Appropriate
evaluations per category should be standardized through policy bodies and coordinated internationally.
Furthermore, notable work at top AI ethics publication venues has not adequately centered on the
least powerful in society [41].
21
Conclusion
Just as generative AI systems undergo performance evaluations, they must also be evaluated for social
impacts. The seven categories in our framework for technical base systems move toward a standard
evaluation framework for listed modalities of a base system. Our analyses of popular evaluation
methods per category can help to improve research in producing novel evaluations. Evaluations under
the “People and Society” category overlaps with existing risk and harms taxonomies for generative
AI systems. The latter evaluation category has limited case studies and must consider challenges
and ethics of determining human responses. Since social impact evaluations can only give limited
information about each impact type, we recommend that all categories are given equal importance,
and that all relevant stakeholders are meaningfully consulted throughout the development, evaluation,
and deployment processes.
Acknowledgements
We would like to thank Adina Williams, Levent Sagun, Kevin Klyman, and Apostol Vassilev for their
input to drafts of this chapter.
Authorship by Section
Bias, Stereotypes, and Representational Harms: Yacine Jernite, Lama Ahmad, Sara Hooker, Irene
Solaiman, Zeerak Talat, Margaret Mitchell, Usman Gohar, Jennifer A Mickel, Dylan Baker, Alina
Leidinger, Felix Friedrich, Anaelia Ovalle, Avijit Ghosh, Hal Daumé III
Cultural Values and Sensitive Content: Irene Solaiman, Zeerak Talat, Alina Leidinger, Isabella Duan,
Margaret Mitchell, Felix Friedrich, Jennifer Mickel
Disparate Performance: Yacine Jernite, Irene Solaiman, Usman Gohar, Jennifer Mickel, Margaret
Mitchell, Arjun Subramonian, Anaelia Ovalle, Avijit Ghosh, Hal Daumé III
Environmental Costs and Carbon Emissions: Sasha Luccioni, Marie-Therese Png, Irene Solaiman,
Usman Gohar, Michelle Lin
Privacy and Data Protection: Ellie Evans, Yacine Jernite, Irene Solaiman, Canyu Chen, Jessica
Newman, Shubham Singh, Isabella Duan, Lukas Struppek, Arjun Subramonian
Financial Costs: Irene Solaiman, Anaelia Ovalle
Data and Content Moderation Labor: Dylan Baker, Yacine Jernite, Alberto Lusoli, Irene Solaiman,
Jennifer Mickel, Arjun Subramonian
Trust in Media and Information: Jessica Newman, Irene Solaiman, Canyu Chen, Arjun Subramonian
Overreliance on Outputs: Jessica Newman, Irene Solaiman
Personal Privacy and Sense of Self : Jessica Newman, Irene Solaiman, Arjun Subramonian
Community Erasure: Dylan Baker, Yacine Jernite, Irene Solaiman, Zeerak Talat
Long-term Amplification and Embedding of Marginalization by Exclusion (and Inclusion): Yacine
Jernite, Irene Solaiman, Zeerak Talat
Abusive or Violence Content: Irene Solaiman, Zeerak Talat
Militarization, Surveillance, and Weaponization: Irene Solaiman, Jessica Newman
Imposing Norms and Values: Dylan Baker, Yacine Jernite, Marie-Therese Png, Irene Solaiman,
Zeerak Talat
Intellectual Property and Ownership: Irene Solaiman, Yacine Jernite
Economy and Labor Market: Alberto Lusoli, Irene Solaiman
Widening Resource Gaps: Irene Solaiman
Environmental Impacts: Sasha Luccioni, Irene Solaiman, Marie-Therese Png, Michelle Lin
22
References
[1]
L. Abbott and C. Grady. A Systematic Review of the Empirical Literature Evaluating IRBs:
What We Know and What We Still Need to Learn. Journal of Empirical Research on
Human Research Ethics, 6(1):3–19, Mar. 2011. ISSN 1556-2646, 1556-2654. doi: 10
.
1525/
jer.2011.6.1.3. URL http://journals.sagepub.com/doi/10.1525/jer.2011.6.1.3.
[2]
G. Abercrombie, A. Curry, T. Dinkar, V. Rieser, and Z. Talat. Mirages. On Anthropomorphism
in Dialogue Systems. In Proceedings of the 2023 Conference on Empirical Methods in Natural
Language Processing, pages 4776–4790, Singapore, 2023. Association for Computational
Linguistics. doi: 10
.
18653/v1/2023
.
emnlp-main
.
290. URL
https://aclanthology.org/
2023.emnlp-main.290.
[3]
D. Acemoglu and P. Restrepo. Tasks, Automation, and the Rise in U.S. Wage Inequality.
Econometrica, 90(5):1973–2016, 2022. ISSN 0012-9682. doi: 10
.
3982/ECTA19815. URL
https://www.econometricsociety.org/doi/10.3982/ECTA19815.
[4]
O. Ahia, J. Kreutzer, and S. Hooker. The Low-Resource Double Bind: An Empirical Study of
Pruning for Low-Resource Machine Translation. In Findings of the Association for Computa-
tional Linguistics: EMNLP 2021, pages 3316–3333, Punta Cana, Dominican Republic, 2021.
Association for Computational Linguistics. doi: 10
.
18653/v1/2021
.
findings-emnlp
.
282. URL
https://aclanthology.org/2021.findings-emnlp.282.
[5]
O. Ahia, S. Kumar, H. Gonen, J. Kasai, D. Mortensen, N. Smith, and Y. Tsvetkov. Do All
Languages Cost the Same? Tokenization in the Era of Commercial Language Models. In
Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,
pages 9904–9923, Singapore, 2023. Association for Computational Linguistics. doi: 10
.
18653/
v1/2023.emnlp-main.614. URL https://aclanthology.org/2023.emnlp-main.614.
[6]
N. Ahmed and M. Wahed. The De-democratization of AI: Deep Learning and the Com-
pute Divide in Artificial Intelligence Research, Oct. 2020. URL
http://arxiv.org/abs/
2010.15581.
[7]
O. Aka, K. Burke, A. Bauerle, C. Greer, and M. Mitchell. Measuring Model Biases in the
Absence of Ground Truth. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and
Society, pages 327–335, Virtual Event USA, July 2021. ACM. ISBN 978-1-4503-8473-5. doi:
10.1145/3461702.3462557. URL https://dl.acm.org/doi/10.1145/3461702.3462557.
[8]
A. Al Odhayani, W. J. Watson, and L. Watson. Behavioural consequences of child abuse.
Canadian Family Physician, 59(8):831–836, Aug. 2013. ISSN 0008-350X. URL
https:
//www.ncbi.nlm.nih.gov/pmc/articles/PMC3743691/.
[9]
All Tech Is Human. AI and Human Rights: Building a Tech Future Aligned With the
Public Interest. Technical report, All Tech Is Human, New York, NY, USA, 2022. URL
https://alltechishuman.org/ai-human-rights-report.
[10]
A. L. Allen. Protecting One’s Own Privacy in a Big Data Economy. Harvard Law Review, 130
(2), 2016. URL
https://harvardlawreview.org/forum/vol-130/protecting-ones-
own-privacy-in-a-big-data-economy/.
[11]
C. Alonso, S. Kothari, and S. Rehman. How Artificial Intelligence Could Widen the Gap
Between Rich and Poor Nations, Dec. 2020. URL
https://www.imf.org/en/Blogs/
Articles/2020/12/02/blog-how-artificial-intelligence-could-widen-the-
gap-between-rich-and-poor-nations.
[12]
Amazon Web Services. Amazon S3 Simple Storage Service Pricing. URL
https:
//aws.amazon.com/s3/pricing/.
[13]
N. Anantrasirichai and D. Bull. Artificial intelligence in the creative industries: A review.
Artificial Intelligence Review, 55(1):589–656, Jan. 2022. ISSN 1573-7462. doi: 10
.
1007/
s10462-021-10039-7. URL https://doi.org/10.1007/s10462-021-10039-7.
23
[14]
C. Andrade. The Limitations of Online Surveys. Indian Journal of Psychological Medicine,
42(6):575–576, Nov. 2020. ISSN 0253-7176, 0975-1564. doi: 10
.
1177/0253717620957496.
URL http://journals.sagepub.com/doi/10.1177/0253717620957496.
[15]
M. Andrus, E. Spitzer, J. Brown, and A. Xiang. What We Can’t Measure, We Can’t Understand:
Challenges to Demographic Data Procurement in the Pursuit of Fairness. In Proceedings of the
2021 ACM Conference on Fairness, Accountability, and Transparency, pages 249–260, Virtual
Event Canada, Mar. 2021. ACM. ISBN 978-1-4503-8309-7. doi: 10
.
1145/3442188
.
3445888.
URL https://dl.acm.org/doi/10.1145/3442188.3445888.
[16]
L. F. W. Anthony, B. Kanding, and R. Selvan. Carbontracker: Tracking and Predicting the
Carbon Footprint of Training Deep Learning Models, July 2020. URL
http://arxiv.org/
abs/2007.03051.
[17]
Anthropic. Introducing the next generation of Claude, 2024. URL
https://
www.anthropic.com/news/claude-3-family.
[18]
C. Ashurst, E. Hine, P. Sedille, and A. Carlier. AI Ethics Statements: Analysis and
Lessons Learnt from NeurIPS Broader Impact Statements. In 2022 ACM Conference on
Fairness, Accountability, and Transparency, pages 2047–2056, Seoul Republic of Korea,
June 2022. ACM. ISBN 978-1-4503-9352-2. doi: 10
.
1145/3531146
.
3533780. URL
https://dl.acm.org/doi/10.1145/3531146.3533780.
[19]
N. Atanasoski and K. Vora. Surrogate Humanity: Race, Robots, and the Politics of Techno-
logical Futures. Perverse Modernities. Duke University Press, Durham, NC, 2019. ISBN
978-1-4780-0386-1.
[20]
R. D. Atkinson. The Case Against Taxing Robots, Apr. 2019. URL
https://
papers.ssrn.com/abstract=3382824.
[21]
M. Atleson. Keep your AI claims in check, Feb. 2023. URL
https://www.ftc.gov/
business-guidance/blog/2023/02/keep-your-ai-claims-check.
[22]
D. Autor. The Labor Market Impacts of Technological Change: From Unbridled Enthusiasm to
Qualified Optimism to Vast Uncertainty, May 2022. URL
https://www.nber.org/papers/
w30074.
[23]
S. K. B, A. Chandrabose, and B. R. Chakravarthi. An Overview of Fairness in Data – Illumi-
nating the Bias in Data Pipeline. In B. R. Chakravarthi, J. P. McCrae, M. Zarrouk, K. Bali, and
P. Buitelaar, editors, Proceedings of the First Workshop on Language Technology for Equal-
ity, Diversity and Inclusion, pages 34–45, Kyiv, Apr. 2021. Association for Computational
Linguistics. URL https://aclanthology.org/2021.ltedi-1.5.
[24]
B-Tech and United Nations Human Rights Office of the High Commissioner. Tax-
onomy of Human Rights Risks Connected to Generative AI. Technical report,
United Nations Human Rights Office of the High Commissioner, 2023. URL
https://www.ohchr.org/sites/default/files/documents/issues/business/
b-tech/taxonomy-GenAI-Human-Rights-Harms.pdf.
[25]
F. O. Baah, A. M. Teitelman, and B. Riegel. Marginalization: Conceptualizing patient
vulnerabilities in the framework of social determinants of health – An integrative review.
Nursing inquiry, 26(1):e12268, Jan. 2019. ISSN 1320-7881. doi: 10
.
1111/nin
.
12268. URL
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6342665/.
[26]
E. Bagdasaryan, O. Poursaeed, and V. Shmatikov. Differential privacy has disparate impact on
model accuracy. In Proceedings of the 33rd International Conference on Neural Information
Processing Systems. Curran Associates Inc., Red Hook, NY, USA, 2019.
[27]
Y. Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirho-
seini, C. McKinnon, C. Chen, C. Olsson, C. Olah, D. Hernandez, D. Drain, D. Ganguli, D. Li,
E. Tran-Johnson, E. Perez, J. Kerr, J. Mueller, J. Ladish, J. Landau, K. Ndousse, K. Lukosuite,
L. Lovitt, M. Sellitto, N. Elhage, N. Schiefer, N. Mercado, N. DasSarma, R. Lasenby, R. Lar-
son, S. Ringer, S. Johnston, S. Kravec, S. E. Showk, S. Fort, T. Lanham, T. Telleen-Lawton,
24
T. Conerly, T. Henighan, T. Hume, S. R. Bowman, Z. Hatfield-Dodds, B. Mann, D. Amodei,
N. Joseph, S. McCandlish, T. Brown, and J. Kaplan. Constitutional AI: Harmlessness from AI
Feedback, Dec. 2022. URL http://arxiv.org/abs/2212.08073.
[28]
J. Barnett. The Ethical Implications of Generative Audio Models: A Systematic Literature
Review. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society,
pages 146–161, Montr\’{e}al QC Canada, Aug. 2023. ACM. ISBN 9798400702310. doi:
10.1145/3600211.3604686. URL https://dl.acm.org/doi/10.1145/3600211.3604686.
[29]
A. M. Barrett, D. Hendrycks, J. Newman, and B. Nonnecke. Actionable Guidance for High-
Consequence AI Risk Management: Towards Standards Addressing AI Catastrophic Risks,
Feb. 2023. URL http://arxiv.org/abs/2206.08966.
[30]
W. Bedingfield. Hollywood Writers Reached an AI Deal That Will Rewrite History,
2023. URL
https://www.wired.com/story/us-writers-strike-ai-provisions-
precedents/.
[31]
H. Belfield and S.-S. Hua. Compute and Antitrust: Regulatory implications of the AI hardware
supply chain, from chip design to cloud APIs. Verfassungsblog, Aug. 2022. URL
https:
//verfassungsblog.de/compute-and-antitrust/.
[32]
N. Benaich, A. Chalmers, O. Sebbouh, and C. Gurau. State of AI Report 2023. Technical
report, State of AI, 2023. URL https://www.stateof.ai/.
[33]
E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell. On the Dangers of Stochastic
Parrots: Can Language Models Be Too Big? . In Proceedings of the 2021 ACM Conference
on Fairness, Accountability, and Transparency, pages 610–623, Virtual Event Canada, Mar.
2021. ACM. ISBN 978-1-4503-8309-7. doi: 10
.
1145/3442188
.
3445922. URL
https:
//dl.acm.org/doi/10.1145/3442188.3445922.
[34]
C. L. Bennett, C. Gleason, M. K. Scheuerman, J. P. Bigham, A. Guo, and A. To. “It’s
Complicated”: Negotiating Accessibility and (Mis)Representation in Image Descriptions of
Race, Gender, and Disability. In Proceedings of the 2021 CHI Conference on Human Factors
in Computing Systems, CHI ’21, pages 1–19, New York, NY, USA, May 2021. Association
for Computing Machinery. ISBN 978-1-4503-8096-6. doi: 10
.
1145/3411764
.
3445498. URL
https://doi.org/10.1145/3411764.3445498.
[35]
J. Berg, M. Furrer, E. Harmon, U. Rani, and M. S. Silberman. Digital labour platforms and
the future of work: Towards decent work in the online world. Report, International Labour
Organization, Sept. 2018. URL https://apo.org.au/node/244461.
[36]
A. Berthelot, E. Caron, M. Jay, and L. Lefèvre. Estimating the environmental impact
of Generative-AI services using an LCA-based methodology. In CIRP LCE 2024 - 31st
Conference on Life Cycle Engineering, pages 1–10, Turin, Italy, June 2024. URL
https:
//inria.hal.science/hal-04346102.
[37]
S. Bhatt, S. Dev, P. Talukdar, S. Dave, and V. Prabhakaran. Re-contextualizing fairness in NLP:
The case of India. In Y. He, H. Ji, S. Li, Y. Liu, and C.-H. Chang, editors, Proceedings of the
2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics
and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long
Papers), pages 727–740, Online only, Nov. 2022. Association for Computational Linguistics.
URL https://aclanthology.org/2022.aacl-main.55.
[38]
BigScience Workshop, T. L. Scao, A. Fan, C. Akiki, E. Pavlick, S. Ili´
c, D. Hesslow, R. Castagné,
A. S. Luccioni, F. Yvon, M. Gallé, J. Tow, A. M. Rush, S. Biderman, A. Webson, P. S. Am-
manamanchi, T. Wang, B. Sagot, N. Muennighoff, A. V. del Moral, O. Ruwase, R. Bawden,
S. Bekman, A. McMillan-Major, I. Beltagy, H. Nguyen, L. Saulnier, S. Tan, P. O. Suarez,
V. Sanh, H. Laurençon, Y. Jernite, J. Launay, M. Mitchell, C. Raffel, A. Gokaslan, A. Simhi,
A. Soroa, A. F. Aji, A. Alfassy, A. Rogers, A. K. Nitzav, C. Xu, C. Mou, C. Emezue, C. Klamm,
C. Leong, D. van Strien, D. I. Adelani, D. Radev, E. G. Ponferrada, E. Levkovizh, E. Kim,
E. B. Natan, F. De Toni, G. Dupont, G. Kruszewski, G. Pistilli, H. Elsahar, H. Benyam-
ina, H. Tran, I. Yu, I. Abdulmumin, I. Johnson, I. Gonzalez-Dios, J. de la Rosa, J. Chim,
25
J. Dodge, J. Zhu, J. Chang, J. Frohberg, J. Tobing, J. Bhattacharjee, K. Almubarak, K. Chen,
K. Lo, L. Von Werra, L. Weber, L. Phan, L. B. allal, L. Tanguy, M. Dey, M. R. Muñoz,
M. Masoud, M. Grandury, M. Šaško, M. Huang, M. Coavoux, M. Singh, M. T.-J. Jiang,
M. C. Vu, M. A. Jauhar, M. Ghaleb, N. Subramani, N. Kassner, N. Khamis, O. Nguyen,
O. Espejel, O. de Gibert, P. Villegas, P. Henderson, P. Colombo, P. Amuok, Q. Lhoest, R. Har-
liman, R. Bommasani, R. L. López, R. Ribeiro, S. Osei, S. Pyysalo, S. Nagel, S. Bose, S. H.
Muhammad, S. Sharma, S. Longpre, S. Nikpoor, S. Silberberg, S. Pai, S. Zink, T. T. Tor-
rent, T. Schick, T. Thrush, V. Danchev, V. Nikoulina, V. Laippala, V. Lepercq, V. Prabhu,
Z. Alyafeai, Z. Talat, A. Raja, B. Heinzerling, C. Si, D. E. Ta¸sar, E. Salesky, S. J. Mielke, W. Y.
Lee, A. Sharma, A. Santilli, A. Chaffin, A. Stiegler, D. Datta, E. Szczechla, G. Chhablani,
H. Wang, H. Pandey, H. Strobelt, J. A. Fries, J. Rozen, L. Gao, L. Sutawika, M. S. Bari,
M. S. Al-shaibani, M. Manica, N. Nayak, R. Teehan, S. Albanie, S. Shen, S. Ben-David, S. H.
Bach, T. Kim, T. Bers, T. Fevry, T. Neeraj, U. Thakker, V. Raunak, X. Tang, Z.-X. Yong,
Z. Sun, S. Brody, Y. Uri, H. Tojarieh, A. Roberts, H. W. Chung, J. Tae, J. Phang, O. Press,
C. Li, D. Narayanan, H. Bourfoune, J. Casper, J. Rasley, M. Ryabinin, M. Mishra, M. Zhang,
M. Shoeybi, M. Peyrounette, N. Patry, N. Tazi, O. Sanseviero, P. von Platen, P. Cornette, P. F.
Lavallée, R. Lacroix, S. Rajbhandari, S. Gandhi, S. Smith, S. Requena, S. Patil, T. Dettmers,
A. Baruwa, A. Singh, A. Cheveleva, A.-L. Ligozat, A. Subramonian, A. Névéol, C. Lovering,
D. Garrette, D. Tunuguntla, E. Reiter, E. Taktasheva, E. Voloshina, E. Bogdanov, G. I. Winata,
H. Schoelkopf, J.-C. Kalo, J. Novikova, J. Z. Forde, J. Clive, J. Kasai, K. Kawamura, L. Hazan,
M. Carpuat, M. Clinciu, N. Kim, N. Cheng, O. Serikov, O. Antverg, O. van der Wal, R. Zhang,
R. Zhang, S. Gehrmann, S. Mirkin, S. Pais, T. Shavrina, T. Scialom, T. Yun, T. Limisiewicz,
V. Rieser, V. Protasov, V. Mikhailov, Y. Pruksachatkun, Y. Belinkov, Z. Bamberger, Z. Kasner,
A. Rueda, A. Pestana, A. Feizpour, A. Khan, A. Faranak, A. Santos, A. Hevia, A. Unldreaj,
A. Aghagol, A. Abdollahi, A. Tammour, A. HajiHosseini, B. Behroozi, B. Ajibade, B. Saxena,
C. M. Ferrandis, D. McDuff, D. Contractor, D. Lansky, D. David, D. Kiela, D. A. Nguyen,
E. Tan, E. Baylor, E. Ozoani, F. Mirza, F. Ononiwu, H. Rezanejad, H. Jones, I. Bhattacharya,
I. Solaiman, I. Sedenko, I. Nejadgholi, J. Passmore, J. Seltzer, J. B. Sanz, L. Dutra, M. Sama-
gaio, M. Elbadri, M. Mieskes, M. Gerchick, M. Akinlolu, M. McKenna, M. Qiu, M. Ghauri,
M. Burynok, N. Abrar, N. Rajani, N. Elkott, N. Fahmy, O. Samuel, R. An, R. Kromann, R. Hao,
S. Alizadeh, S. Shubber, S. Wang, S. Roy, S. Viguier, T. Le, T. Oyebade, T. Le, Y. Yang,
Z. Nguyen, A. R. Kashyap, A. Palasciano, A. Callahan, A. Shukla, A. Miranda-Escalada,
A. Singh, B. Beilharz, B. Wang, C. Brito, C. Zhou, C. Jain, C. Xu, C. Fourrier, D. L. Periñán,
D. Molano, D. Yu, E. Manjavacas, F. Barth, F. Fuhrimann, G. Altay, G. Bayrak, G. Burns,
H. U. Vrabec, I. Bello, I. Dash, J. Kang, J. Giorgi, J. Golde, J. D. Posada, K. R. Sivaraman,
L. Bulchandani, L. Liu, L. Shinzato, M. H. de Bykhovetz, M. Takeuchi, M. Pàmies, M. A.
Castillo, M. Nezhurina, M. Sänger, M. Samwald, M. Cullan, M. Weinberg, M. De Wolf,
M. Mihaljcic, M. Liu, M. Freidank, M. Kang, N. Seelam, N. Dahlberg, N. M. Broad, N. Muell-
ner, P. Fung, P. Haller, R. Chandrasekhar, R. Eisenberg, R. Martin, R. Canalli, R. Su, R. Su,
S. Cahyawijaya, S. Garda, S. S. Deshmukh, S. Mishra, S. Kiblawi, S. Ott, S. Sang-aroonsiri,
S. Kumar, S. Schweter, S. Bharati, T. Laud, T. Gigant, T. Kainuma, W. Kusa, Y. Labrak, Y. S.
Bajaj, Y. Venkatraman, Y. Xu, Y. Xu, Y. Xu, Z. Tan, Z. Xie, Z. Ye, M. Bras, Y. Belkada, and
T. Wolf. BLOOM: A 176B-Parameter Open-Access Multilingual Language Model, June 2023.
URL http://arxiv.org/abs/2211.05100.
[39]
A. Birhane and Z. Talat. It’s incomprehensible: On machine learning and decoloniality.
In S. Lindgren, editor, Handbook of Critical Studies of Artificial Intelligence, pages 128–
140. Edward Elgar Publishing, Nov. 2023. ISBN 978-1-80392-856-2 978-1-80392-855-5.
doi: 10
.
4337/9781803928562
.
00016. URL
https://www.elgaronline.com/view/book/
9781803928562/book-part-9781803928562-16.xml.
[40]
A. Birhane, V. U. Prabhu, and E. Kahembwe. Multimodal datasets: Misogyny, pornography,
and malignant stereotypes, Oct. 2021. URL http://arxiv.org/abs/2110.01963.
[41]
A. Birhane, E. Ruane, T. Laurent, M. S. Brown, J. Flowers, A. Ventresque, and C. L. Dancy.
The Forgotten Margins of AI Ethics. In 2022 ACM Conference on Fairness, Accountability,
and Transparency, pages 948–958, Seoul Republic of Korea, June 2022. ACM. ISBN 978-1-
4503-9352-2. doi: 10
.
1145/3531146
.
3533157. URL
https://dl.acm.org/doi/10.1145/
3531146.3533157.
26
[42]
A. Birhane, V. Prabhu, S. Han, and V. N. Boddeti. On Hate Scaling Laws For Data-Swamps,
June 2023. URL http://arxiv.org/abs/2306.13141.
[43]
A. Birhane, V. Prabhu, S. Han, V. N. Boddeti, and A. S. Luccioni. Into the LAIONs Den:
Investigating Hate in Multimodal Datasets, Nov. 2023. URL
http://arxiv.org/abs/
2311.03449.
[44]
M. Bo, L. Bruun, and V. Boulanin. Retaining Human Responsibility in the Development and
Use of Autonomous Weapon Systems: On Accountability for Violations of International
Humanitarian Law Involving AWS. Technical report, Stockholm International Peace
Research Institute, Oct. 2022. URL
https://www.sipri.org/publications/2022/
other-publications/retaining-human-responsibility-development-and-use-
autonomous-weapon-systems-accountability.
[45]
M. J. Bockarie. We need to end “parachute” research which sidelines the work of African scien-
tists. Quartz, Jan. 2019. URL
https://qz.com/africa/1536355/african-scientists-
are-sidelined-by-parachute-research-teams.
[46]
R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein,
J. Bohg, A. Bosselut, E. Brunskill, E. Brynjolfsson, S. Buch, D. Card, R. Castellon, N. S.
Chatterji, A. S. Chen, K. A. Creel, J. Davis, D. Demszky, C. Donahue, M. Doumbouya,
E. Durmus, S. Ermon, J. Etchemendy, K. Ethayarajh, L. Fei-Fei, C. Finn, T. Gale, L. E.
Gillespie, K. Goel, N. D. Goodman, S. Grossman, N. Guha, T. Hashimoto, P. Henderson,
J. Hewitt, D. E. Ho, J. Hong, K. Hsu, J. Huang, T. F. Icard, S. Jain, D. Jurafsky, P. Kalluri,
S. Karamcheti, G. Keeling, F. Khani, O. Khattab, P. W. Koh, M. S. Krass, R. Krishna,
R. Kuditipudi, A. Kumar, F. Ladhak, M. Lee, T. Lee, J. Leskovec, I. Levent, X. L. Li,
X. Li, T. Ma, A. Malik, C. D. Manning, S. P. Mirchandani, E. Mitchell, Z. Munyikwa,
S. Nair, A. Narayan, D. Narayanan, B. Newman, A. Nie, J. C. Niebles, H. Nilforoshan,
J. F. Nyarko, G. Ogut, L. Orr, I. Papadimitriou, J. S. Park, C. Piech, E. Portelance, C. Potts,
A. Raghunathan, R. Reich, H. Ren, F. Rong, Y. H. Roohani, C. Ruiz, J. Ryan, C. R’e,
D. Sadigh, S. Sagawa, K. Santhanam, A. Shih, K. P. Srinivasan, A. Tamkin, R. Taori, A. W.
Thomas, F. Tramèr, R. E. Wang, W. Wang, B. Wu, J. Wu, Y. Wu, S. M. Xie, M. Yasunaga,
J. You, M. A. Zaharia, M. Zhang, T. Zhang, X. Zhang, Y. Zhang, L. Zheng, K. Zhou, and
P. Liang. On the opportunities and risks of foundation models. ArXiv, 2021. URL
https:
//crfm.stanford.edu/assets/report.pdf.
[47]
R. Bommasani, K. Klyman, S. Longpre, S. Kapoor, N. Maslej, B. Xiong, D. Zhang, and
P. Liang. The Foundation Model Transparency Index, Oct. 2023. URL
http://arxiv.org/
abs/2310.12941.
[48]
R. Bommasani, P. Liang, and T. Lee. Holistic Evaluation of Language Models. Annals of
the New York Academy of Sciences, 1525(1):140–146, July 2023. ISSN 0077-8923, 1749-
6632. doi: 10
.
1111/nyas
.
15007. URL
https://nyaspubs.onlinelibrary.wiley.com/
doi/10.1111/nyas.15007.
[49]
L. Bouza, A. Bugeau, and L. Lannelongue. How to estimate carbon footprint when training
deep learning models? A guide and review. Environmental Research Communications, 5
(11):115014, Nov. 2023. ISSN 2515-7620. doi: 10
.
1088/2515-7620/acf81b. URL
https:
//iopscience.iop.org/article/10.1088/2515-7620/acf81b.
[50]
G. C. Bowker and S. L. Star. Sorting Things out: Classification and Its Consequences. Inside
Technology. MIT Press, Cambridge, Mass., 1. paperback ed., 8. print edition, 2008. ISBN
978-0-262-52295-3 978-0-262-02461-7.
[51]
M. Brereton, P. Roe, R. Schroeter, and A. Lee Hong. Beyond ethnography: Engagement
and reciprocity as foundations for design research out here. In Proceedings of the SIGCHI
Conference on Human Factors in Computing Systems, CHI ’14, pages 1183–1186, New York,
NY, USA, Apr. 2014. Association for Computing Machinery. ISBN 978-1-4503-2473-1. doi:
10.1145/2556288.2557374. URL https://doi.org/10.1145/2556288.2557374.
[52]
J. Brewster, M. Wang, and C. Palmer. How Copycat Sites Use AI to Plagiarize News Ar-
ticles. NewsWeek, 2023. URL
https://www.newsweek.com/how-copycat-sites-use-
ai-plagiarize-news-articles-1835212.
27
[53]
B. Brittain. OpenAI hit with new lawsuits from news outlets over AI training. Reuters,
Feb. 2024. URL
https://www.reuters.com/legal/litigation/openai-hit-with-
new-lawsuits-news-outlets-over-ai-training-2024-02-28/.
[54]
H. Brown, K. Lee, F. Mireshghallah, R. Shokri, and F. Tramèr. What Does it Mean for a
Language Model to Preserve Privacy? In 2022 ACM Conference on Fairness, Accountability,
and Transparency, pages 2280–2292, Seoul Republic of Korea, June 2022. ACM. ISBN 978-
1-4503-9352-2. doi: 10
.
1145/3531146
.
3534642. URL
https://dl.acm.org/doi/10.1145/
3531146.3534642.
[55]
T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan,
P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan,
R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin,
S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei.
Language models are few-shot learners. In Proceedings of the 34th International Conference
on Neural Information Processing Systems, NIPS’20, Red Hook, NY, USA, 2020. Curran
Associates Inc. ISBN 978-1-71382-954-6.
[56]
M. Brundage, S. Avin, J. Wang, H. Belfield, G. Krueger, G. Hadfield, H. Khlaaf, J. Yang,
H. Toner, R. Fong, T. Maharaj, P. W. Koh, S. Hooker, J. Leung, A. Trask, E. Bluemke,
J. Lebensold, C. O’Keefe, M. Koren, T. Ryffel, J. B. Rubinovitz, T. Besiroglu, F. Carugati,
J. Clark, P. Eckersley, S. de Haas, M. Johnson, B. Laurie, A. Ingerman, I. Krawczuk, A. Askell,
R. Cammarota, A. Lohn, D. Krueger, C. Stix, P. Henderson, L. Graham, C. Prunkl, B. Martin,
E. Seger, N. Zilberman, S. Ó. hÉigeartaigh, F. Kroeger, G. Sastry, R. Kagan, A. Weller,
B. Tse, E. Barnes, A. Dafoe, P. Scharre, A. Herbert-Voss, M. Rasser, S. Sodhani, C. Flynn,
T. K. Gilbert, L. Dyer, S. Khan, Y. Bengio, and M. Anderljung. Toward Trustworthy AI
Development: Mechanisms for Supporting Verifiable Claims, Apr. 2020. URL
http://
arxiv.org/abs/2004.07213.
[57]
B. Buchanan, A. Lohn, M. Musser, and K. Sedova. Truth, Lies, and Automation.
Technical report, Center for Security and Emerging Technology, 2021. URL
https:
//cset.georgetown.edu/publication/truth-lies-and-automation/.
[58]
Z. Buçinca, M. B. Malaya, and K. Z. Gajos. To Trust or to Think: Cognitive Forcing
Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of
the ACM on Human-Computer Interaction, 5(CSCW1):1–21, Apr. 2021. ISSN 2573-0142.
doi: 10.1145/3449287. URL http://arxiv.org/abs/2102.09692.
[59]
D. Bui, B. Tang, and K. G. Shin. Do Opt-Outs Really Opt Me Out? In Proceedings of
the 2022 ACM SIGSAC Conference on Computer and Communications Security, CCS ’22,
pages 425–439, New York, NY, USA, Nov. 2022. Association for Computing Machinery.
ISBN 978-1-4503-9450-5. doi: 10
.
1145/3548606
.
3560574. URL
https://dl.acm.org/
doi/10.1145/3548606.3560574.
[60]
W. Burke-White. Human Rights and National Security: The Strategic Correlation. Har-
vard Human Rights Journal, Jan. 2004. URL
https://scholarship.law.upenn.edu/
faculty_scholarship/960.
[61]
California Legislative Service. California Consumer Privacy Act of 2018, 2018. URL
https://oag.ca.gov/privacy/ccpa.
[62]
A. Caliskan, J. J. Bryson, and A. Narayanan. Semantics derived automatically from language
corpora contain human-like biases. Science, 356(6334):183–186, Apr. 2017. ISSN 0036-8075.
doi: 10.1126/science.aal4230.
[63]
M. Calo. The Boundaries of Privacy Harm. 86 Indiana Law Journal 1131 (2011), 86(3), July
2011. ISSN 0019-6665. URL
https://www.repository.law.indiana.edu/ilj/vol86/
iss3/8.
[64]
R. Capurro. The Promising Field of Intercultural Information Ethics. In J. J. Frühbauer,
T. Hausmanninger, and R. Capurro, editors, Localizing the Internet, pages 9–18. Brill | Fink,
Jan. 2018. ISBN 978-3-7705-4200-0 978-3-8467-4200-6. doi: 10
.
30965/9783846742006
_
002.
URL https://brill.com/view/book/edcoll/9783846742006/BP000002.xml.
28
[65]
N. Carlini, C. Liu, Ú. Erlingsson, J. Kos, and D. Song. The secret sharer: Evaluating and testing
unintended memorization in neural networks. In Proceedings of the 28th USENIX Conference
on Security Symposium, SEC’19, pages 267–284, USA, 2019. USENIX Association. ISBN
978-1-939133-06-9.
[66]
N. Carlini, J. Hayes, M. Nasr, M. Jagielski, V. Sehwag, F. Tramèr, B. Balle, D. Ippolito, and
E. Wallace. Extracting training data from diffusion models. In Proceedings of the 32nd
USENIX Conference on Security Symposium, SEC ’23, pages 5253–5270, USA, Aug. 2023.
USENIX Association. ISBN 978-1-939133-37-3.
[67]
N. Carlini, D. Ippolito, M. Jagielski, K. Lee, F. Tramer, and C. Zhang. Quantifying Memoriza-
tion Across Neural Language Models. In The Eleventh International Conference on Learning
Representations. ArXiv, Mar. 2023. URL http://arxiv.org/abs/2202.07646.
[68]
Center for Technology and Society. Americans’ Views on Generative Artificial Intelligence,
Hate and Harassment, 2023. URL
https://www.adl.org/resources/blog/americans-
views-generative-artificial-intelligence-hate-and-harassment.
[69]
A. Chan, R. Salganik, A. Markelius, C. Pang, N. Rajkumar, D. Krasheninnikov, L. Lan-
gosco, Z. He, Y. Duan, M. Carroll, M. Lin, A. Mayhew, K. Collins, M. Molamoham-
madi, J. Burden, W. Zhao, S. Rismani, K. Voudouris, U. Bhatt, A. Weller, D. Krueger,
and T. Maharaj. Harms from Increasingly Agentic Algorithmic Systems. In 2023 ACM
Conference on Fairness, Accountability, and Transparency, pages 651–666, Chicago IL
USA, June 2023. ACM. ISBN 9798400701924. doi: 10
.
1145/3593013
.
3594033. URL
https://dl.acm.org/doi/10.1145/3593013.3594033.
[70]
C. Chen and K. Shu. Can LLM-Generated Misinformation Be Detected? In The
Twelfth International Conference on Learning Representations, Oct. 2023. URL
https:
//openreview.net/forum?id=ccxD4mtkTU.
[71]
C. Chen and K. Shu. Combating Misinformation in the Age of LLMs: Opportunities and
Challenges, Nov. 2023. URL http://arxiv.org/abs/2311.05656.
[72]
M. Chen. Artists and Illustrators Are Suing Three A.I. Art Generators for
Scraping and ’Collaging’ Their Work Without Consent, Jan. 2023. URL
https://news.artnet.com/art-world/class-action-lawsuit-ai-generators-
deviantart-midjourney-stable-diffusion-2246770.
[73]
M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. d. O. Pinto, J. Kaplan, H. Edwards, Y. Burda,
N. Joseph, G. Brockman, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry,
P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Winter,
P. Tillet, F. P. Such, D. Cummings, M. Plappert, F. Chantzis, E. Barnes, A. Herbert-Voss, W. H.
Guss, A. Nichol, A. Paino, N. Tezak, J. Tang, I. Babuschkin, S. Balaji, S. Jain, W. Saunders,
C. Hesse, A. N. Carr, J. Leike, J. Achiam, V. Misra, E. Morikawa, A. Radford, M. Knight,
M. Brundage, M. Murati, K. Mayer, P. Welinder, B. McGrew, D. Amodei, S. McCandlish,
I. Sutskever, and W. Zaremba. Evaluating Large Language Models Trained on Code, July
2021. URL http://arxiv.org/abs/2107.03374.
[74] M. Cheng, E. Durmus, and D. Jurafsky. Marked Personas: Using Natural Language Prompts
to Measure Stereotypes in Language Models. In Proceedings of the 61st Annual Meeting of
the Association for Computational Linguistics (Volume 1: Long Papers), pages 1504–1532,
Toronto, Canada, 2023. Association for Computational Linguistics. doi: 10
.
18653/v1/2023
.
acl-
long.84. URL https://aclanthology.org/2023.acl-long.84.
[75]
W.-L. Chiang, L. Zheng, Y. Sheng, A. N. Angelopoulos, T. Li, D. Li, H. Zhang, B. Zhu,
M. Jordan, J. E. Gonzalez, and I. Stoica. Chatbot Arena: An Open Platform for Evaluating
LLMs by Human Preference, Mar. 2024. URL http://arxiv.org/abs/2403.04132.
[76]
S. G. Chiarella, G. Torromino, D. M. Gagliardi, D. Rossi, F. Babiloni, and G. Cartocci.
Investigating the negative bias towards artificial intelligence: Effects of prior assignment
of AI-authorship on the aesthetic appreciation of abstract paintings. Computers in Human
Behavior, 137:107406, Dec. 2022. ISSN 0747-5632. doi: 10
.
1016/j
.
chb
.
2022
.
107406. URL
https://www.sciencedirect.com/science/article/pii/S074756322200228X.
29
[77]
J. Chien, K. R. McKee, J. Kay, and W. Isaac. Recourse for reclamation: Chatting with
generative language models, Mar. 2024. URL http://arxiv.org/abs/2403.14467.
[78]
J. Cho, A. Zala, and M. Bansal. DALL-Eval: Probing the reasoning skills and social biases of
text-to-image generation models. In ICCV, 2023.
[79]
D. Y. Choi and J. H. Kang. Net Job Creation in an Increasingly Autonomous Economy:
The Challenge of a Generation. Journal of Management Inquiry, 28(3):300–305, July 2019.
ISSN 1056-4926. doi: 10
.
1177/1056492619827372. URL
https://doi.org/10.1177/
1056492619827372.
[80]
R. Chowdhury. Technology-Facilitated Gender-Based Violencein an Era of Generative
AI. World Trends in Freedom of Expression and Media Development Series. UNESCO,
France, 2023. ISBN 978-92-3-100631-9. URL
https://unesdoc.unesco.org/ark:
/48223/pf0000387483.
[81]
M. Cifor, P. Garcia, T. L. Cowan, J. Rault, J. Sutherland, A. L. Hoffman, N. Salehi, and
L. Nakamura. Feminist Data Manifest-No, 2019. URL https://www.manifestno.com.
[82]
T. Claburn. AI assistants help developers produce code that’s insecure. The Register, 2022.
URL https://www.theregister.com/2022/12/21/ai_assistants_bad_code/.
[83]
CodeCarbon. CodeCarbon. mlco2, May 2023. URL
https://github.com/mlco2/
codecarbon.
[84]
D. Coffey. M
¯
aori are trying to save their language from Big Tech. Wired, 2021. ISSN
1059-1028. URL https://www.wired.com/story/maori-language-tech/.
[85]
N. Couldry and U. A. Mejias. The decolonial turn in data and technology research: What
is at stake and where is it heading? Information, Communication & Society, 26(4):786–
802, Mar. 2023. ISSN 1369-118X. doi: 10
.
1080/1369118X
.
2021
.
1986102. URL
https:
//doi.org/10.1080/1369118X.2021.1986102.
[86]
K. Crenshaw. Mapping the Margins: Intersectionality, Identity Politics, and Violence against
Women of Color. Stanford Law Review, 43(6):1241, July 1991. ISSN 00389765. doi:
10.2307/1229039. URL https://www.jstor.org/stable/1229039?origin=crossref.
[87]
B. Dang, M. J. Riedl, and M. Lease. But Who Protects the Moderators? The Case of
Crowdsourced Image Moderation, Jan. 2020. URL http://arxiv.org/abs/1804.10999.
[88]
N. C. Dang, M. N. Moreno-García, and F. De La Prieta. Sentiment Analysis Based on Deep
Learning: A Comparative Study. Electronics, 9(3):483, Mar. 2020. ISSN 2079-9292. doi:
10.3390/electronics9030483. URL https://www.mdpi.com/2079-9292/9/3/483.
[89]
Databricks. Databricks pricing, Fri, 10/21/2022 - 20:37. URL
https://
www.databricks.com/product/pricing.
[90]
T. Davidson and D. Bhattacharya. Examining Racial Bias in an Online Abuse Corpus with
Structural Topic Modeling. In ICWSM Data Challenge, May 2020. URL
http://arxiv.org/
abs/2005.13041.
[91]
T. Davidson, D. Bhattacharya, and I. Weber. Racial Bias in Hate Speech and Abusive
Language Detection Datasets. In Proceedings of the Third Workshop on Abusive Language
Online, pages 25–35, Florence, Italy, 2019. Association for Computational Linguistics. doi:
10.18653/v1/W19-3504. URL https://www.aclweb.org/anthology/W19-3504.
[92]
Department of Defense. DOD Adopts Ethical Principles for Artificial Intelligence,
2020. URL
https://www.defense.gov/News/Releases/Release/Article/2091996/
dod-adopts-ethical-principles-for-artificial-intelligence/.
[93]
Department of International Cooperation Ministry of Science and Technology. Next Generation
Artificial Intelligence Development Plan. Technical report, Department of International
Cooperation,Ministry of Science and Technology (MOST), Beijing, China, 2017. URL
http:
//fi.china-embassy.gov.cn/eng/kxjs/201710/P020210628714286134479.pdf.
30
[94]
S. Dev, M. Monajatipoor, A. Ovalle, A. Subramonian, J. Phillips, and K.-W. Chang. Harms of
Gender Exclusivity and Challenges in Non-Binary Representation in Language Technologies.
In M.-F. Moens, X. Huang, L. Specia, and S. W.-t. Yih, editors, Proceedings of the 2021
Conference on Empirical Methods in Natural Language Processing, pages 1968–1994, Online
and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational Linguistics.
doi: 10
.
18653/v1/2021
.
emnlp-main
.
150. URL
https://aclanthology.org/2021.emnlp-
main.150.
[95]
S. Dev, J. Goyal, D. Tewari, S. Dave, and V. Prabhakaran. Building socio-culturally
inclusive stereotype resources with community engagement. In A. Oh, T. Neu-
mann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neu-
ral Information Processing Systems, volume 36, pages 4365–4381. Curran Associates,
Inc., 2023. URL
https://proceedings.neurips.cc/paper_files/paper/2023/file/
0dc91de822b71c66a7f54fa121d8cbb9-Paper-Datasets_and_Benchmarks.pdf.
[96]
H. Devinney, J. Björklund, and H. Björklund. Semi-supervised topic modeling for gender
bias discovery in English and Swedish. In M. R. Costa-jussà, C. Hardmeier, W. Radford,
and K. Webster, editors, Proceedings of the Second Workshop on Gender Bias in Natural
Language Processing, pages 79–92, Barcelona, Spain (Online), Dec. 2020. Association for
Computational Linguistics. URL https://aclanthology.org/2020.gebnlp-1.8.
[97]
M. Díaz, I. Kivlichan, R. Rosen, D. Baker, R. Amironesei, V. Prabhakaran, and E. Denton.
CrowdWorkSheets: Accounting for Individual and Collective Identities Underlying Crowd-
sourced Dataset Annotation. In 2022 ACM Conference on Fairness, Accountability, and
Transparency, pages 2342–2351, Seoul Republic of Korea, June 2022. ACM. ISBN 978-1-
4503-9352-2. doi: 10
.
1145/3531146
.
3534647. URL
https://dl.acm.org/doi/10.1145/
3531146.3534647.
[98]
E. Dinan, G. Abercrombie, A. S. Bergman, S. Spruit, D. Hovy, Y.-L. Boureau, and V. Rieser.
Anticipating Safety Issues in E2E Conversational AI: Framework and Tooling, July 2021. URL
http://arxiv.org/abs/2107.03451.
[99]
J. Dodge, M. Sap, A. Marasovi´
c, W. Agnew, G. Ilharco, D. Groeneveld, M. Mitchell, and
M. Gardner. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean
Crawled Corpus. In M.-F. Moens, X. Huang, L. Specia, and S. W.-t. Yih, editors, Proceedings
of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1286–
1305, Online and Punta Cana, Dominican Republic, Nov. 2021. Association for Computational
Linguistics. doi: 10
.
18653/v1/2021
.
emnlp-main
.
98. URL
https://aclanthology.org/
2021.emnlp-main.98.
[100]
M. Douglas. Purity and Danger: An Analysis of the Concepts of Pollution and Taboo.
Routledge, London, repr edition, 1978. ISBN 978-0-7100-1299-9 978-0-7100-8827-7.
[101]
C. Douwes, P. Esling, and J.-P. Briot. Energy Consumption of Deep Generative Audio Models,
Oct. 2021. URL http://arxiv.org/abs/2107.02621.
[102]
Dreyfus. Generative AI and the protection of intellectual property rights, May
2023. URL
https://www.dreyfus.fr/en/2023/05/22/generative-ai-balancing-
innovation-and-intellectual-property-rights-protection/.
[103]
M. Duan, A. Suri, N. Mireshghallah, S. Min, W. Shi, L. Zettlemoyer, Y. Tsvetkov, Y. Choi,
D. Evans, and H. Hajishirzi. Do Membership Inference Attacks Work on Large Language
Models?, Feb. 2024. URL http://arxiv.org/abs/2402.07841.
[104]
Y. Elazar, A. Bhagia, I. Magnusson, A. Ravichander, D. Schwenk, A. Suhr, P. Walsh, D. Groen-
eveld, L. Soldaini, S. Singh, H. Hajishirzi, N. A. Smith, and J. Dodge. What’s In My Big
Data?, Mar. 2024. URL http://arxiv.org/abs/2310.20707.
[105]
C. Ess. Ethical pluralism and global information ethics. Ethics and Information Technology, 8
(4):215–226, Nov. 2006. ISSN 1572-8439. doi: 10
.
1007/s10676-006-9113-3. URL
https:
//doi.org/10.1007/s10676-006-9113-3.
31
[106]
European Commission. Article 7 GDPR. Conditions for consent | GDPR-Text.com, 2016.
URL https://gdpr-text.com/read/article-7/.
[107]
European Commission. Proposal for a Regulation laying down harmonised rules
on artificial intelligence | Shaping Europe’s digital future, Apr. 2021. URL
https://digital-strategy.ec.europa.eu/en/library/proposal-regulation-
laying-down-harmonised-rules-artificial-intelligence.
[108]
European Institute for Gender Equality. Cyber violence is a growing threat,
especially for women and girls, 2017. URL
https://eige.europa.eu/
newsroom/news/cyber-violence-growing-threat-especially-women-and-
girls?language_content_entity=en.
[109] Fairwork. Fairwork. URL https://fair.work/en/fw/homepage/.
[110]
H. Farid. Creating, Using, Misusing, and Detecting Deep Fakes. Journal of Online Trust
and Safety, 1(4), Sept. 2022. ISSN 2770-3142. doi: 10
.
54501/jots
.
v1i4
.
56. URL
https:
//tsjournal.org/index.php/jots/article/view/56.
[111]
Federal Trade Commission. Protections Against Discrimination and Other Prohib-
ited Practices, 2003. URL
https://www.ftc.gov/policy-notices/no-fear-act/
protections-against-discrimination.
[112]
Federal Trade Commission. FTC Proposes New Protections to Combat AI Impersonation
of Individuals, Feb. 2024. URL
https://www.ftc.gov/news-events/news/press-
releases/2024/02/ftc-proposes-new-protections-combat-ai-impersonation-
individuals.
[113]
P. Feldman, A. Dant, and D. Rosenbluth. Ethics, Rules of Engagement, and AI: Neural
Narrative Mapping Using Large Transformer Language Models, Feb. 2022. URL
http:
//arxiv.org/abs/2202.02647.
[114]
V. Feldman. Does learning require memorization? a short tale about a long tail. In
Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing,
pages 954–959, Chicago IL USA, June 2020. ACM. ISBN 978-1-4503-6979-4. doi:
10.1145/3357713.3384290. URL https://dl.acm.org/doi/10.1145/3357713.3384290.
[115]
D. Feliba. In Latin America, data center plans fuel water worries, 2023. URL
https:
//www.reuters.com/article/idUSL8N3AU1PY/.
[116]
G. Fergusson, C. Fitzgerald, C. Frascella, M. Iorio, T. McBrien, C. Schroeder, B. Winters,
and E. Zhou. Generating Harms: Generative AI’s Impact & Paths Forward. Technical
report, Electronic Privacy Information Center, 2023. URL
https://epic.org/documents/
generating-harms-generative-ais-impact-paths-forward/.
[117]
A. Field, S. L. Blodgett, Z. Talat, and Y. Tsvetkov. A Survey of Race, Racism, and Anti-Racism
in NLP. In Proceedings of the 59th Annual Meeting of the Association for Computational
Linguistics and the 11th International Joint Conference on Natural Language Processing
(Volume 1: Long Papers), pages 1905–1925, Online, 2021. Association for Computational
Linguistics. doi: 10
.
18653/v1/2021
.
acl-long
.
149. URL
https://aclanthology.org/
2021.acl-long.149.
[118]
E. E. Fitzsimmons-Craft, W. W. Chan, A. C. Smith, M.-L. Firebaugh, L. A. Fowler, N. Topooco,
B. DePietro, D. E. Wilfley, C. B. Taylor, and N. C. Jacobson. Effectiveness of a chatbot for
eating disorders prevention: A randomized clinical trial. International Journal of Eating
Disorders, 55(3):343–353, 2022. ISSN 1098-108X. doi: 10
.
1002/eat
.
23662. URL
https:
//onlinelibrary.wiley.com/doi/abs/10.1002/eat.23662.
[119]
S. A. Friedler, C. Scheidegger, and S. Venkatasubramanian. The (Im)possibility of fairness: Dif-
ferent value systems require different mechanisms for fair decision making. Communications
of the ACM, 64(4):136–143, Apr. 2021. ISSN 0001-0782, 1557-7317. doi: 10
.
1145/3433949.
URL https://dl.acm.org/doi/10.1145/3433949.
32
[120]
B. Friedman, P. H. Kahn, A. Borning, and A. Huldtgren. Value Sensitive Design and Infor-
mation Systems. In N. Doorn, D. Schuurbiers, I. van de Poel, and M. E. Gorman, editors,
Early Engagement and New Technologies: Opening up the Laboratory, pages 55–95. Springer
Netherlands, Dordrecht, 2013. ISBN 978-94-007-7844-3. doi: 10
.
1007/978-94-007-7844-3
_
4.
URL https://doi.org/10.1007/978-94-007-7844-3_4.
[121]
F. Friedrich, M. Brack, L. Struppek, D. Hintersdorf, P. Schramowski, S. Luccioni, and K. Ker-
sting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness, July 2023.
URL http://arxiv.org/abs/2302.10893.
[122]
F. Friedrich, K. Hämmerl, P. Schramowski, J. Libovicky, K. Kersting, and A. Fraser. Multilin-
gual Text-to-Image Generation Magnifies Gender Stereotypes and Prompt Engineering May
Not Help You, Jan. 2024. URL http://arxiv.org/abs/2401.16092.
[123]
R. Gandikota, J. Materzy´
nska, J. Fiotto-Kaufman, and D. Bau. Erasing Concepts from
Diffusion Models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV),
pages 2426–2436, Paris, France, Oct. 2023. IEEE. ISBN 9798350307184. doi: 10
.
1109/
ICCV51070.2023.00230. URL https://ieeexplore.ieee.org/document/10378568/.
[124]
L. Gao, J. Tow, B. Abbasi, S. Biderman, S. Black, A. DiPofi, C. Foster, L. Golding, J. Hsu,
A. Le Noac’h, H. Li, K. McDonell, N. Muennighoff, C. Ociepa, J. Phang, L. Reynolds,
H. Schoelkopf, A. Skowron, L. Sutawika, E. Tang, A. Thite, B. Wang, K. Wang, and A. Zou. A
framework for few-shot language model evaluation, Dec. 2023. URL
https://zenodo.org/
records/10256836.
[125]
M. Gaske. Corporate Officers’ Fiduciary Duty to Monitor Generative Artificial Intelligence,
Dec. 2023. URL https://papers.ssrn.com/abstract=4664899.
[126]
W. Gaviria Rojas, S. Diamos, K. Kini, D. Kanter, V. Janapa Reddi, and C. Coleman. The dollar
street dataset: Images representing the geographic and socioeconomic diversity of the world.
In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in
Neural Information Processing Systems, volume 35, pages 12979–12990. Curran Associates,
Inc., 2022. URL
https://proceedings.neurips.cc/paper_files/paper/2022/file/
5474d9d43c0519aa176276ff2c1ca528-Paper-Datasets_and_Benchmarks.pdf.
[127]
T. Gebru, J. Morgenstern, B. Vecchione, J. W. Vaughan, H. Wallach, H. D. Iii, and K. Crawford.
Datasheets for datasets. Communications of the ACM, 64(12):86–92, Dec. 2021. ISSN
0001-0782, 1557-7317. doi: 10
.
1145/3458723. URL
https://dl.acm.org/doi/10.1145/
3458723.
[128]
R. S. Geiger, K. Yu, Y. Yang, M. Dai, J. Qiu, R. Tang, and J. Huang. Garbage in, garbage
out?: Do machine learning application papers in social computing report where human-
labeled training data comes from? In Proceedings of the 2020 Conference on Fairness,
Accountability, and Transparency, pages 325–336, Barcelona Spain, Jan. 2020. ACM. ISBN
978-1-4503-6936-7. doi: 10
.
1145/3351095
.
3372862. URL
https://dl.acm.org/doi/
10.1145/3351095.3372862.
[129]
A. Ghosh, R. Dutt, and C. Wilson. When Fair Ranking Meets Uncertain Inference. In
Proceedings of the 44th International ACM SIGIR Conference on Research and Development
in Information Retrieval, pages 1033–1043, Virtual Event Canada, July 2021. ACM. ISBN 978-
1-4503-8037-9. doi: 10
.
1145/3404835
.
3462850. URL
https://dl.acm.org/doi/10.1145/
3404835.3462850.
[130]
N. Gillespie. Trust in artificial intelligence - KPMG Global, Jan. 2024. URL
https://kpmg.com/xx/en/home/insights/2023/09/trust-in-artificial-
intelligence.html.
[131]
A. Glaese, N. McAleese, M. Tr˛ebacz, J. Aslanides, V. Firoiu, T. Ewalds, M. Rauh, L. Wei-
dinger, M. Chadwick, P. Thacker, L. Campbell-Gillingham, J. Uesato, P.-S. Huang, R. Co-
manescu, F. Yang, A. See, S. Dathathri, R. Greig, C. Chen, D. Fritz, J. S. Elias, R. Green,
S. Mokrá, N. Fernando, B. Wu, R. Foley, S. Young, I. Gabriel, W. Isaac, J. Mellor, D. Hassabis,
K. Kavukcuoglu, L. A. Hendricks, and G. Irving. Improving alignment of dialogue agents via
targeted human judgements, Sept. 2022. URL http://arxiv.org/abs/2209.14375.
33
[132]
W. Godel, Z. Sanderson, K. Aslett, J. Nagler, R. Bonneau, N. Persily, and J. Tucker. Moderating
with the Mob: Evaluating the Efficacy of Real-Time Crowdsourced Fact-Checking. Journal of
Online Trust and Safety, 1(1), Oct. 2021. ISSN 2770-3142. doi: 10
.
54501/jots
.
v1i1
.
15. URL
https://tsjournal.org/index.php/jots/article/view/15.
[133]
U. Gohar and L. Cheng. A Survey on Intersectional Fairness in Machine Learning: Notions,
Mitigation, and Challenges. In Proceedings of the Thirty-Second International Joint Confer-
ence on Artificial Intelligence, pages 6619–6627, Macau, SAR China, Aug. 2023. International
Joint Conferences on Artificial Intelligence Organization. ISBN 978-1-956792-03-4. doi:
10.24963/ijcai.2023/742. URL https://www.ijcai.org/proceedings/2023/742.
[134]
U. Gohar, S. Biswas, and H. Rajan. Towards understanding fairness and its composition in
ensemble machine learning. In 2023 IEEE/ACM 45th International Conference on Software
Engineering (ICSE), pages 1533–1545. IEEE, 2023.
[135]
U. Gohar, M. C. Hunter, A. Marczak-Czajka, R. R. Lutz, M. B. Cohen, and J. Cleland-Huang.
Towards engineering fair and equitable software systems for managing low-altitude airspace
authorizations. arXiv preprint arXiv:2401.07353, 2024.
[136]
S. Goldfarb-Tarrant, R. Marchant, R. Muñoz Sánchez, M. Pandya, and A. Lopez. Intrinsic
Bias Metrics Do Not Correlate with Application Bias. In Proceedings of the 59th Annual
Meeting of the Association for Computational Linguistics and the 11th International Joint
Conference on Natural Language Processing (Volume 1: Long Papers), pages 1926–1940,
Online, 2021. Association for Computational Linguistics. doi: 10
.
18653/v1/2021
.
acl-long
.
150.
URL https://aclanthology.org/2021.acl-long.150.
[137]
S. Goldfarb-Tarrant, E. Ungless, E. Balkir, and S. L. Blodgett. This prompt is measuring
<
mask
>
: Evaluating bias evaluation in language models. In Findings of the Association for
Computational Linguistics: ACL 2023, pages 2209–2225, Toronto, Canada, 2023. Association
for Computational Linguistics. doi: 10
.
18653/v1/2023
.
findings-acl
.
139. URL
https://
aclanthology.org/2023.findings-acl.139.
[138]
J. A. Goldstein, G. Sastry, M. Musser, R. DiResta, M. Gentzel, and K. Sedova. Generative
Language Models and Automated Influence Operations: Emerging Threats and Potential
Mitigations, Jan. 2023. URL http://arxiv.org/abs/2301.04246.
[139]
J. A. Goldstein, J. Chao, S. Grossman, A. Stamos, and M. Tomz. How persuasive is AI-
generated propaganda? PNAS Nexus, 3(2):pgae034, Feb. 2024. ISSN 2752-6542. doi:
10
.
1093/pnasnexus/pgae034. URL
https://academic.oup.com/pnasnexus/article/
doi/10.1093/pnasnexus/pgae034/7610937.
[140]
Google Cloud. Pricing | Vertex AI | Google Cloud. URL
https://cloud.google.com/
vertex-ai/pricing.
[141]
Y. Green, A. Gully, Y. Roth, A. Roy, J. A. Tucker, and A. Wanless. Evidence-Based
Misinformation Interventions: Challenges and Opportunities for Measurement and Col-
laboration, 2023. URL
https://carnegieendowment.org/2023/01/09/evidence-
based-misinformation-interventions-challenges-and-opportunities-for-
measurement-and-collaboration-pub-88661.
[142]
X. Gu, C. Du, T. Pang, C. Li, M. Lin, and Y. Wang. On Memorization in Diffusion Models,
Oct. 2023. URL http://arxiv.org/abs/2310.02664.
[143]
A. M. Guess, P. Barberá, S. Munzert, and J. Yang. The consequences of online partisan media.
Proceedings of the National Academy of Sciences, 118(14):e2013464118, Apr. 2021. ISSN
0027-8424, 1091-6490. doi: 10
.
1073/pnas
.
2013464118. URL
https://pnas.org/doi/
full/10.1073/pnas.2013464118.
[144]
U. Gupta, Y. G. Kim, S. Lee, J. Tse, H.-H. S. Lee, G.-Y. Wei, D. Brooks, and C.-J. Wu.
Chasing Carbon: The Elusive Environmental Footprint of Computing. IEEE Micro, 42(4):
37–47, July 2022. ISSN 0272-1732, 1937-4143. doi: 10
.
1109/MM
.
2022
.
3163226. URL
https://ieeexplore.ieee.org/document/9744492/.
34
[145]
C. Haerpfer, R. Inglehart, A. Moreno, C. Welzel, K. Kizilova, J. Diez-Medrano, M. La-
gos, P. Norris, E. Ponarin, and B. Puranen. World Values Survey Wave 7 (2017-
2022) Cross-National Data-Set, 2022. URL
http://www.worldvaluessurvey.org/
WVSDocumentationWV7.jsp.
[146]
O. L. Haimson, D. Delmonaco, P. Nie, and A. Wegner. Disproportionate Removals and
Differing Content Moderation Experiences for Conservative, Transgender, and Black Social
Media Users: Marginalization and Moderation Gray Areas. Proceedings of the ACM on
Human-Computer Interaction, 5(CSCW2):466:1–466:35, Oct. 2021. doi: 10
.
1145/3479610.
URL https://dl.acm.org/doi/10.1145/3479610.
[147]
T. Hartvigsen, S. Gabriel, H. Palangi, M. Sap, D. Ray, and E. Kamar. ToxiGen: A Large-
Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection. In
S. Muresan, P. Nakov, and A. Villavicencio, editors, Proceedings of the 60th Annual Meeting
of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3309–3326,
Dublin, Ireland, May 2022. Association for Computational Linguistics. doi: 10
.
18653/v1/
2022.acl-long.234. URL https://aclanthology.org/2022.acl-long.234.
[148]
W. D. Heaven. Why Meta’s latest large language model survived only three days online.
MIT Technology Review, 2022. URL
https://www.technologyreview.com/2022/11/
18/1063487/meta-large-language-model-ai-only-survived-three-days-gpt-
3-science/.
[149]
P. Henderson, J. Hu, J. Romoff, E. Brunskill, D. Jurafsky, and J. Pineau. Towards the
Systematic Reporting of the Energy and Carbon Footprints of Machine Learning. Journal of
Machine Learning Research, 21(248):1–43, 2020. ISSN 1533-7928. URL
http://jmlr.org/
papers/v21/20-312.html.
[150]
D. Hintersdorf, L. Struppek, D. Neider, and K. Kersting. Defending Our Privacy With
Backdoors, Feb. 2024. URL http://arxiv.org/abs/2310.08320.
[151]
M. Hirsh. How AI Will Revolutionize Warfare, 2023. URL
https://foreignpolicy.com/
2023/04/11/ai-arms-race-artificial-intelligence-chatgpt-military-
technology/.
[152]
L. Hofeditz, M. Mirbabaie, J. Holstein, and S. Stieglitz. Do You trust an AI-Journalist? A
Credibility Analysis of News Content with AI-Authorship. ECIS 2021 Research Papers, June
2021. URL https://aisel.aisnet.org/ecis2021_rp/50.
[153]
G. Hofstede. Dimensionalizing Cultures: The Hofstede Model in Context. Online Readings in
Psychology and Culture, 2(1), Dec. 2011. ISSN 2307-0919. doi: 10
.
9707/2307-0919
.
1014.
URL https://scholarworks.gvsu.edu/orpc/vol2/iss1/8.
[154]
G. Hofstede. Culture’s Consequences: Comparing Values, Behaviors, Institutions, and
Organizations across Nations. Sage, Thousand Oaks, Calif., 2. ed. [nachdr.] edition, 2013.
ISBN 978-0-8039-7324-4 978-0-8039-7323-7.
[155]
Home Security Heroes. STATE OF DEEPFAKES: Realities, Threats, and Im-
pact., 2023. URL
https://blog.biocomm.ai/2024/02/18/report-2023-state-of-
deepfakes-realities-threats-and-impact/.
[156]
S. Hooker. Moving beyond “algorithmic bias is a data problem”. Patterns, 2(4):
100241, Apr. 2021. ISSN 26663899. doi: 10
.
1016/j
.
patter
.
2021
.
100241. URL
https:
//linkinghub.elsevier.com/retrieve/pii/S2666389921000611.
[157]
S. Hooker, N. Moorosi, G. Clark, S. Bengio, and E. Denton. Characterising Bias in Compressed
Models, Dec. 2020. URL http://arxiv.org/abs/2010.03058.
[158]
House of Commons. Digital Charter Implementation Act, 2022. URL
https://www.parl.ca/
DocumentViewer/en/44-1/bill/C-27/first-reading.
35
[159]
D. Hovy and S. L. Spruit. The Social Impact of Natural Language Processing. In Proceedings
of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short
Papers), pages 591–598, Berlin, Germany, 2016. Association for Computational Linguistics.
doi: 10.18653/v1/P16-2096. URL http://aclweb.org/anthology/P16-2096.
[160]
H. Huang, T. Tang, D. Zhang, X. Zhao, T. Song, Y. Xia, and F. Wei. Not All Lan-
guages Are Created Equal in LLMs: Improving Multilingual Capability by Cross-Lingual-
Thought Prompting. In Findings of the Association for Computational Linguistics: EMNLP
2023, pages 12365–12394, Singapore, 2023. Association for Computational Linguis-
tics. doi: 10
.
18653/v1/2023
.
findings-emnlp
.
826. URL
https://aclanthology.org/
2023.findings-emnlp.826.
[161]
Human Rights Watch. Agenda for action: Alternative processes for negotiating a killer
robots treaty, 2022. URL
https://www.hrw.org/report/2022/11/10/agenda-action/
alternative-processes-negotiating-killer-robots-treaty.
[162]
B. Hutchinson and M. Mitchell. 50 Years of Test (Un)fairness: Lessons for Machine Learning.
In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 49–58,
Atlanta GA USA, Jan. 2019. ACM. ISBN 978-1-4503-6125-5. doi: 10
.
1145/3287560
.
3287600.
URL https://dl.acm.org/doi/10.1145/3287560.3287600.
[163]
B. Hutchinson, N. Rostamzadeh, C. Greer, K. Heller, and V. Prabhakaran. Evaluation Gaps
in Machine Learning Practice. In 2022 ACM Conference on Fairness, Accountability, and
Transparency, pages 1859–1876, Seoul Republic of Korea, June 2022. ACM. ISBN 978-1-
4503-9352-2. doi: 10
.
1145/3531146
.
3533233. URL
https://dl.acm.org/doi/10.1145/
3531146.3533233.
[164]
A. N. Institute. Zero Trust AI Governance, Aug. 2023. URL
https://
ainowinstitute.org/publication/zero-trust-ai-governance.
[165]
Institute of Medicine (US) Committee on Understanding and Eliminating Racial and Ethnic
Disparities in Health Care. Unequal Treatment: Confronting Racial and Ethnic Disparities
in Health Care. National Academies Press (US), Washington (DC), 2003. URL
http:
//www.ncbi.nlm.nih.gov/books/NBK220358/
. Eds. Smedley, Brian D. and Stith, Adrienne
Y. and Nelson, Alan R.
[166]
Y. Ishibashi and H. Shimodaira. Knowledge Sanitization of Large Language Models, Mar.
2024. URL http://arxiv.org/abs/2309.11852.
[167]
J. James, V. Yogarajan, I. Shields, C. Watson, P. Keegan, K. Mahelona, and P.-L. Jones.
Language Models for Code-switch Detection of te reo M
¯
aori and English in a Low-resource
Setting. In M. Carpuat, M.-C. de Marneffe, and I. V. Meza Ruiz, editors, Findings of the
Association for Computational Linguistics: NAACL 2022, pages 650–660, Seattle, United
States, July 2022. Association for Computational Linguistics. doi: 10
.
18653/v1/2022
.
findings-
naacl.49. URL https://aclanthology.org/2022.findings-naacl.49.
[168]
Y. Jernite. Let’s talk about biases in machine learning! Ethics and Society Newsletter #2, 2022.
URL https://huggingface.co/blog/ethics-soc-2.
[169]
Y. Jernite, H. Nguyen, S. Biderman, A. Rogers, M. Masoud, V. Danchev, S. Tan, A. S.
Luccioni, N. Subramani, I. Johnson, G. Dupont, J. Dodge, K. Lo, Z. Talat, D. Radev,
A. Gokaslan, S. Nikpoor, P. Henderson, R. Bommasani, and M. Mitchell. Data Gover-
nance in the Age of Large-Scale Data-Driven Language Technology. In 2022 ACM Conference
on Fairness, Accountability, and Transparency, pages 2206–2222, Seoul Republic of Ko-
rea, June 2022. ACM. ISBN 978-1-4503-9352-2. doi: 10
.
1145/3531146
.
3534637. URL
https://dl.acm.org/doi/10.1145/3531146.3534637.
[170]
Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. Bang, D. Chen, H. S. Chan,
W. Dai, A. Madotto, and P. Fung. Survey of Hallucination in Natural Language Generation.
ACM Computing Surveys, 55(12):1–38, Dec. 2023. ISSN 0360-0300, 1557-7341. doi:
10.1145/3571730. URL http://arxiv.org/abs/2202.03629.
36
[171]
H. Jiang, D. Beeferman, B. Roy, and D. Roy. CommunityLM: Probing partisan worldviews
from language models. In N. Calzolari, C.-R. Huang, H. Kim, J. Pustejovsky, L. Wanner, K.-S.
Choi, P.-M. Ryu, H.-H. Chen, L. Donatelli, H. Ji, S. Kurohashi, P. Paggio, N. Xue, S. Kim,
Y. Hahm, Z. He, T. K. Lee, E. Santus, F. Bond, and S.-H. Na, editors, Proceedings of the
29th International Conference on Computational Linguistics, pages 6818–6826, Gyeongju,
Republic of Korea, Oct. 2022. International Committee on Computational Linguistics. URL
https://aclanthology.org/2022.coling-1.593.
[172]
H. H. Jiang, L. Brown, J. Cheng, M. Khan, A. Gupta, D. Workman, A. Hanna, J. Flowers,
and T. Gebru. AI Art and its Impact on Artists. In Proceedings of the 2023 AAAI/ACM
Conference on AI, Ethics, and Society, AIES ’23, pages 363–374, New York, NY, USA,
Aug. 2023. Association for Computing Machinery. ISBN 9798400702310. doi: 10
.
1145/
3600211.3604681. URL https://dl.acm.org/doi/10.1145/3600211.3604681.
[173]
S. Jindal. Responsible Sourcing of Data Enrichment Services, June 2021. URL
https:
//partnershiponai.org/responsible-sourcing-considerations/.
[174]
P. Joshi, S. Santy, A. Budhiraja, K. Bali, and M. Choudhury. The State and Fate of Lin-
guistic Diversity and Inclusion in the NLP World. In Proceedings of the 58th Annual
Meeting of the Association for Computational Linguistics, pages 6282–6293, Online, 2020.
Association for Computational Linguistics. doi: 10
.
18653/v1/2020
.
acl-main
.
560. URL
https://www.aclweb.org/anthology/2020.acl-main.560.
[175]
L. H. Kaack, P. L. Donti, E. Strubell, G. Kamiya, F. Creutzig, and D. Rolnick. Aligning
artificial intelligence with climate change mitigation. Nature Climate Change, 12(6):518–
527, June 2022. ISSN 1758-678X, 1758-6798. doi: 10
.
1038/s41558-022-01377-7. URL
https://www.nature.com/articles/s41558-022-01377-7.
[176]
P. Kalluri. Don’t ask if artificial intelligence is good or fair, ask how it shifts power. Nature, 583
(7815):169–169, July 2020. ISSN 0028-0836, 1476-4687. doi: 10
.
1038/d41586-020-02003-2.
URL http://www.nature.com/articles/d41586-020-02003-2.
[177]
M. Kamal and W. J. Newman. Revenge Pornography: Mental Health Implications and Related
Legislation. The Journal of the American Academy of Psychiatry and the Law, 44(3):359–367,
Sept. 2016. ISSN 1943-3662.
[178]
R. Kamikubo, U. Dwivedi, and H. Kacorri. Sharing Practices for Datasets Related to Accessi-
bility and Aging. In Proceedings of the 23rd International ACM SIGACCESS Conference on
Computers and Accessibility, pages 1–16, Virtual Event USA, Oct. 2021. ACM. ISBN 978-1-
4503-8306-6. doi: 10
.
1145/3441852
.
3471208. URL
https://dl.acm.org/doi/10.1145/
3441852.3471208.
[179]
M. E. Kaminski. Regulating the Risks of AI. Boston University Law Review, 103(1347), 2022.
ISSN 1556-5068. doi: 10
.
2139/ssrn
.
4195066. URL
https://www.ssrn.com/abstract=
4195066.
[180]
M. Kan. US Lawmakers Push FTC to Crack Down on VPNs That Use Deceptive Practices. PC
Mag, 2022. URL
https://www.pcmag.com/news/us-lawmakers-push-ftc-to-crack-
down-on-vpns-that-use-deceptive-practices.
[181]
F. Karimi. ‘Mom, these bad men have me’: She believes scammers cloned her daughter’s
voice in a fake kidnapping, Apr. 2023. URL
https://www.cnn.com/2023/04/29/us/ai-
scam-calls-kidnapping-cec/index.html.
[182]
J. Kaur, R. A. Dara, C. Obimbo, F. Song, and K. Menard. A comprehensive keyword analysis
of online privacy policies. Information Security Journal: A Global Perspective, 27(5-6):
260–275, Nov. 2018. ISSN 1939-3555, 1939-3547. doi: 10
.
1080/19393555
.
2019
.
1606368.
URL https://www.tandfonline.com/doi/full/10.1080/19393555.2019.1606368.
[183]
D. Kaushik, Z. C. Lipton, and A. J. London. Resolving the Human-subjects Status of Machine
Learning’s Crowdworkers: What ethical framework should govern the interaction of ML
researchers and crowdworkers? Queue, 21(6):101–127, Dec. 2023. ISSN 1542-7730, 1542-
7749. doi: 10.1145/3639452. URL https://dl.acm.org/doi/10.1145/3639452.
37
[184]
K. Kelley. Good practice in the conduct and reporting of survey research. International Journal
for Quality in Health Care, 15(3):261–266, May 2003. ISSN 13534505, 14643677. doi:
10
.
1093/intqhc/mzg031. URL
https://academic.oup.com/intqhc/article-lookup/
doi/10.1093/intqhc/mzg031.
[185]
Z. Khan and Y. Fu. One Label, One Billion Faces: Usage and Consistency of Racial Categories
in Computer Vision. In Proceedings of the 2021 ACM Conference on Fairness, Accountability,
and Transparency, pages 587–597, Virtual Event Canada, Mar. 2021. ACM. ISBN 978-1-
4503-8309-7. doi: 10
.
1145/3442188
.
3445920. URL
https://dl.acm.org/doi/10.1145/
3442188.3445920.
[186]
H. Khlaaf, P. Mishkin, J. Achiam, G. Krueger, and M. Brundage. A Hazard Analysis Frame-
work for Code Synthesis Large Language Models, July 2022. URL
http://arxiv.org/abs/
2207.14157.
[187]
D. Kiela, M. Bartolo, Y. Nie, D. Kaushik, A. Geiger, Z. Wu, B. Vidgen, G. Prasad,
A. Singh, P. Ringshia, Z. Ma, T. Thrush, S. Riedel, Z. Talat, P. Stenetorp, R. Jia, M. Bansal,
C. Potts, and A. Williams. Dynabench: Rethinking Benchmarking in NLP. In Proceed-
ings of the 2021 Conference of the North American Chapter of the Association for Com-
putational Linguistics: Human Language Technologies, pages 4110–4124, Online, 2021.
Association for Computational Linguistics. doi: 10
.
18653/v1/2021
.
naacl-main
.
324. URL
https://www.aclweb.org/anthology/2021.naacl-main.324.
[188]
S. Kim, S. Yun, H. Lee, M. Gubri, S. Yoon, and S. J. Oh. ProPILE: Prob-
ing privacy leakage in large language models. In A. Oh, T. Neumann, A. Glober-
son, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Informa-
tion Processing Systems, volume 36, pages 20750–20762. Curran Associates, Inc.,
2023. URL
https://proceedings.neurips.cc/paper_files/paper/2023/file/
420678bb4c8251ab30e765bc27c3b047-Paper-Conference.pdf.
[189]
J. King and C. Meinhardt. Rethinking Privacy in the AI Era: Policy Provo-
cations for a Data-Centric World. Technical report, Stanford, 2024. URL
https://hai.stanford.edu/white-paper-rethinking-privacy-ai-era-policy-
provocations-data-centric-world.
[190]
J. Kirchenbauer, J. Geiping, Y. Wen, J. Katz, I. Miers, and T. Goldstein. A Watermark for
Large Language Models, June 2023. URL http://arxiv.org/abs/2301.10226.
[191]
K. Klyman. Acceptable use policies for foundation models: Considerations for policymakers
and developers, Apr. 2024. URL https://crfm.stanford.edu/2024/04/08/aups.html.
[192]
W.-Y. Ko, D. D’souza, K. Nguyen, R. Balestriero, and S. Hooker. FAIR-Ensemble: When
Fairness Naturally Emerges From Deep Ensembling, Dec. 2023. URL
http://arxiv.org/
abs/2303.00586.
[193]
A. Korinek and J. Vipra. Market concentration implications of foundation mod-
els: The Invisible Hand of ChatGPT. Brookings Institution, Sept. 2023. URL
https://policycommons.net/artifacts/4864844/market-concentration-
implications-of-foundation-models/5702045/.
[194]
S. M. Labott, T. P. Johnson, M. Fendrich, and N. C. Feeny. Emotional Risks to Respon-
dents in Survey Research: Some Empirical Evidence. Journal of Empirical Research on
Human Research Ethics, 8(4):53–66, Oct. 2013. ISSN 1556-2646, 1556-2654. doi: 10
.
1525/
jer.2013.8.4.53. URL http://journals.sagepub.com/doi/10.1525/jer.2013.8.4.53.
[195]
A. Lacoste, A. Luccioni, V. Schmidt, and T. Dandres. Quantifying the Carbon Emissions of
Machine Learning, Nov. 2019. URL http://arxiv.org/abs/1910.09700.
[196]
J. Lalor, Y. Yang, K. Smith, N. Forsgren, and A. Abbasi. Benchmarking Intersectional Biases in
NLP. In Proceedings of the 2022 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, pages 3598–3609, Seattle,
United States, 2022. Association for Computational Linguistics. doi: 10
.
18653/v1/2022
.
naacl-
main.263. URL https://aclanthology.org/2022.naacl-main.263.
38
[197]
J. Larson, S. Mattu, L. Kirchner, and J. Angwin. How We Analyzed the COMPAS
Recidivism Algorithm, 2016. URL
https://www.propublica.org/article/how-we-
analyzed-the-compas-recidivism-algorithm.
[198]
J.-U. Lee, H. Puerto, B. van Aken, Y. Arase, J. Z. Forde, L. Derczynski, A. Rücklé, I. Gurevych,
R. Schwartz, E. Strubell, and J. Dodge. Surveying (Dis)Parities and Concerns of Compute
Hungry NLP Research, Nov. 2023. URL http://arxiv.org/abs/2306.16900.
[199]
K. Lee, D. Ippolito, A. Nystrom, C. Zhang, D. Eck, C. Callison-Burch, and N. Carlini.
Deduplicating Training Data Makes Language Models Better. In Proceedings of the 60th
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),
pages 8424–8445, Dublin, Ireland, 2022. Association for Computational Linguistics. doi:
10.18653/v1/2022.acl-long.577. URL https://aclanthology.org/2022.acl-long.577.
[200]
J. Lepawsky. No insides on the outsides. Discard Studies, 0(0), Sept. 2019. URL
https:
//discardstudies.com/2019/09/23/no-insides-on-the-outsides/.
[201]
D. I. Levine. Automation as Part of the Solution. Journal of Management Inquiry, 28(3):
316–318, July 2019. ISSN 1056-4926. doi: 10
.
1177/1056492619827375. URL
https:
//doi.org/10.1177/1056492619827375.
[202]
B. Li, S. Haider, and C. Callison-Burch. This Land is {Your, My} Land: Evaluating Geopoliti-
cal Biases in Language Models, Apr. 2024. URL http://arxiv.org/abs/2305.14610.
[203]
C. Li. OpenAI’s GPT-3 Language Model: A Technical Overview, June 2020. URL
https:
//lambdalabs.com/blog/demystifying-gpt-3.
[204]
H. Li, D. Guo, W. Fan, M. Xu, J. Huang, F. Meng, and Y. Song. Multi-step Jailbreaking
Privacy Attacks on ChatGPT. In Findings of the Association for Computational Linguis-
tics: EMNLP 2023, pages 4138–4153, Singapore, 2023. Association for Computational Lin-
guistics. doi: 10
.
18653/v1/2023
.
findings-emnlp
.
272. URL
https://aclanthology.org/
2023.findings-emnlp.272.
[205]
N. Li, A. Pan, A. Gopal, S. Yue, D. Berrios, A. Gatti, J. D. Li, A.-K. Dombrowski, S. Goel,
L. Phan, G. Mukobi, N. Helm-Burger, R. Lababidi, L. Justen, A. B. Liu, M. Chen, I. Barrass,
O. Zhang, X. Zhu, R. Tamirisa, B. Bharathi, A. Khoja, Z. Zhao, A. Herbert-Voss, C. B.
Breuer, A. Zou, M. Mazeika, Z. Wang, P. Oswal, W. Liu, A. A. Hunt, J. Tienken-Harder, K. Y.
Shih, K. Talley, J. Guan, R. Kaplan, I. Steneker, D. Campbell, B. Jokubaitis, A. Levinson,
J. Wang, W. Qian, K. K. Karmakar, S. Basart, S. Fitz, M. Levine, P. Kumaraguru, U. Tupakula,
V. Varadharajan, Y. Shoshitaishvili, J. Ba, K. M. Esvelt, A. Wang, and D. Hendrycks. The
WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning, Mar. 2024.
URL http://arxiv.org/abs/2403.03218.
[206]
T. Li, D. Khashabi, T. Khot, A. Sabharwal, and V. Srikumar. UNQOVERing Stereotyping
Biases via Underspecified Questions. In T. Cohn, Y. He, and Y. Liu, editors, Findings of the
Association for Computational Linguistics: EMNLP 2020, pages 3475–3489, Online, Nov.
2020. Association for Computational Linguistics. doi: 10
.
18653/v1/2020
.
findings-emnlp
.
311.
URL https://aclanthology.org/2020.findings-emnlp.311.
[207]
W. Liang, M. Yuksekgonul, Y. Mao, E. Wu, and J. Zou. GPT detectors are biased
against non-native English writers. Patterns, 4(7):100779, July 2023. ISSN 26663899.
doi: 10
.
1016/j
.
patter
.
2023
.
100779. URL
https://linkinghub.elsevier.com/retrieve/
pii/S2666389923001307.
[208]
Q. V. Liao and Z. Xiao. Rethinking Model Evaluation as Narrowing the Socio-Technical Gap,
June 2023. URL http://arxiv.org/abs/2306.03100.
[209]
M. Liboiron. Pollution Is Colonialism. Duke University Press, Durham, 2021. ISBN 978-1-
4780-2144-5.
[210]
B. Liu, L. Wang, C. Lyu, Y. Zhang, J. Su, S. Shi, and Z. Tu. On the Cultural Gap in Text-to-
Image Generation, July 2023. URL http://arxiv.org/abs/2307.02971.
39
[211]
G. Liu, X. Wang, L. Yuan, Y. Chen, and H. Peng. Examining LLMs’ Uncertainty Expression
Towards Questions Outside Parametric Knowledge, Feb. 2024. URL
http://arxiv.org/
abs/2311.09731.
[212]
L. T. Liu, S. Dean, E. Rolf, M. Simchowitz, and M. Hardt. Delayed Impact of Fair Machine
Learning. Proceedings of the Twenty-Eighth International Joint Conference on Artificial
Intelligence, pages 6196–6200, 2019. URL
https://www.ijcai.org/proceedings/2019/
862.
[213]
Z. Liu, U. Iqbal, and N. Saxena. Opted Out, Yet Tracked: Are Regulations Enough to Protect
Your Privacy? Proceedings on Privacy Enhancing Technologies, 2024. ISSN 2299-0984. URL
https://petsymposium.org/popets/2024/popets-2024-0016.php.
[214]
Z. Liu, P. Schaldenbrand, B.-C. Okogwu, W. Peng, Y. Yun, A. Hundt, J. Kim, and J. Oh.
SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation, Jan. 2024. URL
http://arxiv.org/abs/2401.08053.
[215]
A. S. Luccioni, C. Akiki, M. Mitchell, and Y. Jernite. Stable Bias: Analyzing Societal Repre-
sentations in Diffusion Models, Nov. 2023. URL http://arxiv.org/abs/2303.11408.
[216]
A. S. Luccioni, S. Viguier, and A.-L. Ligozat. Estimating the Carbon Footprint of BLOOM,
a 176B Parameter Language Model. Journal of Machine Learning Research, 24(253):1–15,
2023. ISSN 1533-7928. URL http://jmlr.org/papers/v24/23-0069.html.
[217]
S. Luccioni, Y. Jernite, and M. Mitchell. Introducing the Data Measurements Tool: An
Interactive Tool for Looking at Datasets, 2021. URL
https://huggingface.co/blog/
data-measurements-tool.
[218]
S. Luccioni, Y. Jernite, and E. Strubell. Power hungry processing: Watts driving the cost of
ai deployment? In Proceedings of the 2024 ACM Conference on Fairness, Accountability,
and Transparency, FAccT ’24, page 85–99, New York, NY, USA, 2024. Association for
Computing Machinery. ISBN 9798400704505. doi: 10
.
1145/3630106
.
3658542. URL
https://doi.org/10.1145/3630106.3658542.
[219]
V. Malik, S. Dev, A. Nishi, N. Peng, and K.-W. Chang. Socially Aware Bias Measurements for
Hindi Language Representations. In Proceedings of the 2022 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies,
pages 1041–1052, Seattle, United States, 2022. Association for Computational Linguistics.
doi: 10
.
18653/v1/2022
.
naacl-main
.
76. URL
https://aclanthology.org/2022.naacl-
main.76.
[220]
G. Marcus and R. Southen. Generative AI Has a Visual Plagiarism Problem, 2024. URL
https://spectrum.ieee.org/midjourney-copyright.
[221]
V. Marian. Opinion | AI could cause a mass-extinction of languages — and ways of thinking.
Washington Post, Apr. 2023. ISSN 0190-8286. URL
https://www.washingtonpost.com/
opinions/2023/04/19/ai-chatgpt-language-extinction/.
[222]
A. Markdalen, M. Roberts, B. Schwartz, S. KVJ, M. Oost, S. Jones, L. Mitnick, S. Cherian,
R. Engels, R. Tolido, M. A. Sowa, R. Puttur, V. Perhirin, S. Andersson, and J. Buvat. Creative
and generative AI, Dec. 2023. URL
https://www.capgemini.com/insights/research-
library/creative-and-generative-ai/.
[223]
E. Markey. Markey, Heinrich, Eshoo, Beyer Introduce Legislation to Inves-
tigate, Measure Environmental Impacts of Artificial Intelligence, 2024. URL
https://www.markey.senate.gov/news/press-releases/markey-heinrich-eshoo-
beyer-introduce-legislation-to-investigate-measure-environmental-
impacts-of-artificial-intelligence.
[224]
K. Martin. The penalty for privacy violations: How privacy violations impact trust online.
Journal of Business Research, 82:103–116, Jan. 2018. ISSN 0148-2963. doi: 10
.
1016/
j
.
jbusres
.
2017
.
08
.
034. URL
https://www.sciencedirect.com/science/article/pii/
S0148296317302965.
40
[225]
C. May, A. Wang, S. Bordia, S. R. Bowman, and R. Rudinger. On Measuring Social Biases
in Sentence Encoders. In Proceedings of the 2019 Conference of the North, pages 622–628,
Minneapolis, Minnesota, 2019. Association for Computational Linguistics. doi: 10
.
18653/v1/
N19-1063. URL http://aclweb.org/anthology/N19-1063.
[226]
J. Mendelsohn, R. Le Bras, Y. Choi, and M. Sap. From Dogwhistles to Bullhorns: Unveiling
Coded Rhetoric with Language Models. In Proceedings of the 61st Annual Meeting of the
Association for Computational Linguistics (Volume 1: Long Papers), pages 15162–15180,
Toronto, Canada, 2023. Association for Computational Linguistics. doi: 10
.
18653/v1/2023
.
acl-
long.845. URL https://aclanthology.org/2023.acl-long.845.
[227]
M. Miceli, M. Schuessler, and T. Yang. Between Subjectivity and Imposition: Power Dynamics
in Data Annotation for Computer Vision. Proceedings of the ACM on Human-Computer
Interaction, 4(CSCW2):1–25, Oct. 2020. ISSN 2573-0142. doi: 10
.
1145/3415186. URL
https://dl.acm.org/doi/10.1145/3415186.
[228]
J. Mickel. Racial/Ethnic Categories in AI and Algorithmic Fairness: Why They Matter and
What They Represent, Apr. 2024. URL http://arxiv.org/abs/2404.06717.
[229]
Microsoft Threat Intelligence. Staying ahead of threat actors in the age of AI,
Feb. 2024. URL
https://www.microsoft.com/en-us/security/blog/2024/02/14/
staying-ahead-of-threat-actors-in-the-age-of-ai/.
[230]
Ministry of Foreign Affairs. REAIM 2023, 2023. URL
https://www.government.nl/
ministries/ministry-of-foreign-affairs/activiteiten/reaim.
[231]
N. Mireshghallah, H. Kim, X. Zhou, Y. Tsvetkov, M. Sap, R. Shokri, and Y. Choi. Can LLMs
Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity
Theory. In The Twelfth International Conference on Learning Representations, Oct. 2023.
URL https://openreview.net/forum?id=gmg7t8b4s0.
[232]
R. Modhvadia. How do people feel about AI? Technical report, Ada Lovelace Institute, 2023.
URL https://www.adalovelaceinstitute.org/report/public-attitudes-ai/.
[233]
S. Mohamed, M.-T. Png, and W. Isaac. Decolonial AI: Decolonial Theory as Sociotech-
nical Foresight in Artificial Intelligence. Philosophy & Technology, 33(4):659–684, Dec.
2020. ISSN 2210-5433, 2210-5441. doi: 10
.
1007/s13347-020-00405-8. URL
https:
//link.springer.com/10.1007/s13347-020-00405-8.
[234]
I. Mollas, Z. Chrysopoulou, S. Karlos, and G. Tsoumakas. ETHOS: An Online Hate Speech
Detection Dataset. Complex & Intelligent Systems, 8(6):4663–4678, Dec. 2022. ISSN
2199-4536, 2198-6053. doi: 10
.
1007/s40747-021-00608-2. URL
http://arxiv.org/abs/
2006.08328.
[235]
A. Monea. The Digital Closet: How the Internet Became Straight. The
MIT Press, Apr. 2022. ISBN 978-0-262-36913-8. doi: 10
.
7551/mitpress/
12551
.
001
.
0001. URL
https://direct.mit.edu/books/oa-monograph/5305/The-
Digital-ClosetHow-the-Internet-Became-Straight.
[236]
S. G. Monserrate. The Cloud Is Material: On the Environmental Impacts of Computa-
tion and Data Storage. MIT Case Studies in Social and Ethical Responsibilities of Com-
puting, (Winter 2022), Jan. 2022. doi: 10
.
21428/2c646de5
.
031d4553. URL
https:
//mit-serc.pubpub.org/pub/the-cloud-is-material/release/2.
[237]
B. Moore. Privacy: Studies in Social and Cultural History. Routledge Revivals. Routledge,
Abingdon, 2018. ISBN 978-1-351-69676-0.
[238]
C. Morris. AI chatbot will replace human helpline workers at National Eating Disorder Asso-
ciation. Fortune Well, 2023. URL
https://fortune.com/well/2023/05/26/national-
eating-disorder-association-ai-chatbot-tessa/.
[239]
M. Muro and S. Liu. The geography of AI. Technical report, Brookings, 2021. URL
https://www.brookings.edu/articles/the-geography-of-ai/.
41
[240]
M. Murphy. Predators Exploit AI Tools to Generate Images of Child Abuse.
Bloomberg.com, May 2023. URL
https://www.bloomberg.com/news/articles/2023-
05-23/predators-exploit-ai-tools-to-depict-abuse-prompting-warnings.
[241]
M. Nadeem, A. Bethke, and S. Reddy. StereoSet: Measuring stereotypical bias in pretrained
language models, Apr. 2020. URL http://arxiv.org/abs/2004.09456.
[242]
R. Naik and B. Nushi. Social Biases through the Text-to-Image Generation Lens. In
Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 786–
808, Montr\’{e}al QC Canada, Aug. 2023. ACM. ISBN 9798400702310. doi: 10
.
1145/
3600211.3604711. URL https://dl.acm.org/doi/10.1145/3600211.3604711.
[243]
N. Nangia, C. Vania, R. Bhalerao, and S. R. Bowman. CrowS-Pairs: A Challenge Dataset for
Measuring Social Biases in Masked Language Models. In B. Webber, T. Cohn, Y. He, and
Y. Liu, editors, Proceedings of the 2020 Conference on Empirical Methods in Natural Language
Processing (EMNLP), pages 1953–1967, Online, Nov. 2020. Association for Computational
Linguistics. doi: 10
.
18653/v1/2020
.
emnlp-main
.
154. URL
https://aclanthology.org/
2020.emnlp-main.154.
[244]
M. Nasr, N. Carlini, J. Hayase, M. Jagielski, A. F. Cooper, D. Ippolito, C. A. Choquette-Choo,
E. Wallace, F. Tramèr, and K. Lee. Scalable Extraction of Training Data from (Production)
Language Models, Nov. 2023. URL http://arxiv.org/abs/2311.17035.
[245]
National Institute of Standards and Technology. AI Risk Management Framework: AI RMF
(1.0). Technical Report error: NIST AI 100-1, National Institute of Standards and Technology,
Gaithersburg, MD, 2023. URL
https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-
1.pdf.
[246]
Neural Information Processing Systems. NeurIPS Code of Ethics, 2023. URL
https:
//neurips.cc/Conferences/2023/EthicsGuidelines.
[247]
J. Newman. A Taxonomy of Trustworthiness for Artificial Intelligence. Tech-
nical report, Center for Long-Term Cybersecurity, Berkeley, 2023. URL
https://cltc.berkeley.edu/publication/a-taxonomy-of-trustworthiness-
for-artificial-intelligence/.
[248]
D. Nikolaiev. Behind the Millions: Estimating the Scale of Large Language Models, June
2023. URL
https://towardsdatascience.com/behind-the-millions-estimating-
the-scale-of-large-language-models-97bd7287fb6b.
[249]
H. Nissenbaum. Privacy in Context: Technology, Policy, and the Integrity of Social Life.
Stanford University Press, Stanford, 2009. ISBN 978-0-8047-5236-7.
[250]
J. Niu, W. Tang, F. Xu, X. Zhou, and Y. Song. Global Research on Artificial Intelligence
from 1990–2014: Spatially-Explicit Bibliometric Analysis. ISPRS International Journal of
Geo-Information, 5(5):66, May 2016. ISSN 2220-9964. doi: 10
.
3390/ijgi5050066. URL
https://www.mdpi.com/2220-9964/5/5/66.
[251]
S. Noy and W. Zhang. Experimental evidence on the productivity effects of generative artificial
intelligence. Science, 381(6654):187–192, July 2023. doi: 10
.
1126/science
.
adh2586. URL
https://www.science.org/doi/10.1126/science.adh2586.
[252]
D. Nozza, F. Bianchi, and D. Hovy. HONEST: Measuring Hurtful Sentence Completion in
Language Models. In Proceedings of the 2021 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language Technologies, pages
2398–2406, Online, 2021. Association for Computational Linguistics. doi: 10
.
18653/v1/
2021.naacl-main.191. URL https://aclanthology.org/2021.naacl-main.191.
[253]
OECD Policy Observatory. OECD’s live repository of AI strategies & policies, 2021. URL
https://oecd.ai/en/dashboards.
[254]
OECD Policy Observatory. OECD Framework for the Classification of AI Systems: A tool for
effective AI policies, Apr. 2023. URL https://oecd.ai/en/classification.
42
[255]
K. Ogueji, O. Ahia, G. Onilude, S. Gehrmann, S. Hooker, and J. Kreutzer. Intriguing
Properties of Compression on Multilingual Models. In Y. Goldberg, Z. Kozareva, and
Y. Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural
Language Processing, pages 9092–9110, Abu Dhabi, United Arab Emirates, Dec. 2022. As-
sociation for Computational Linguistics. doi: 10
.
18653/v1/2022
.
emnlp-main
.
619. URL
https://aclanthology.org/2022.emnlp-main.619.
[256]
V. Ooi and G. Goh. Taxation of Automation and Artificial Intelligence as a Tool of Labour
Policy. SSRN Electronic Journal, 2018. ISSN 1556-5068. doi: 10
.
2139/ssrn
.
3322306. URL
https://www.ssrn.com/abstract=3322306.
[257] OpenAI. Pricing. URL https://openai.com/pricing.
[258]
OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida,
J. Altenschmidt, S. Altman, S. Anadkat, R. Avila, I. Babuschkin, S. Balaji, V. Balcom,
P. Baltescu, H. Bao, M. Bavarian, J. Belgum, I. Bello, J. Berdine, G. Bernadett-Shapiro,
C. Berner, L. Bogdonoff, O. Boiko, M. Boyd, A.-L. Brakman, G. Brockman, T. Brooks,
M. Brundage, K. Button, T. Cai, R. Campbell, A. Cann, B. Carey, C. Carlson, R. Carmichael,
B. Chan, C. Chang, F. Chantzis, D. Chen, S. Chen, R. Chen, J. Chen, M. Chen, B. Chess,
C. Cho, C. Chu, H. W. Chung, D. Cummings, J. Currier, Y. Dai, C. Decareaux, T. Degry,
N. Deutsch, D. Deville, A. Dhar, D. Dohan, S. Dowling, S. Dunning, A. Ecoffet, A. Eleti,
T.Eloundou, D. Farhi, L. Fedus, N. Felix, S. P. Fishman, J. Forte, I. Fulford, L. Gao, E. Georges,
C. Gibson, V. Goel, T. Gogineni, G. Goh, R. Gontijo-Lopes, J. Gordon, M. Grafstein, S. Gray,
R. Greene, J. Gross, S. S. Gu, Y. Guo, C. Hallacy, J. Han, J. Harris, Y. He, M. Heaton,
J. Heidecke, C. Hesse, A. Hickey, W. Hickey, P. Hoeschele, B. Houghton, K. Hsu, S. Hu,
X. Hu, J. Huizinga, S. Jain, S. Jain, J. Jang, A. Jiang, R. Jiang, H. Jin, D. Jin, S. Jomoto,
B. Jonn, H. Jun, T. Kaftan, Ł. Kaiser, A. Kamali, I. Kanitscheider, N. S. Keskar, T. Khan,
L. Kilpatrick, J. W. Kim, C. Kim, Y. Kim, J. H. Kirchner, J. Kiros, M. Knight, D. Kokotajlo,
Ł. Kondraciuk, A. Kondrich, A. Konstantinidis, K. Kosic, G. Krueger, V. Kuo, M. Lampe,
I. Lan, T. Lee, J. Leike, J. Leung, D. Levy, C. M. Li, R. Lim, M. Lin, S. Lin, M. Litwin,
T. Lopez, R. Lowe, P. Lue, A. Makanju, K. Malfacini, S. Manning, T. Markov, Y. Markovski,
B. Martin, K. Mayer, A. Mayne, B. McGrew, S. M. McKinney, C. McLeavey, P. McMillan,
J. McNeil, D. Medina, A. Mehta, J. Menick, L. Metz, A. Mishchenko, P. Mishkin, V. Monaco,
E. Morikawa, D. Mossing, T. Mu, M. Murati, O. Murk, D. Mély, A. Nair, R. Nakano,
R. Nayak, A. Neelakantan, R. Ngo, H. Noh, L. Ouyang, C. O’Keefe, J. Pachocki, A. Paino,
J. Palermo, A. Pantuliano, G. Parascandolo, J. Parish, E. Parparita, A. Passos, M. Pavlov,
A. Peng, A. Perelman, F. d. A. B. Peres, M. Petrov, H. P. d. O. Pinto, Michael, Pokorny,
M. Pokrass, V. H. Pong, T. Powell, A. Power, B. Power, E. Proehl, R. Puri, A. Radford, J. Rae,
A. Ramesh, C. Raymond, F. Real, K. Rimbach, C. Ross, B. Rotsted, H. Roussez, N. Ryder,
M. Saltarelli, T. Sanders, S. Santurkar, G. Sastry, H. Schmidt, D. Schnurr, J. Schulman,
D. Selsam, K. Sheppard, T. Sherbakov, J. Shieh, S. Shoker, P. Shyam, S. Sidor, E. Sigler,
M. Simens, J. Sitkin, K. Slama, I. Sohl, B. Sokolowsky, Y. Song, N. Staudacher, F. P. Such,
N. Summers, I. Sutskever, J. Tang, N. Tezak, M. B. Thompson, P. Tillet, A. Tootoonchian,
E. Tseng, P. Tuggle, N. Turley, J. Tworek, J. F. C. Uribe, A. Vallone, A. Vijayvergiya, C. Voss,
C. Wainwright, J. J. Wang, A. Wang, B. Wang, J. Ward, J. Wei, C. J. Weinmann, A. Welihinda,
P. Welinder, J. Weng, L. Weng, M. Wiethoff, D. Willner, C. Winter, S. Wolrich, H. Wong,
L. Workman, S. Wu, J. Wu, M. Wu, K. Xiao, T. Xu, S. Yoo, K. Yu, Q. Yuan, W. Zaremba,
R. Zellers, C. Zhang, M. Zhang, S. Zhao, T. Zheng, J. Zhuang, W. Zhuk, and B. Zoph. GPT-4
Technical Report, Mar. 2024. URL http://arxiv.org/abs/2303.08774.
[259]
Organizers of QueerInAI and A. Subramonian. NAIAC Briefing, 2023. URL
https://
www.queerinai.com/naiac-briefing.
[260]
A. Ovalle, P. Goyal, J. Dhamala, Z. Jaggers, K.-W. Chang, A. Galstyan, R. Zemel, and
R. Gupta. “I’m fully who I am”: Towards Centering Transgender and Non-Binary Voices
to Measure Biases in Open Language Generation. In 2023 ACM Conference on Fairness,
Accountability, and Transparency, pages 1246–1266, Chicago IL USA, June 2023. ACM.
ISBN 9798400701924. doi: 10
.
1145/3593013
.
3594078. URL
https://dl.acm.org/doi/
10.1145/3593013.3594078.
43
[261]
A. Ovalle, A. Subramonian, V. Gautam, G. Gee, and K.-W. Chang. Factoring the Matrix
of Domination: A Critical Review and Reimagination of Intersectionality in AI Fairness.
In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pages 496–
511, Montr\’{e}al QC Canada, Aug. 2023. ACM. ISBN 9798400702310. doi: 10
.
1145/
3600211.3604705. URL https://dl.acm.org/doi/10.1145/3600211.3604705.
[262]
R. Parasuraman and V. Riley. Humans and Automation: Use, Misuse, Disuse, Abuse. Human
Factors, 39(2):230–253, June 1997. ISSN 0018-7208. doi: 10
.
1518/001872097778543886.
URL https://doi.org/10.1518/001872097778543886.
[263]
P. S. Park, S. Goldstein, A. O’Gara, M. Chen, and D. Hendrycks. AI Deception: A Survey
of Examples, Risks, and Potential Solutions, Aug. 2023. URL
http://arxiv.org/abs/
2308.14752.
[264]
A. Parrish, A. Chen, N. Nangia, V. Padmakumar, J. Phang, J. Thompson, P. M. Htut, and
S. Bowman. BBQ: A hand-built bias benchmark for question answering. In S. Mure-
san, P. Nakov, and A. Villavicencio, editors, Findings of the Association for Compu-
tational Linguistics: ACL 2022, pages 2086–2105, Dublin, Ireland, May 2022. Asso-
ciation for Computational Linguistics. doi: 10
.
18653/v1/2022
.
findings-acl
.
165. URL
https://aclanthology.org/2022.findings-acl.165.
[265]
J. Paschen. Investigating the emotional appeal of fake news using artificial intelligence
and human contributions. Journal of Product & Brand Management, 29(2):223–233, May
2019. ISSN 1061-0421, 1061-0421. doi: 10
.
1108/JPBM-12-2018-2179. URL
https://
www.emerald.com/insight/content/doi/10.1108/JPBM-12-2018-2179/full/html.
[266]
S. Passi and M. Vorvoreanu. Overreliance on AI: Literature Review. Technical report,
Microsoft Corporation, June 2022. URL
https://www.microsoft.com/en-us/research/
publication/overreliance-on-ai-literature-review/.
[267]
G. Pennycook, T. D. Cannon, and D. G. Rand. Prior exposure increases perceived accuracy of
fake news. Journal of Experimental Psychology: General, 147(12):1865–1880, Dec. 2018.
ISSN 1939-2222, 0096-3445. doi: 10
.
1037/xge0000465. URL
https://doi.apa.org/doi/
10.1037/xge0000465.
[268]
G. Pennycook, Z. Epstein, M. Mosleh, A. A. Arechar, D. Eckles, and D. G. Rand. Shifting
attention to accuracy can reduce misinformation online. Nature, 592(7855):590–595, Apr.
2021. ISSN 0028-0836, 1476-4687. doi: 10
.
1038/s41586-021-03344-2. URL
https:
//www.nature.com/articles/s41586-021-03344-2.
[269]
B. Perrigo. Inside Facebook’s African Sweatshop. Time Magazine, 2022. URL
https://time.com/6147458/facebook-africa-content-moderation-employee-
treatment/.
[270] Perspective API. Perspective API. URL https://www.perspectiveapi.com/#/home.
[271]
M. Phuong, M. Aitchison, E. Catt, S. Cogan, A. Kaskasoli, V. Krakovna, D. Lindner, M. Rahtz,
Y. Assael, S. Hodkinson, H. Howard, T. Lieberum, R. Kumar, M. A. Raad, A. Webson, L. Ho,
S. Lin, S. Farquhar, M. Hutter, G. Deletang, A. Ruoss, S. El-Sayed, S. Brown, A. Dragan,
R. Shah, A. Dafoe, and T. Shevlane. Evaluating Frontier Models for Dangerous Capabilities,
Apr. 2024. URL http://arxiv.org/abs/2403.13793.
[272]
V. W. Polonski. AI trust and AI fears: A media debate that could divide society,
2018. URL
https://www.oii.ox.ac.uk/news-events/ai-trust-and-ai-fears-a-
media-debate-that-could-divide-society.
[273]
B. Potts. Frontiers of multimodal learning: A responsible AI approach, Sept.
2023. URL
https://www.microsoft.com/en-us/research/blog/frontiers-of-
multimodal-learning-a-responsible-ai-approach/.
[274]
L. Pozzobon, B. Ermis, P. Lewis, and S. Hooker. On the Challenges of Using Black-Box APIs
for Toxicity Evaluation in Research, Apr. 2023. URL
http://arxiv.org/abs/2304.12397
.
44
[275]
V. Prabhakaran, Z. Waseem, S. Akiwowo, and B. Vidgen. Online Abuse and Human Rights:
WOAH Satellite Session at RightsCon 2020. In Proceedings of the Fourth Workshop on Online
Abuse and Harms, pages 1–6, Online, 2020. Association for Computational Linguistics. doi:
10.18653/v1/2020.alw-1.1. URL https://www.aclweb.org/anthology/2020.alw- 1.1.
[276]
V. Prabhakaran, A. Mostafazadeh Davani, and M. Diaz. On Releasing Annotator-Level
Labels and Information in Datasets. In Proceedings of The Joint 15th Linguistic Annotation
Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 133–
138, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. doi:
10.18653/v1/2021.law-1.14. URL https://aclanthology.org/2021.law- 1.14.
[277]
A. Priyanshu, S. Vijay, A. Kumar, R. Naidu, and F. Mireshghallah. Are Chatbots Ready for
Privacy-Sensitive Applications? An Investigation into Input Regurgitation and Prompt-Induced
Sanitization, May 2023. URL http://arxiv.org/abs/2305.15008.
[278]
R. Qadri, R. Shelby, C. L. Bennett, and E. Denton. AI’s Regimes of Representation: A
Community-centered Study of Text-to-Image Models in South Asia. In 2023 ACM Conference
on Fairness, Accountability, and Transparency, pages 506–517, Chicago IL USA, June 2023.
ACM. ISBN 9798400701924. doi: 10
.
1145/3593013
.
3594016. URL
https://dl.acm.org/
doi/10.1145/3593013.3594016.
[279]
S. Quach, P. Thaichon, K. D. Martin, S. Weaven, and R. W. Palmatier. Digital technologies:
Tensions in privacy and data. Journal of the Academy of Marketing Science, 50(6):1299–
1323, Nov. 2022. ISSN 1552-7824. doi: 10
.
1007/s11747-022-00845-y. URL
https:
//doi.org/10.1007/s11747-022-00845-y.
[280]
O. O. Queerinai, A. Ovalle, A. Subramonian, A. Singh, C. Voelcker, D. J. Sutherland, D. Lo-
catelli, E. Breznik, F. Klubicka, H. Yuan, H. J, H. Zhang, J. Shriram, K. Lehman, L. Soldaini,
M. Sap, M. P. Deisenroth, M. L. Pacheco, M. Ryskina, M. Mundt, M. Agarwal, N. Mclean,
P. Xu, A. Pranav, R. Korpan, R. Ray, S. Mathew, S. Arora, S. John, T. Anand, V. Agrawal,
W. Agnew, Y. Long, Z. J. Wang, Z. Talat, A. Ghosh, N. Dennler, M. Noseworthy, S. Jha,
E. Baylor, A. Joshi, N. Y. Bilenko, A. Mcnamara, R. Gontijo-Lopes, A. Markham, E. Dong,
J. Kay, M. Saraswat, N. Vytla, and L. Stark. Queer In AI: A Case Study in Community-
Led Participatory AI. In 2023 ACM Conference on Fairness, Accountability, and Trans-
parency, pages 1882–1895, Chicago IL USA, June 2023. ACM. ISBN 9798400701924. doi:
10.1145/3593013.3594134. URL https://dl.acm.org/doi/10.1145/3593013.3594134.
[281]
I. D. Raji, A. Smart, R. N. White, M. Mitchell, T. Gebru, B. Hutchinson, J. Smith-
Loud, D. Theron, and P. Barnes. Closing the AI accountability gap: Defining an end-
to-end framework for internal algorithmic auditing. In Proceedings of the 2020 Con-
ference on Fairness, Accountability, and Transparency, pages 33–44, Barcelona Spain,
Jan. 2020. ACM. ISBN 978-1-4503-6936-7. doi: 10
.
1145/3351095
.
3372873. URL
https://dl.acm.org/doi/10.1145/3351095.3372873.
[282]
I. D. Raji, E. M. Bender, A. Paullada, E. Denton, and A. Hanna. AI and the Ev-
erything in the Whole Wide World Benchmark. In Proceedings of the Neural Infor-
mation Processing Systems Track on Datasets and Benchmarks. Curran, 2021. URL
https://datasets-benchmarks-proceedings.neurips.cc/paper_files/paper/
2021/file/084b6fbb10729ed4da8c3d3f5a3ae7c9-Paper-round2.pdf.
[283]
A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever.
Zero-shot text-to-image generation. In M. Meila and T. Zhang, editors, Proceedings of
the 38th International Conference on Machine Learning, volume 139 of Proceedings of
Machine Learning Research, pages 8821–8831. PMLR, 2021-07-18/2021-07-24. URL
https:
//proceedings.mlr.press/v139/ramesh21a.html.
[284]
P. Rao and M. Taboada. Gender Bias in the News: A Scalable Topic Modelling and Visu-
alization Framework. Frontiers in Artificial Intelligence, 4:664737, June 2021. ISSN 2624-
8212. doi: 10
.
3389/frai
.
2021
.
664737. URL
https://www.frontiersin.org/articles/
10.3389/frai.2021.664737/full.
45
[285]
A. Rapp, L. Curti, and A. Boldi. The human side of human-chatbot interaction: A systematic
literature review of ten years of research on text-based chatbots. International Journal
of Human-Computer Studies, 151:102630, July 2021. ISSN 1071-5819. doi: 10
.
1016/
j
.
ijhcs
.
2021
.
102630. URL
https://www.sciencedirect.com/science/article/pii/
S1071581921000483.
[286]
M. Rauh, J. Mellor, J. Uesato, P.-S. Huang, J. Welbl, L. Weidinger, S. Dathathri, A. Glaese,
G. Irving, I. Gabriel, W. Isaac, and L. A. Hendricks. Characteristics of harmful text: Towards
rigorous benchmarking of language models. In Proceedings of the 36th International Con-
ference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY, USA, 2024.
Curran Associates Inc. ISBN 978-1-71387-108-8.
[287]
T. Ray. Common but Different Futures: AI Inequity and Climate Change, 2021.
URL
https://www.orfonline.org/research/common-but-different-futures-ai-
inequity-and-climate-change.
[288]
S. Read and I. Shine. You’ve probably heard of Scope 1, 2 and 3 emissions, but what are Scope
4 emissions?, Sept. 2022. URL
https://www.weforum.org/agenda/2022/09/scope-4-
emissions-climate-greenhouse-business/.
[289]
D. Reisman, J. Schultz, K. Crawford, and M. Whittaker. Algorithmic Impact Assessments
Report: A Practical Framework for Public Agency Accountability, Apr. 2018. URL
https://
ainowinstitute.org/publication/algorithmic-impact-assessments-report-2.
[290]
N. Rekabsaz, R. West, J. Henderson, and A. Hanbury. Measuring Societal Biases from Text
Corpora with Smoothed First-Order Co-occurrence. Proceedings of the International AAAI
Conference on Web and Social Media, 15:549–560, May 2021. ISSN 2334-0770, 2162-
3449. doi: 10
.
1609/icwsm
.
v15i1
.
18083. URL
https://ojs.aaai.org/index.php/ICWSM/
article/view/18083.
[291]
Y. D. D.-N.-. Rep. Clarke. Algorithmic Accountability Act of 2022, Feb. 2022. URL
https:
//www.congress.gov/bill/117th-congress/house-bill/6580/text.
[292]
Republic of Korea. Input by the Government of the Republic of Korea on the Themes of
an Expert Consultation on the Practical Application of the United Nations Guiding Prin-
ciples on Business and Human Rights to the Activities of Technology Companies. Tech-
nical report, 2022. URL
https://www.ohchr.org/sites/default/files/2022-03/
RepublicofKorea.pdf.
[293]
J. Ricker, S. Damm, T. Holz, and A. Fischer. Towards the Detection of Diffusion Model
Deepfakes:. In Proceedings of the 19th International Joint Conference on Computer Vision,
Imaging and Computer Graphics Theory and Applications, pages 446–457, Rome, Italy,
2024. SCITEPRESS - Science and Technology Publications. ISBN 978-989-758-679-8. doi:
10
.
5220/0012422000003660. URL
https://www.scitepress.org/DigitalLibrary/
Link.aspx?doi=10.5220/0012422000003660.
[294]
R. Righi, S. Samoili, M. López Cobo, M. Vázquez-Prada Baillet, M. Cardona, and G. De Prato.
The AI techno-economic complex System: Worldwide landscape, thematic subdomains and
technological collaborations. Telecommunications Policy, 44(6):101943, July 2020. ISSN
0308-5961. doi: 10
.
1016/j
.
telpol
.
2020
.
101943. URL
https://www.sciencedirect.com/
science/article/pii/S0308596120300355.
[295]
S. T. Roberts. Behind the Screen: Content Moderation in the Shadows of Social Media. Yale
University Press, New Haven, 2019. ISBN 978-0-300-23588-3.
[296]
R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image
synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, June 2022.
[297]
J. Sablosky. “Dangerous organizations: Facebook’s content moderation decisions and
ethnic visibility in Myanmar”. Media, Culture & Society, 43(6):1017–1042, Sept. 2021.
ISSN 0163-4437. doi: 10
.
1177/0163443720987751. URL
https://doi.org/10.1177/
0163443720987751.
46
[298]
V. S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi. Can AI-Generated
Text be Reliably Detected?, Feb. 2024. URL http://arxiv.org/abs/2303.11156.
[299]
S. Safavi, H. Gan, I. Mporas, and R. Sotudeh. Fraud Detection in Voice-Based Iden-
tity Authentication Applications and Services. In 2016 IEEE 16th International Confer-
ence on Data Mining Workshops (ICDMW), pages 1074–1081, Barcelona, Spain, Dec.
2016. IEEE. ISBN 978-1-5090-5910-2. doi: 10
.
1109/ICDMW
.
2016
.
0155. URL
http:
//ieeexplore.ieee.org/document/7836786/.
[300]
C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan,
S. S. Mahdavi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, and M. Norouzi. Photorealistic
Text-to-Image Diffusion Models with Deep Language Understanding, May 2022. URL
http://arxiv.org/abs/2205.11487.
[301]
N. Sambasivan, E. Arnesen, B. Hutchinson, T. Doshi, and V. Prabhakaran. Re-imagining
Algorithmic Fairness in India and Beyond. In Proceedings of the 2021 ACM Conference
on Fairness, Accountability, and Transparency, pages 315–328, Virtual Event Canada, Mar.
2021. ACM. ISBN 978-1-4503-8309-7. doi: 10
.
1145/3442188
.
3445896. URL
https:
//dl.acm.org/doi/10.1145/3442188.3445896.
[302]
S. Santurkar, E. Durmus, F. Ladhak, C. Lee, P. Liang, and T. Hashimoto. Whose Opinions
Do Language Models Reflect? 2023. doi: 10
.
48550/ARXIV
.
2303
.
17548. URL
https:
//arxiv.org/abs/2303.17548.
[303]
M. Sap, D. Card, S. Gabriel, Y. Choi, and N. A. Smith. The Risk of Racial Bias in Hate Speech
Detection. In A. Korhonen, D. Traum, and L. Màrquez, editors, Proceedings of the 57th
Annual Meeting of the Association for Computational Linguistics, pages 1668–1678, Florence,
Italy, July 2019. Association for Computational Linguistics. doi: 10
.
18653/v1/P19-1163. URL
https://aclanthology.org/P19-1163.
[304]
M. Sap, S. Swayamdipta, L. Vianna, X. Zhou, Y. Choi, and N. A. Smith. Annotators
with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection. In
M. Carpuat, M.-C. de Marneffe, and I. V. Meza Ruiz, editors, Proceedings of the 2022
Conference of the North American Chapter of the Association for Computational Linguis-
tics: Human Language Technologies, pages 5884–5906, Seattle, United States, July 2022.
Association for Computational Linguistics. doi: 10
.
18653/v1/2022
.
naacl-main
.
431. URL
https://aclanthology.org/2022.naacl-main.431.
[305]
M. K. Scheuerman, A. Hanna, and E. Denton. Do Datasets Have Politics? Disciplinary Values
in Computer Vision Dataset Development. Proceedings of the ACM on Human-Computer
Interaction, 5(CSCW2):1–37, Oct. 2021. ISSN 2573-0142. doi: 10
.
1145/3476058. URL
https://dl.acm.org/doi/10.1145/3476058.
[306]
K. Schoenberg. History Can Help Us Chart AI’s Future, Jan. 2024. URL
https:
//issues.org/ai-history-future-li/.
[307] P. Schramowski, M. Brack, B. Deiseroth, and K. Kersting. Safe Latent Diffusion: Mitigating
Inappropriate Degeneration in Diffusion Models, Apr. 2023. URL
http://arxiv.org/abs/
2211.05105.
[308]
R. Schwartz, A. Vassilev, K. Greene, L. Perine, A. Burt, and P. Hall. Towards a standard
for identifying and managing bias in artificial intelligence. Technical Report NIST SP 1270,
National Institute of Standards and Technology (U.S.), Gaithersburg, MD, Mar. 2022. URL
https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf.
[309]
A. See, A. Pappu, R. Saxena, A. Yerukola, and C. D. Manning. Do Massively Pretrained
Language Models Make Better Storytellers? In M. Bansal and A. Villavicencio, editors,
Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL),
pages 843–861, Hong Kong, China, Nov. 2019. Association for Computational Linguistics.
doi: 10.18653/v1/K19-1079. URL https://aclanthology.org/K19-1079.
[310]
S. Shafayat, E. Kim, J. Oh, and A. Oh. Multi-FAct: Assessing Multilingual LLMs’
Multi-Regional Knowledge using FActScore, Mar. 2024. URL
http://arxiv.org/abs/
2402.18045.
47
[311]
S. Shankar, Y. Halpern, E. Breck, J. Atwood, J. Wilson, and D. Sculley. No Classification
without Representation: Assessing Geodiversity Issues in Open Data Sets for the Developing
World, Nov. 2017. URL http://arxiv.org/abs/1711.08536.
[312]
O. Sharir, B. Peleg, and Y. Shoham. The Cost of Training NLP Models: A Concise Overview,
Apr. 2020. URL http://arxiv.org/abs/2004.08900.
[313]
R. Shelby, S. Rismani, K. Henne, Aj. Moon, N. Rostamzadeh, P. Nicholas, N. Yilla, J. Gallegos,
A. Smart, E. Garcia, and G. Virk. Sociotechnical Harms of Algorithmic Systems: Scoping a
Taxonomy for Harm Reduction, July 2023. URL http://arxiv.org/abs/2210.05791.
[314]
S. Shen, M. Zhang, W. Chen, A. Bialkowski, and M. Xu. Words Can Be Confusing: Stereotype
Bias Removal in Text Classification at the Word Level. In H. Kashima, T. Ide, and W.-C. Peng,
editors, Advances in Knowledge Discovery and Data Mining, pages 99–111, Cham, 2023.
Springer Nature Switzerland. ISBN 978-3-031-33383-5. doi: 10
.
1007/978-3-031-33383-5
_
8.
[315]
J. Shi, Y. Liu, P. Zhou, and L. Sun. BadGPT: Exploring Security Vulnerabilities of ChatGPT via
Backdoor Attacks to InstructGPT, Feb. 2023. URL http://arxiv.org/abs/2304.12298.
[316]
R. J. Shiller. Narrative Economics: How Stories Go Viral & Drive Major Economic Events.
Princeton University Press, Princeton, 2019. ISBN 978-0-691-18229-2.
[317]
K. Shilton, E. Moss, S. A. Gilbert, M. J. Bietz, C. Fiesler, J. Metcalf, J. Vitak, and M. Zimmer.
Excavating awareness and power in data science: A manifesto for trustworthy pervasive data
research. Big Data & Society, 8(2):20539517211040759, July 2021. ISSN 2053-9517. doi:
10.1177/20539517211040759. URL https://doi.org/10.1177/20539517211040759.
[318]
J. Shin and S. Chan-Olmsted. User perceptions and trust of explainable machine learning fake
news detectors. International Journal of Communication, 17(0), 2022. ISSN 1932-8036. URL
https://ijoc.org/index.php/ijoc/article/view/19534.
[319]
G. Simmons. Moral Mimicry: Large Language Models Produce Moral Rationalizations
Tailored to Political Identity. In Proceedings of the 61st Annual Meeting of the Association for
Computational Linguistics (Volume 4: Student Research Workshop), pages 282–297, Toronto,
Canada, 2023. Association for Computational Linguistics. doi: 10
.
18653/v1/2023
.
acl-srw
.
40.
URL https://aclanthology.org/2023.acl-srw.40.
[320]
A. Simpson. On Ethnographic Refusal: Indigeneity, ‘Voice’ and Colonial Citizenship.
Junctures: The Journal for Thematic Dialogue, (9), 2007. ISSN 1179-8912. URL
https://junctures.org/index.php/junctures/article/view/66.
[321]
I. Solaiman. The Gradient of Generative AI Release: Methods and Considerations. In 2023
ACM Conference on Fairness, Accountability, and Transparency, pages 111–122, Chicago
IL USA, June 2023. ACM. ISBN 9798400701924. doi: 10
.
1145/3593013
.
3593981. URL
https://dl.acm.org/doi/10.1145/3593013.3593981.
[322]
I. Solaiman and C. Dennison. Process for Adapting Language Models to Society (PALMS) with
Values-Targeted Datasets. In Advances in Neural Information Processing Systems, volume 34,
pages 5861–5873. Curran Associates, Inc., 2021. URL
https://proceedings.neurips.cc/
paper/2021/hash/2e855f9489df0712b4bd8ea9e2848c5a-Abstract.html.
[323]
I. Solaiman, M. Brundage, J. Clark, A. Askell, A. Herbert-Voss, J. Wu, A. Radford, G. Krueger,
J. W. Kim, S. Kreps, M. McCain, A. Newhouse, J. Blazakis, K. McGuffie, and J. Wang.
Release Strategies and the Social Impacts of Language Models, Nov. 2019. URL
http:
//arxiv.org/abs/1908.09203.
[324]
D. Solove and D. K. Citron. Privacy Harms. GW Law Faculty Publications & Other Works,
Jan. 2021. URL https://scholarship.law.gwu.edu/faculty_publications/1534.
[325]
D. J. Solove. A Taxonomy of Privacy. University of Pennsylvania Law Review, 154(3):
477–564, 2006. ISSN 0041-9907. doi: 10
.
2307/40041279. URL
https://www.jstor.org/
stable/40041279.
[326] Spawning AI. Have I Been Trained. URL https://haveibeentrained.com/.
48
[327]
G. Spitale, N. Biller-Andorno, and F. Germani. AI model GPT-3 (dis)informs us better than
humans. Science Advances, 9(26):eadh1850, June 2023. ISSN 2375-2548. doi: 10
.
1126/
sciadv.adh1850. URL https://www.science.org/doi/10.1126/sciadv.adh1850.
[328]
A. Srivastava, A. Rastogi, A. Rao, A. A. M. Shoeb, A. Abid, A. Fisch, A. R. Brown, A. Santoro,
A. Gupta, A. Garriga-Alonso, A. Kluska, A. Lewkowycz, A. Agarwal, A. Power, A. Ray,
A. Warstadt, A. W. Kocurek, A. Safaya, A. Tazarv, A. Xiang, A. Parrish, A. Nie, A. Hussain,
A. Askell, A. Dsouza, A. Slone, A. Rahane, A. S. Iyer, A. Andreassen, A. Madotto, A. Santilli,
A. Stuhlmüller, A. Dai, A. La, A. Lampinen, A. Zou, A. Jiang, A. Chen, A. Vuong, A. Gupta,
A. Gottardi, A. Norelli, A. Venkatesh, A. Gholamidavoodi, A. Tabassum, A. Menezes,
A. Kirubarajan, A. Mullokandov, A. Sabharwal, A. Herrick, A. Efrat, A. Erdem, A. Karaka¸s,
B. R. Roberts, B. S. Loe, B. Zoph, B. Bojanowski, B. Özyurt, B. Hedayatnia, B. Neyshabur,
B. Inden, B. Stein, B. Ekmekci, B. Y. Lin, B. Howald, B. Orinion, C. Diao, C. Dour, C. Stinson,
C. Argueta, C. F. Ramírez, C. Singh, C. Rathkopf, C. Meng, C. Baral, C. Wu, C. Callison-Burch,
C. Waites, C. Voigt, C. D. Manning, C. Potts, C. Ramirez, C. E. Rivera, C. Siro, C. Raffel,
C. Ashcraft, C. Garbacea, D. Sileo, D. Garrette, D. Hendrycks, D. Kilman, D. Roth, D. Free-
man, D. Khashabi, D. Levy, D. M. González, D. Perszyk, D. Hernandez, D. Chen, D. Ippolito,
D. Gilboa, D. Dohan, D. Drakard, D. Jurgens, D. Datta, D. Ganguli, D. Emelin, D. Kleyko,
D. Yuret, D. Chen, D. Tam, D. Hupkes, D. Misra, D. Buzan, D. C. Mollo, D. Yang, D.-H.
Lee, D. Schrader, E. Shutova, E. D. Cubuk, E. Segal, E. Hagerman, E. Barnes, E. Donoway,
E. Pavlick, E. Rodola, E. Lam, E. Chu, E. Tang, E. Erdem, E. Chang, E. A. Chi, E. Dyer,
E. Jerzak, E. Kim, E. E. Manyasi, E. Zheltonozhskii, F. Xia, F. Siar, F. Martínez-Plumed,
F. Happé, F. Chollet, F. Rong, G. Mishra, G. I. Winata, G. de Melo, G. Kruszewski, G. Parascan-
dolo, G. Mariani, G. Wang, G. Jaimovitch-López, G. Betz, G. Gur-Ari, H. Galijasevic, H. Kim,
H. Rashkin, H. Hajishirzi, H. Mehta, H. Bogar, H. Shevlin, H. Schütze, H. Yakura, H. Zhang,
H. M. Wong, I. Ng, I. Noble, J. Jumelet, J. Geissinger, J. Kernion, J. Hilton, J. Lee, J. F. Fisac,
J. B. Simon, J. Koppel, J. Zheng, J. Zou, J. Koco´
n, J. Thompson, J. Wingfield, J. Kaplan,
J. Radom, J. Sohl-Dickstein, J. Phang, J. Wei, J. Yosinski, J. Novikova, J. Bosscher, J. Marsh,
J. Kim, J. Taal, J. Engel, J. Alabi, J. Xu, J. Song, J. Tang, J. Waweru, J. Burden, J. Miller,
J. U. Balis, J. Batchelder, J. Berant, J. Frohberg, J. Rozen, J. Hernandez-Orallo, J. Boudeman,
J. Guerr, J. Jones, J. B. Tenenbaum, J. S. Rule, J. Chua, K. Kanclerz, K. Livescu, K. Krauth,
K. Gopalakrishnan, K. Ignatyeva, K. Markert, K. D. Dhole, K. Gimpel, K. Omondi, K. Mathew-
son, K. Chiafullo, K. Shkaruta, K. Shridhar, K. McDonell, K. Richardson, L. Reynolds, L. Gao,
L. Zhang, L. Dugan, L. Qin, L. Contreras-Ochando, L.-P. Morency, L. Moschella, L. Lam,
L. Noble, L. Schmidt, L. He, L. O. Colón, L. Metz, L. K. ¸Senel, M. Bosma, M. Sap, M. ter Ho-
eve, M. Farooqi, M. Faruqui, M. Mazeika, M. Baturan, M. Marelli, M. Maru, M. J. R. Quintana,
M. Tolkiehn, M. Giulianelli, M. Lewis, M. Potthast, M. L. Leavitt, M. Hagen, M. Schubert,
M. O. Baitemirova, M. Arnaud, M. McElrath, M. A. Yee, M. Cohen, M. Gu, M. Ivanitskiy,
M. Starritt, M. Strube, M. Sw˛edrowski, M. Bevilacqua, M. Yasunaga, M. Kale, M. Cain, M. Xu,
M. Suzgun, M. Walker, M. Tiwari, M. Bansal, M. Aminnaseri, M. Geva, M. Gheini, M. V. T,
N. Peng, N. A. Chi, N. Lee, N. G.-A. Krakover, N. Cameron, N. Roberts, N. Doiron, N. Mar-
tinez, N. Nangia, N. Deckers, N. Muennighoff, N. S. Keskar, N. S. Iyer, N. Constant, N. Fiedel,
N. Wen, O. Zhang, O. Agha, O. Elbaghdadi, O. Levy, O. Evans, P. A. M. Casares, P. Doshi,
P. Fung, P. P. Liang, P. Vicol, P. Alipoormolabashi, P. Liao, P. Liang, P. Chang, P. Eckersley,
P. M. Htut, P. Hwang, P. Miłkowski, P. Patil, P. Pezeshkpour, P. Oli, Q. Mei, Q. Lyu, Q. Chen,
R. Banjade, R. E. Rudolph, R. Gabriel, R. Habacker, R. Risco, R. Millière, R. Garg, R. Barnes,
R. A. Saurous, R. Arakawa, R. Raymaekers, R. Frank, R. Sikand, R. Novak, R. Sitelew,
R. LeBras, R. Liu, R. Jacobs, R. Zhang, R. Salakhutdinov, R. Chi, R. Lee, R. Stovall, R. Tee-
han, R. Yang, S. Singh, S. M. Mohammad, S. Anand, S. Dillavou, S. Shleifer, S. Wiseman,
S. Gruetter, S. R. Bowman, S. S. Schoenholz, S. Han, S. Kwatra, S. A. Rous, S. Ghazarian,
S. Ghosh, S. Casey, S. Bischoff, S. Gehrmann, S. Schuster, S. Sadeghi, S. Hamdan, S. Zhou,
S. Srivastava, S. Shi, S. Singh, S. Asaadi, S. S. Gu, S. Pachchigar, S. Toshniwal, S. Upadhyay,
Shyamolima, Debnath, S. Shakeri, S. Thormeyer, S. Melzi, S. Reddy, S. P. Makini, S.-H. Lee,
S. Torene, S. Hatwar, S. Dehaene, S. Divic, S. Ermon, S. Biderman, S. Lin, S. Prasad, S. T.
Piantadosi, S. M. Shieber, S. Misherghi, S. Kiritchenko, S. Mishra, T. Linzen, T. Schuster,
T. Li, T. Yu, T. Ali, T. Hashimoto, T.-L. Wu, T. Desbordes, T. Rothschild, T. Phan, T. Wang,
T. Nkinyili, T. Schick, T. Kornev, T. Tunduny, T. Gerstenberg, T. Chang, T. Neeraj, T. Khot,
T. Shultz, U. Shaham, V. Misra, V. Demberg, V. Nyamai, V. Raunak, V. Ramasesh, V. U.
Prabhu, V. Padmakumar, V. Srikumar, W. Fedus, W. Saunders, W. Zhang, W. Vossen, X. Ren,
49
X. Tong, X. Zhao, X. Wu, X. Shen, Y. Yaghoobzadeh, Y. Lakretz, Y. Song, Y. Bahri, Y. Choi,
Y. Yang, Y. Hao, Y. Chen, Y. Belinkov, Y. Hou, Y. Hou, Y. Bai, Z. Seid, Z. Zhao, Z. Wang, Z. J.
Wang, Z. Wang, and Z. Wu. Beyond the Imitation Game: Quantifying and extrapolating the
capabilities of language models, June 2023. URL http://arxiv.org/abs/2206.04615.
[329]
Stanford Institute for Human-Centered Artificial Intelligence. Artificial Intelligence Index.
URL https://aiindex.stanford.edu/.
[330]
G. Stein, J. C. Cresswell, R. Hosseinzadeh, Y. Sui, B. L. Ross, V. Villecroze, Z. Liu, A. L.
Caterini, J. E. T. Taylor, and G. Loaiza-Ganem. Exposing flaws of generative model evaluation
metrics and their unfair treatment of diffusion models, Oct. 2023. URL
http://arxiv.org/
abs/2306.04675.
[331]
E. Strubell, A. Ganesh, and A. McCallum. Energy and Policy Considerations for Deep Learning
in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational
Linguistics, pages 3645–3650, Florence, Italy, 2019. Association for Computational Linguistics.
doi: 10.18653/v1/P19-1355. URL https://www.aclweb.org/anthology/P19-1355.
[332]
L. Struppek, D. Hintersdorf, F. Friedrich, M. Br, P. Schramowski, and K. Kersting. Exploiting
Cultural Biases via Homoglyphs in Text-to-Image Synthesis. Journal of Artificial Intelligence
Research, 78:1017–1068, Dec. 2023. ISSN 1076-9757. doi: 10
.
1613/jair
.
1
.
15388. URL
http://www.jair.org/index.php/jair/article/view/15388.
[333]
L. Struppek, D. Hintersdorf, and K. Kersting. Rickrolling the Artist: Injecting Back-
doors into Text Encoders for Text-to-Image Synthesis. In 2023 IEEE/CVF International
Conference on Computer Vision (ICCV), pages 4561–4573, Paris, France, Oct. 2023.
IEEE. ISBN 9798350307184. doi: 10
.
1109/ICCV51070
.
2023
.
00423. URL
https:
//ieeexplore.ieee.org/document/10377762/.
[334]
A. Subramonian, X. Yuan, H. Daumé Iii, and S. L. Blodgett. It Takes Two to Tango: Navigating
Conceptualizations of NLP Tasks and Measurements of Performance. In Findings of the
Association for Computational Linguistics: ACL 2023, pages 3234–3279, Toronto, Canada,
2023. Association for Computational Linguistics. doi: 10
.
18653/v1/2023
.
findings-acl
.
202.
URL https://aclanthology.org/2023.findings-acl.202.
[335]
L. Sun, Y. Huang, H. Wang, S. Wu, Q. Zhang, Y. Li, C. Gao, Y. Huang, W. Lyu, Y. Zhang,
X. Li, Z. Liu, Y. Liu, Y. Wang, Z. Zhang, B. Vidgen, B. Kailkhura, C. Xiong, C. Xiao, C. Li,
E. Xing, F. Huang, H. Liu, H. Ji, H. Wang, H. Zhang, H. Yao, M. Kellis, M. Zitnik, M. Jiang,
M. Bansal, J. Zou, J. Pei, J. Liu, J. Gao, J. Han, J. Zhao, J. Tang, J. Wang, J. Vanschoren,
J. Mitchell, K. Shu, K. Xu, K.-W. Chang, L. He, L. Huang, M. Backes, N. Z. Gong, P. S.
Yu, P.-Y. Chen, Q. Gu, R. Xu, R. Ying, S. Ji, S. Jana, T. Chen, T. Liu, T. Zhou, W. Wang,
X. Li, X. Zhang, X. Wang, X. Xie, X. Chen, X. Wang, Y. Liu, Y. Ye, Y. Cao, Y. Chen,
and Y. Zhao. TrustLLM: Trustworthiness in Large Language Models, Mar. 2024. URL
http://arxiv.org/abs/2401.05561.
[336]
H. Suresh and J. Guttag. A Framework for Understanding Sources of Harm throughout
the Machine Learning Life Cycle. In Equity and Access in Algorithms, Mechanisms, and
Optimization, pages 1–9, – NY USA, Oct. 2021. ACM. ISBN 978-1-4503-8553-4. doi:
10.1145/3465416.3483305. URL https://dl.acm.org/doi/10.1145/3465416.3483305.
[337]
Z. Talat and A. Lauscher. Back to the Future: On Potential Histories in NLP, Oct. 2022. URL
http://arxiv.org/abs/2210.06245.
[338]
Z. Talat, J. Thorne, and J. Bingel. Bridging the Gaps: Multi Task Learning for Domain Transfer
of Hate Speech Detection: Multi-task Learning for Domain Transfer of Hate Speech Detection.
In J. Golbeck, editor, Online Harassment, pages 29–55. Springer International Publishing,
Cham, 2018. ISBN 978-3-319-78582-0 978-3-319-78583-7. doi: 10
.
1007/978-3-319-78583-
7_3. URL https://link.springer.com/10.1007/978-3-319-78583-7_3.
[339]
Z. Talat, S. Lulz, J. Bingel, and I. Augenstein. Disembodied Machine Learning: On the Illusion
of Objectivity in NLP. Jan. 2021. URL http://arxiv.org/abs/2101.11974.
50
[340]
Z. Talat, A. Névéol, S. Biderman, M. Clinciu, M. Dey, S. Longpre, S. Luccioni, M. Masoud,
M. Mitchell, D. Radev, S. Sharma, A. Subramonian, J. Tae, S. Tan, D. Tunuguntla, and
O. Van Der Wal. You reap what you sow: On the Challenges of Bias Evaluation Under
Multilingual Settings. In Proceedings of BigScience Episode #5 – Workshop on Challenges
& Perspectives in Creating Large Language Models, pages 26–41, virtual+Dublin, 2022.
Association for Computational Linguistics. doi: 10
.
18653/v1/2022
.
bigscience-1
.
3. URL
https://aclanthology.org/2022.bigscience-1.3.
[341]
R. Taori and T. B. Hashimoto. Data feedback loops: Model-driven amplification of dataset
biases. In Proceedings of the 40th International Conference on Machine Learning, ICML’23,
¡conf-loc¿, ¡city¿Honolulu¡/city¿, ¡state¿Hawaii¡/state¿, ¡country¿USA¡/country¿, ¡/conf-loc¿,
2023. JMLR.org.
[342]
V. Taras, P. Steel, and B. L. Kirkman. Does Country Equate with Culture? Beyond Geography
in the Search for Cultural Boundaries. Management International Review, 56(4):455–487,
Aug. 2016. ISSN 0938-8249, 1861-8901. doi: 10
.
1007/s11575-016-0283-x. URL
http:
//link.springer.com/10.1007/s11575-016-0283-x.
[343]
G. Team, R. Anil, S. Borgeaud, Y. Wu, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M.
Dai, A. Hauth, K. Millican, D. Silver, S. Petrov, M. Johnson, I. Antonoglou, J. Schrittwieser,
A. Glaese, J. Chen, E. Pitler, T. Lillicrap, A. Lazaridou, O. Firat, J. Molloy, M. Isard, P. R.
Barham, T. Hennigan, B. Lee, F. Viola, M. Reynolds, Y. Xu, R. Doherty, E. Collins, C. Meyer,
E. Rutherford, E. Moreira, K. Ayoub, M. Goel, G. Tucker, E. Piqueras, M. Krikun, I. Barr,
N. Savinov, I. Danihelka, B. Roelofs, A. White, A. Andreassen, T. von Glehn, L. Yagati,
M. Kazemi, L. Gonzalez, M. Khalman, J. Sygnowski, A. Frechette, C. Smith, L. Culp, L. Pro-
leev, Y. Luan, X. Chen, J. Lottes, N. Schucher, F. Lebron, A. Rrustemi, N. Clay, P. Crone,
T. Kocisky, J. Zhao, B. Perz, D. Yu, H. Howard, A. Bloniarz, J. W. Rae, H. Lu, L. Sifre,
M. Maggioni, F. Alcober, D. Garrette, M. Barnes, S. Thakoor, J. Austin, G. Barth-Maron,
W. Wong, R. Joshi, R. Chaabouni, D. Fatiha, A. Ahuja, R. Liu, Y. Li, S. Cogan, J. Chen, C. Jia,
C. Gu, Q. Zhang, J. Grimstad, A. J. Hartman, M. Chadwick, G. S. Tomar, X. Garcia, E. Senter,
E. Taropa, T. S. Pillai, J. Devlin, M. Laskin, D. d. L. Casas, D. Valter, C. Tao, L. Blanco, A. P.
Badia, D. Reitter, M. Chen, J. Brennan, C. Rivera, S. Brin, S. Iqbal, G. Surita, J. Labanowski,
A. Rao, S. Winkler, E. Parisotto, Y. Gu, K. Olszewska, Y. Zhang, R. Addanki, A. Miech,
A. Louis, L. E. Shafey, D. Teplyashin, G. Brown, E. Catt, N. Attaluri, J. Balaguer, J. Xiang,
P. Wang, Z. Ashwood, A. Briukhov, A. Webson, S. Ganapathy, S. Sanghavi, A. Kannan, M.-W.
Chang, A. Stjerngren, J. Djolonga, Y. Sun, A. Bapna, M. Aitchison, P. Pejman, H. Michalewski,
T. Yu, C. Wang, J. Love, J. Ahn, D. Bloxwich, K. Han, P. Humphreys, T. Sellam, J. Brad-
bury, V. Godbole, S. Samangooei, B. Damoc, A. Kaskasoli, S. M. R. Arnold, V. Vasudevan,
S. Agrawal, J. Riesa, D. Lepikhin, R. Tanburn, S. Srinivasan, H. Lim, S. Hodkinson, P. Shyam,
J. Ferret, S. Hand, A. Garg, T. L. Paine, J. Li, Y. Li, M. Giang, A. Neitz, Z. Abbas, S. York,
M. Reid, E. Cole, A. Chowdhery, D. Das, D. Rogozi´
nska, V. Nikolaev, P. Sprechmann, Z. Nado,
L. Zilka, F. Prost, L. He, M. Monteiro, G. Mishra, C. Welty, J. Newlan, D. Jia, M. Allamanis,
C. H. Hu, R. de Liedekerke, J. Gilmer, C. Saroufim, S. Rijhwani, S. Hou, D. Shrivastava,
A. Baddepudi, A. Goldin, A. Ozturel, A. Cassirer, Y. Xu, D. Sohn, D. Sachan, R. K. Amplayo,
C. Swanson, D. Petrova, S. Narayan, A. Guez, S. Brahma, J. Landon, M. Patel, R. Zhao,
K. Villela, L. Wang, W. Jia, M. Rahtz, M. Giménez, L. Yeung, H. Lin, J. Keeling, P. Georgiev,
D. Mincu, B. Wu, S. Haykal, R. Saputro, K. Vodrahalli, J. Qin, Z. Cankara, A. Sharma,
N. Fernando, W. Hawkins, B. Neyshabur, S. Kim, A. Hutter, P. Agrawal, A. Castro-Ros,
G. van den Driessche, T. Wang, F. Yang, S.-y. Chang, P. Komarek, R. McIlroy, M. Luˇ
ci´
c,
G. Zhang, W. Farhan, M. Sharman, P. Natsev, P. Michel, Y. Cheng, Y. Bansal, S. Qiao, K. Cao,
S. Shakeri, C. Butterfield, J. Chung, P. K. Rubenstein, S. Agrawal, A. Mensch, K. Soparkar,
K. Lenc, T. Chung, A. Pope, L. Maggiore, J. Kay, P. Jhakra, S. Wang, J. Maynez, M. Phuong,
T. Tobin, A. Tacchetti, M. Trebacz, K. Robinson, Y. Katariya, S. Riedel, P. Bailey, K. Xiao,
N. Ghelani, L. Aroyo, A. Slone, N. Houlsby, X. Xiong, Z. Yang, E. Gribovskaya, J. Adler,
M. Wirth, L. Lee, M. Li, T. Kagohara, J. Pavagadhi, S. Bridgers, A. Bortsova, S. Ghemawat,
Z. Ahmed, T. Liu, R. Powell, V. Bolina, M. Iinuma, P. Zablotskaia, J. Besley, D.-W. Chung,
T. Dozat, R. Comanescu, X. Si, J. Greer, G. Su, M. Polacek, R. L. Kaufman, S. Tokumine,
H. Hu, E. Buchatskaya, Y. Miao, M. Elhawaty, A. Siddhant, N. Tomasev, J. Xing, C. Greer,
H. Miller, S. Ashraf, A. Roy, Z. Zhang, A. Ma, A. Filos, M. Besta, R. Blevins, T. Klimenko,
C.-K. Yeh, S. Changpinyo, J. Mu, O. Chang, M. Pajarskas, C. Muir, V. Cohen, C. L. Lan,
51
K. Haridasan, A. Marathe, S. Hansen, S. Douglas, R. Samuel, M. Wang, S. Austin, C. Lan,
J. Jiang, J. Chiu, J. A. Lorenzo, L. L. Sjösund, S. Cevey, Z. Gleicher, T. Avrahami, A. Boral,
H. Srinivasan, V. Selo, R. May, K. Aisopos, L. Hussenot, L. B. Soares, K. Baumli, M. B.
Chang, A. Recasens, B. Caine, A. Pritzel, F. Pavetic, F. Pardo, A. Gergely, J. Frye, V. Ra-
masesh, D. Horgan, K. Badola, N. Kassner, S. Roy, E. Dyer, V. Campos, A. Tomala, Y. Tang,
D. E. Badawy, E. White, B. Mustafa, O. Lang, A. Jindal, S. Vikram, Z. Gong, S. Caelles,
R. Hemsley, G. Thornton, F. Feng, W. Stokowiec, C. Zheng, P. Thacker, Ç. Ünlü, Z. Zhang,
M. Saleh, J. Svensson, M. Bileschi, P. Patil, A. Anand, R. Ring, K. Tsihlas, A. Vezer, M. Selvi,
T. Shevlane, M. Rodriguez, T. Kwiatkowski, S. Daruki, K. Rong, A. Dafoe, N. FitzGerald,
K. Gu-Lemberg, M. Khan, L. A. Hendricks, M. Pellat, V. Feinberg, J. Cobon-Kerr, T. Sainath,
M. Rauh, S. H. Hashemi, R. Ives, Y. Hasson, Y. Li, E. Noland, Y. Cao, N. Byrd, L. Hou,
Q. Wang, T. Sottiaux, M. Paganini, J.-B. Lespiau, A. Moufarek, S. Hassan, K. Shivaku-
mar, J. van Amersfoort, A. Mandhane, P. Joshi, A. Goyal, M. Tung, A. Brock, H. Sheahan,
V. Misra, C. Li, N. Raki´
cevi´
c, M. Dehghani, F. Liu, S. Mittal, J. Oh, S. Noury, E. Sezener,
F. Huot, M. Lamm, N. De Cao, C. Chen, G. Elsayed, E. Chi, M. Mahdieh, I. Tenney, N. Hua,
I. Petrychenko, P. Kane, D. Scandinaro, R. Jain, J. Uesato, R. Datta, A. Sadovsky, O. Bunyan,
D. Rabiej, S. Wu, J. Zhang, G. Vasudevan, E. Leurent, M. Alnahlawi, I. Georgescu, N. Wei,
I. Zheng, B. Chan, P. G. Rabinovitch, P. Stanczyk, Y. Zhang, D. Steiner, S. Naskar, M. Az-
zam, M. Johnson, A. Paszke, C.-C. Chiu, J. S. Elias, A. Mohiuddin, F. Muhammad, J. Miao,
A. Lee, N. Vieillard, S. Potluri, J. Park, E. Davoodi, J. Zhang, J. Stanway, D. Garmon, A. Kar-
markar, Z. Dong, J. Lee, A. Kumar, L. Zhou, J. Evens, W. Isaac, Z. Chen, J. Jia, A. Levskaya,
Z. Zhu, C. Gorgolewski, P. Grabowski, Y. Mao, A. Magni, K. Yao, J. Snaider, N. Casagrande,
P. Suganthan, E. Palmer, G. Irving, E. Loper, M. Faruqui, I. Arkatkar, N. Chen, I. Shafran,
M. Fink, A. Castaño, I. Giannoumis, W. Kim, M. Rybi´
nski, A. Sreevatsa, J. Prendki, D. So-
ergel, A. Goedeckemeyer, W. Gierke, M. Jafari, M. Gaba, J. Wiesner, D. G. Wright, Y. Wei,
H. Vashisht, Y. Kulizhskaya, J. Hoover, M. Le, L. Li, C. Iwuanyanwu, L. Liu, K. Ramirez,
A. Khorlin, A. Cui, T. LIN, M. Georgiev, M. Wu, R. Aguilar, K. Pallo, A. Chakladar, A. Repina,
X. Wu, T. van der Weide, P. Ponnapalli, C. Kaplan, J. Simsa, S. Li, O. Dousse, F. Yang, J. Piper,
N. Ie, M. Lui, R. Pasumarthi, N. Lintz, A. Vijayakumar, L. N. Thiet, D. Andor, P. Valen-
zuela, C. Paduraru, D. Peng, K. Lee, S. Zhang, S. Greene, D. D. Nguyen, P. Kurylowicz,
S. Velury, S. Krause, C. Hardin, L. Dixon, L. Janzer, K. Choo, Z. Feng, B. Zhang, A. Singhal,
T. Latkar, M. Zhang, Q. Le, E. A. Abellan, D. Du, D. McKinnon, N. Antropova, T. Boluk-
basi, O. Keller, D. Reid, D. Finchelstein, M. A. Raad, R. Crocker, P. Hawkins, R. Dadashi,
C. Gaffney, S. Lall, K. Franko, E. Filonov, A. Bulanova, R. Leblond, V. Yadav, S. Chung,
H. Askham, L. C. Cobo, K. Xu, F. Fischer, J. Xu, C. Sorokin, C. Alberti, C.-C. Lin, C. Evans,
H. Zhou, A. Dimitriev, H. Forbes, D. Banarse, Z. Tung, J. Liu, M. Omernick, C. Bishop,
C. Kumar, R. Sterneck, R. Foley, R. Jain, S. Mishra, J. Xia, T. Bos, G. Cideron, E. Amid,
F. Piccinno, X. Wang, P. Banzal, P. Gurita, H. Noga, P. Shah, D. J. Mankowitz, A. Polozov,
N. Kushman, V. Krakovna, S. Brown, M. Bateni, D. Duan, V. Firoiu, M. Thotakuri, T. Natan,
A. Mohananey, M. Geist, S. Mudgal, S. Girgin, H. Li, J. Ye, O. Roval, R. Tojo, M. Kwong,
J. Lee-Thorp, C. Yew, Q. Yuan, S. Bagri, D. Sinopalnikov, S. Ramos, J. Mellor, A. Sharma,
A. Severyn, J. Lai, K. Wu, H.-T. Cheng, D. Miller, N. Sonnerat, D. Vnukov, R. Greig, J. Beattie,
E. Caveness, L. Bai, J. Eisenschlos, A. Korchemniy, T. Tsai, M. Jasarevic, W. Kong, P. Dao,
Z. Zheng, F. Liu, F. Yang, R. Zhu, M. Geller, T. H. Teh, J. Sanmiya, E. Gladchenko, N. Trdin,
A. Sozanschi, D. Toyama, E. Rosen, S. Tavakkol, L. Xue, C. Elkind, O. Woodman, J. Carpenter,
G. Papamakarios, R. Kemp, S. Kafle, T. Grunina, R. Sinha, A. Talbert, A. Goyal, D. Wu,
D. Owusu-Afriyie, C. Du, C. Thornton, J. Pont-Tuset, P. Narayana, J. Li, S. Fatehi, J. Wieting,
O. Ajmeri, B. Uria, T. Zhu, Y. Ko, L. Knight, A. Héliou, N. Niu, S. Gu, C. Pang, D. Tran,
Y. Li, N. Levine, A. Stolovich, N. Kalb, R. Santamaria-Fernandez, S. Goenka, W. Yustalim,
R. Strudel, A. Elqursh, B. Lakshminarayanan, C. Deck, S. Upadhyay, H. Lee, M. Dusenberry,
Z. Li, X. Wang, K. Levin, R. Hoffmann, D. Holtmann-Rice, O. Bachem, S. Yue, S. Arora,
E. Malmi, D. Mirylenka, Q. Tan, C. Koh, S. H. Yeganeh, S. Põder, S. Zheng, F. Pongetti,
M. Tariq, Y. Sun, L. Ionita, M. Seyedhosseini, P. Tafti, R. Kotikalapudi, Z. Liu, A. Gulati,
J. Liu, X. Ye, B. Chrzaszcz, L. Wang, N. Sethi, T. Li, B. Brown, S. Singh, W. Fan, A. Parisi,
J. Stanton, C. Kuang, V. Koverkathu, C. A. Choquette-Choo, Y. Li, T. J. Lu, A. Ittycheriah,
P. Shroff, P. Sun, M. Varadarajan, S. Bahargam, R. Willoughby, D. Gaddy, I. Dasgupta,
G. Desjardins, M. Cornero, B. Robenek, B. Mittal, B. Albrecht, A. Shenoy, F. Moiseev,
H. Jacobsson, A. Ghaffarkhah, M. Rivière, A. Walton, C. Crepy, A. Parrish, Y. Liu, Z. Zhou,
C. Farabet, C. Radebaugh, P. Srinivasan, C. van der Salm, A. Fidjeland, S. Scellato, E. Latorre-
52
Chimoto, H. Klimczak-Pluci´
nska, D. Bridson, D. de Cesare, T. Hudson, P. Mendolicchio,
L. Walker, A. Morris, I. Penchev, M. Mauger, A. Guseynov, A. Reid, S. Odoom, L. Loher,
V. Cotruta, M. Yenugula, D. Grewe, A. Petrushkina, T. Duerig, A. Sanchez, S. Yadlowsky,
A. Shen, A. Globerson, A. Kurzrok, L. Webb, S. Dua, D. Li, P. Lahoti, S. Bhupatiraju, D. Hurt,
H. Qureshi, A. Agarwal, T. Shani, M. Eyal, A. Khare, S. R. Belle, L. Wang, C. Tekur, M. S.
Kale, J. Wei, R. Sang, B. Saeta, T. Liechty, Y. Sun, Y. Zhao, S. Lee, P. Nayak, D. Fritz,
M. R. Vuyyuru, J. Aslanides, N. Vyas, M. Wicke, X. Ma, T. Bilal, E. Eltyshev, D. Balle,
N. Martin, H. Cate, J. Manyika, K. Amiri, Y. Kim, X. Xiong, K. Kang, F. Luisier, N. Tripura-
neni, D. Madras, M. Guo, A. Waters, O. Wang, J. Ainslie, J. Baldridge, H. Zhang, G. Pruthi,
J. Bauer, F. Yang, R. Mansour, J.Gelman, Y. Xu, G. Polovets, J. Liu, H. Cai, W. Chen, X. Sheng,
E. Xue, S. Ozair, A. Yu, C. Angermueller, X. Li, W. Wang, J. Wiesinger, E. Koukoumidis,
Y. Tian, A. Iyer, M. Gurumurthy, M. Goldenson, P. Shah, M. K. Blake, H. Yu, A. Urbanowicz,
J. Palomaki, C. Fernando, K. Brooks, K. Durden, H. Mehta, N. Momchev, E. Rahimtoroghi,
M. Georgaki, A. Raul, S. Ruder, M. Redshaw, J. Lee, K. Jalan, D. Li, G. Perng, B. Hecht-
man, P. Schuh, M. Nasr, M. Chen, K. Milan, V. Mikulik, T. Strohman, J. Franco, T. Green,
D. Hassabis, K. Kavukcuoglu, J. Dean, and O. Vinyals. Gemini: A Family of Highly Capable
Multimodal Models, Dec. 2023. URL http://arxiv.org/abs/2312.11805.
[344]
The GDELT Project. The Unintended Consequences & Harms Of Multimodal LLM
Debiasing: Detection Vs Generation, 2023. URL https://blog.gdeltproject.org/the-
unintended-consequences-harms-of-multimodal-llm-debiasing-detection-
vs-generation/.
[345]
The Glaze Project, S. Shan, B. Zhao, H. Zheng, W. Ding, A. Ha, J. Passananti, S. Wu,
R. Bhaskar, J. Cryan, and E. Wenger. Glaze - Protecting Artists from Generative AI. URL
https://glaze.cs.uchicago.edu/.
[346]
The International Association of Privacy Professionals. Global Privacy Law and DPA Directory,
2024. URL https://iapp.org/resources/global-privacy-directory/.
[347]
The Ministry of Economy, Trade and Industry. Governance Guidelines for Implementation of
AI Principles Ver. 1.1" Compiled, 2022. URL
https://www.meti.go.jp/press/2021/01/
20220125001/20220124003.html.
[348]
D. Thiel, M. Stroebel, and R. Portnoff. Generative ML and CSAM: Implications and
Mitigations. 2023. doi: 10
.
25740/jv206yg3793. URL
https://purl.stanford.edu/
jv206yg3793.
[349]
Thorn. Generative AI: Now is the Time for Safety By Design, May 2023. URL
https:
//www.thorn.org/blog/now-is-the-time-for-safety-by-design/.
[350]
N. Thylstrup and Z. Talat. Detecting ‘Dirt’ and ‘Toxicity’: Rethinking Content Moderation
as Pollution Behaviour. SSRN Electronic Journal, 2020. ISSN 1556-5068. doi: 10
.
2139/
ssrn.3709719. URL https://www.ssrn.com/abstract=3709719.
[351]
N. Todorovic and A. Chaudhuri. Using AI to help organizations detect and report child sexual
abuse material online, Sept. 2018. URL
https://blog.google/around-the-globe/
google-europe/using-ai-help-organizations-detect-and-report-child-
sexual-abuse-material-online/.
[352]
J. A. Tomain. Online Privacy and the First Amendment: An Opt-In Approach to Data
Processing, Feb. 2014. URL https://papers.ssrn.com/abstract=2573206.
[353]
J. Tomlinson. Cultural Imperialism. In G. Ritzer, editor, The Wiley-Blackwell Encyclopedia
of Globalization. Wiley, 1 edition, Feb. 2012. ISBN 978-1-4051-8824-1 978-0-470-67059-0.
doi: 10
.
1002/9780470670590
.
wbeog129. URL
https://onlinelibrary.wiley.com/doi/
10.1002/9780470670590.wbeog129.
[354]
H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière,
N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample. LLaMA:
Open and Efficient Foundation Language Models, Feb. 2023. URL
http://arxiv.org/abs/
2302.13971.
53
[355]
Treasury Board of Canada Secretariat. Algorithmic Impact Assessment Tool, Mar.
2021. URL
https://www.canada.ca/en/government/system/digital-government/
digital-government-innovations/responsible-use-ai/algorithmic-impact-
assessment.html.
[356]
UN Human Rights Office of the High Commissioner. Non-discrimination (ND) Enhancing
equality and countering discrimination, 2021. URL
https://www2.ohchr.org/english/
ohchrreport2021/documents/Non-DiscriminationND.pdf.
[357]
United Kingdom’s Department for Digital, Culture, Media & Sport and D. Collins. UK
sets out proposals for new AI rulebook to unleash innovation and boost public trust in
the technology, 2022. URL
https://www.gov.uk/government/news/uk-sets-out-
proposals-for-new-ai-rulebook-to-unleash-innovation-and-boost-public-
trust-in-the-technology.
[358]
United Nations Framework Convention on Climate Change. Methodologies and Tools to Evalu-
ate Climate Change Impacts and Adaptation. URL
https://unfccc.int/methodologies-
and-tools-to-evaluate-climate-change-impacts-and-adaptation-2.
[359]
U.S. Army Acquisition Support Center. U.S. army acquisition support center, 2024. URL
https://asc.army.mil/web/news-generation-generation/.
[360]
U.S. Department of State. Political Declaration on Responsible Military Use of Artificial Intel-
ligence and Autonomy, 2024. URL
https://www.state.gov/political-declaration-
on-responsible-military-use-of-artificial-intelligence-and-autonomy/.
[361]
C. Vaccari and A. Chadwick. Deepfakes and Disinformation: Exploring the Impact of Synthetic
Political Video on Deception, Uncertainty, and Trust in News. Social Media + Society, 6(1):
205630512090340, Jan. 2020. ISSN 2056-3051, 2056-3051. doi: 10
.
1177/2056305120903408.
URL http://journals.sagepub.com/doi/10.1177/2056305120903408.
[362]
M. Veale and R. Binns. Fairer machine learning in the real world: Mitigating discrim-
ination without collecting sensitive data. Big Data & Society, 4(2):205395171774353,
Dec. 2017. ISSN 2053-9517, 2053-9517. doi: 10
.
1177/2053951717743530. URL
http://journals.sagepub.com/doi/10.1177/2053951717743530.
[363]
A. Venigalla and L. Li. Mosaic LLMs: GPT-3 quality for, Thu, 09/29/2022 - 13:28. URL
https://www.databricks.com/blog/gpt-3-quality-for-500k.
[364]
B. Vidgen, A. Agrawal, A. M. Ahmed, V. Akinwande, N. Al-Nuaimi, N. Alfaraj, E. Alhajjar,
L. Aroyo, T. Bavalatti, B. Blili-Hamelin, et al. Introducing v0. 5 of the ai safety benchmark
from mlcommons. arXiv preprint arXiv:2404.12241, 2024.
[365]
J. Vipra and S. M. West. Computational Power and AI. Technical report, AI Now Institute,
New York, NY, USA, 2023. URL
https://ainowinstitute.org/publication/policy/
compute-and-ai.
[366]
D. Walsh. The legal issues presented by generative AI, 2023. URL
https:
//mitsloan.mit.edu/ideas-made-to-matter/legal-issues-presented-
generative-ai.
[367]
A. Wang, A. Singh, J. Michael, F. Hill, O. Levy, and S. Bowman. GLUE: A Multi-Task
Benchmark and Analysis Platform for Natural Language Understanding. In Proceedings of
the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for
NLP, pages 353–355, Brussels, Belgium, 2018. Association for Computational Linguistics.
doi: 10.18653/v1/W18-5446. URL http://aclweb.org/anthology/W18-5446.
[368]
A. Wang, V. V. Ramaswamy, and O. Russakovsky. Towards Intersectionality in Machine
Learning: Including More Identities, Handling Underrepresentation, and Performing Eval-
uation. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages
336–349, Seoul Republic of Korea, June 2022. ACM. ISBN 978-1-4503-9352-2. doi:
10.1145/3531146.3533101. URL https://dl.acm.org/doi/10.1145/3531146.3533101.
54
[369]
A. Wang, X. Bai, S. Barocas, and S. L. Blodgett. Measuring machine learning harms from
stereotypes: Requires understanding who is being harmed by which errors in what ways, Feb.
2024. URL http://arxiv.org/abs/2402.04420.
[370]
Z. Waseem. Are You a Racist or Am I Seeing Things? Annotator Influence on Hate Speech
Detection on Twitter. In Proceedings of the First Workshop on NLP and Computational Social
Science, pages 138–142, Austin, Texas, 2016. Association for Computational Linguistics. doi:
10.18653/v1/W16-5618. URL http://aclweb.org/anthology/W16-5618.
[371]
L. Weidinger, J. Mellor, M. Rauh, C. Griffin, J. Uesato, P.-S. Huang, M. Cheng, M. Glaese,
B. Balle, A. Kasirzadeh, Z. Kenton, S. Brown, W. Hawkins, T. Stepleton, C. Biles, A. Birhane,
J. Haas, L. Rimell, L. A. Hendricks, W. Isaac, S. Legassick, G. Irving, and I. Gabriel. Ethical
and social risks of harm from Language Models, Dec. 2021. URL
http://arxiv.org/abs/
2112.04359.
[372]
L. Weidinger, K. R. McKee, R. Everett, S. Huang, T. O. Zhu, M. J. Chadwick, C. Summerfield,
and I. Gabriel. Using the Veil of Ignorance to align AI systems with principles of justice. Pro-
ceedings of the National Academy of Sciences, 120(18):e2213709120, May 2023. doi: 10
.
1073/
pnas.2213709120. URL https://www.pnas.org/doi/10.1073/pnas.2213709120.
[373]
L. Weidinger, M. Rauh, N. Marchal, A. Manzini, L. A. Hendricks, J. Mateos-Garcia,
S. Bergman, J. Kay, C. Griffin, B. Bariach, I. Gabriel, V. Rieser, and W. Isaac. Sociotechnical
Safety Evaluation of Generative AI Systems, Oct. 2023. URL
http://arxiv.org/abs/
2310.11986.
[374]
R. Weitzer. Racial discrimination in the criminal justice system: Findings and problems
in the literature. Journal of Criminal Justice, 24(4):309–322, Jan. 1996. ISSN 0047-2352.
doi: 10
.
1016/0047-2352(96)00015-3. URL
https://www.sciencedirect.com/science/
article/pii/0047235296000153.
[375]
Y. Wen, Y. Liu, C. Chen, and L. Lyu. Detecting, Explaining, and Mitigating Memorization in
Diffusion Models. In The Twelfth International Conference on Learning Representations, Oct.
2023. URL https://openreview.net/forum?id=84n3UwkH7b.
[376]
F. Westin and S. Chiasson. Opt out of privacy or "go home": Understanding reluctant privacy
behaviours through the FoMO-centric design paradigm. In Proceedings of the New Security
Paradigms Workshop, NSPW ’19, pages 57–67, New York, NY, USA, Jan. 2020. Association
for Computing Machinery. ISBN 978-1-4503-7647-1. doi: 10
.
1145/3368860
.
3368865. URL
https://doi.org/10.1145/3368860.3368865.
[377]
C. C. Williams and A. Efendic. Evaluating the relationship between marginalization and
participation in undeclared work: Lessons from Bosnia and Herzegovina. Southeast Euro-
pean and Black Sea Studies, 21(3):481–499, July 2021. ISSN 1468-3857. doi: 10
.
1080/
14683857.2021.1928419. URL https://doi.org/10.1080/14683857.2021.1928419.
[378]
L. Winner. Do Artifacts Have Politics? Daedalus, 109(1), 1980. URL
http://
www.jstor.org/stable/20024652.
[379]
A. Wolfers. “National Security” as an Ambiguous Symbol. Political Science Quarterly,
67(4):481–502, Dec. 1952. ISSN 0032-3195, 1538-165X. doi: 10
.
2307/2145138. URL
https://academic.oup.com/psq/article/67/4/481/7245057.
[380]
C.-J. Wu, R. Raghavendra, U. Gupta, B. Acun, N. Ardalani, K. Maeng, G. Chang, F. Aga,
J. Huang, C. Bai, M. Gschwind, A. Gupta, M. Ott, A. Melnikov, S. Candido, D. Brooks,
G. Chauhan, B. Lee, H.-H. Lee, B. Akyildiz, M. Balandat, J. Spisak, R. Jain, M. Rabbat, and
K. Hazelwood. Sustainable AI: Environmental implications, challenges and opportunities. In
D. Marculescu, Y. Chi, and C. Wu, editors, Proceedings of Machine Learning and Systems,
volume 4, pages 795–813, 2022. URL
https://proceedings.mlsys.org/paper_files/
paper/2022/file/462211f67c7d858f663355eff93b745e-Paper.pdf.
[381]
S. Wyllie, I. Shumailov, and N. Papernot. Fairness Feedback Loops: Training on Synthetic
Data Amplifies Bias, Mar. 2024. URL http://arxiv.org/abs/2403.07857.
55
[382]
A. Xu, E. Pathak, E. Wallace, S. Gururangan, M. Sap, and D. Klein. Detoxifying Language
Models Risks Marginalizing Minority Voices. In K. Toutanova, A. Rumshisky, L. Zettlemoyer,
D. Hakkani-Tur, I. Beltagy, S. Bethard, R. Cotterell, T. Chakraborty, and Y. Zhou, editors,
Proceedings of the 2021 Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, pages 2390–2397, Online, June
2021. Association for Computational Linguistics. doi: 10
.
18653/v1/2021
.
naacl-main
.
190.
URL https://aclanthology.org/2021.naacl-main.190.
[383]
Z. Yan, Y. Zhang, X. Yuan, S. Lyu, and B. Wu. DeepfakeBench: A compre-
hensive benchmark of deepfake detection. In A. Oh, T. Neumann, A. Glober-
son, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Infor-
mation Processing Systems, volume 36, pages 4534–4565. Curran Associates, Inc.,
2023. URL
https://proceedings.neurips.cc/paper_files/paper/2023/file/
0e735e4b4f07de483cbe250130992726-Paper-Datasets_and_Benchmarks.pdf.
[384]
A. G. Yasar, A. Chong, E. Dong, T. Gilbert, S. Hladikova, C. Mougan, X. Shen, S. Singh, A.-A.
Stoica, and S. Thais. Integration of generative ai in the digital markets act: Contestability and
fairness from a cross-disciplinary perspective. SSRN Electronic Journal, 2024. ISSN 1556-
5068. doi: 10.2139/ssrn.4769439. URL http://dx.doi.org/10.2139/ssrn.4769439.
[385]
J. Zaller and S. Feldman. A Simple Theory of the Survey Response: Answering Ques-
tions versus Revealing Preferences. American Journal of Political Science, 36(3):579, Aug.
1992. ISSN 00925853. doi: 10
.
2307/2111583. URL
https://www.jstor.org/stable/
2111583?origin=crossref.
[386]
S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V.
Lin, T. Mihaylov, M. Ott, S. Shleifer, K. Shuster, D. Simig, P. S. Koura, A. Sridhar, T. Wang,
and L. Zettlemoyer. OPT: Open Pre-trained Transformer Language Models, June 2022. URL
http://arxiv.org/abs/2205.01068.
[387]
D. Zhuang, X. Zhang, S. Song, and S. Hooker. Randomness in Neural Network Training:
Characterizing the Impact of Tooling. Proceedings of Machine Learning and Systems, 4:316–
336, Apr. 2022. URL
https://proceedings.mlsys.org/paper_files/paper/2022/
hash/427e0e886ebf87538afdf0badb805b7f-Abstract.html.
56
