GPTs are GPTs: An Early Look at the Labor Market Impact
Potential of Large Language Models
Tyna Eloundou1, Sam Manning1,2, Pamela Mishkin∗1, and Daniel Rock3
3University of Pennsylvania
March 20, 2023
We investigate the potential implications of Generative Pre-trained Transformer (GPT) models and
related technologies on the U.S. labor market. Using a new rubric, we assess occupations based on their
correspondence with GPT capabilities, incorporating both human expertise and classiﬁcations from GPT-4.
Our ﬁndings indicate that approximately 80% of the U.S. workforce could have at least 10% of their work
tasks aﬀected by the introduction of GPTs, while around 19% of workers may see at least 50% of their
tasks impacted. The inﬂuence spans all wage levels, with higher-income jobs potentially facing greater
exposure. Notably, the impact is not limited to industries with higher recent productivity growth. We
conclude that Generative Pre-trained Transformers exhibit characteristics of general-purpose technologies
(GPTs), suggesting that as these models could have notable economic, social, and policy implications.
As shown in Figure 1, recent years, months, and weeks have seen remarkable progress in the ﬁeld of generative
AI and large language models (LLMs). While the public often associates LLMs with various iterations of the
Generative Pre-trained Transformer (GPT), LLMs can be trained using a range of architectures, and are not
limited to transformer-based models (Devlin et al., 2019). LLMs can process and produce various forms of
sequential data, including assembly language, protein sequences and chess games, extending beyond natural
language applications alone. In this paper, we use LLMs and GPTs somewhat interchangeably, and specify in
our rubric that these should be considered similar to the GPT-family of models available via ChatGPT or
the OpenAI Playground (which at the time of labeling included models in the GPT-3.5 family but not in the
GPT-4 family). We examine GPTs with text- and code-generating abilities and employ the term "generative
AI" to additionally include modalities such as images or audio.
Our study is motivated less by the progress of these models alone though, and more by the breadth,
scale, and capabilities we’ve seen in the complementary technologies developed around them. The role of
complementary technologies remains to be seen, but maximizing the impact of LLMs appears contingent
on integrating them with larger systems (Bresnahan, 2019; Agrawal et al., 2021). While we focus much of
this discussion on the generative capabilities of LLMs, there may be new types of software and machine
communication made possible by use of LLMs for other tasks including things like embeddings which make
∗Corresponding author (email@example.com). Authors contributed equally and are listed alphabetically.
arXiv:2303.10130v1 [econ.GN] 17 Mar 2023
Figure 1: To get a sense of how quickly model capabilities are progressing – consider the jump in exam
performance between GPT-3.5 and GPT-4 (OpenAI, 2023b).
it possible to build custom search applications or tasks like summarization and classiﬁcation where it can be
unclear where to draw the distinction over what is or is not generative.
To contextualize this progression and complement labor impact forecasts of technology, we propose a
new rubric for understanding LLM capabilities and their potential eﬀects on jobs. This rubric (A.1) measures
the overall exposure of tasks to GPTs, following the spirit of prior work on quantifying exposure to machine
learning (Brynjolfsson et al., 2018; Felten et al., 2018; Webb, 2020). We deﬁne exposure as a proxy for
potential economic impact without distinguishing between labor-augmenting or labor-displacing eﬀects. We
employ human annotators and GPT-4 itself as a classiﬁer to apply this rubric to occupational data in the U.S.
economy, primarily sourced from the O*NET database.1 2
To construct our primary exposure dataset, we collected both human annotations and GPT-4 classiﬁcations,
using a prompt tuned for agreement with a sample of labels from the authors. We observe similar agreement
levels in GPT-4 responses and between human and machine evaluations, when aggregated to the task level.
This measure reﬂects an estimate of the technical capacity to make human labor more eﬃcient; however,
social, economic, regulatory, or other determinants imply that technical feasibility does not guarantee labor
productivity or automation outcomes. Our analysis indicates that approximately 19% of jobs have at least
50% of their tasks exposed when considering both current model capabilities and anticipated tools built upon
them. Human assessments suggest that only 3% of U.S. workers have over half of their tasks exposed to
GPT when considering existing language and code capabilities without additional software or modalities.
Accounting for other generative models and complementary technologies, our human estimates indicate that
This is distinct from recent social science research that makes use of advanced language models to simulate human behavior
(Horton, 2023; Sorensen et al., 2022)
While our exposure rubric does not necessarily tie the concept of language models to any particular model, we were strongly
motivated by our observed capabilities of GPT-4 and the suite of capabilities we saw in development with OpenAI’s launch partners
up to 49% of workers could have half or more of their tasks exposed to LLMs.
Our ﬁndings consistently show across both human and GPT-4 annotations that most occupations exhibit
some degree of exposure to LLMs, with varying exposure levels across diﬀerent types of work. Occupations
with higher wages generally present with high exposure, a result contrary to similar evaluations of overall
machine learning exposure (Brynjolfsson et al., 2023). When regressing exposure measures on skillsets using
O*NET’s skill rubric, we discover that roles heavily reliant on science and critical thinking skills show a
negative correlation with exposure, while programming and writing skills are positively associated with
LLM exposure. Following Autor et al. (2022a), we examine barriers to entry by "job zones" and ﬁnd that
occupational exposure to LLMs weakly increases with the diﬃculty of job preparation. In other words,
workers facing higher (lower) barriers to entry in their jobs tend to experience more (less) exposure to LLMs.
We further compare our measurements to previous eﬀorts documenting the distribution of automation
exposure in the economy and ﬁnd broadly consistent results. Most other technology exposure measures we
examine are statistically signiﬁcantly correlated with our preferred exposure measure, while measures of
manual routineness and robotics exposure show negative correlations. The variance explained by these earlier
eﬀorts (Acemoglu and Autor, 2011a; Frey and Osborne, 2017; Brynjolfsson et al., 2018; Felten et al., 2018;
Webb, 2020; Brynjolfsson et al., 2023), along with wage controls, ranges from 60 to 72%, indicating that 28
to 40% of the variation in our AI exposure measure remains unaccounted for by previous technology exposure
We analyze exposure by industry and discover that information processing industries (4-digit NAICS)
exhibit high exposure, while manufacturing, agriculture, and mining demonstrate lower exposure. The
connection between productivity growth in the past decade and overall GPT exposure appears weak, suggesting
a potential optimistic case that future productivity gains from LLMs may not exacerbate possible cost disease
eﬀects (Baumol, 2012). 3
Our analysis indicates that the impacts of LLMs like GPT-4, are likely to be pervasive. While LLMs
have consistently improved in capabilities over time, their growing economic eﬀect is expected to persist and
increase even if we halt the development of new capabilities today. We also ﬁnd that the potential impact of
LLMs expands signiﬁcantly when we take into account the development of complementary technologies.
Collectively, these characteristics imply that Generative Pre-trained Transformers (GPTs) are general-purpose
(Bresnahan and Trajtenberg, 1995; Lipsey et al., 2005). (Goldfarb et al., 2023) argue
that machine learning as a broad category is likely a general-purpose technology. Our evidence supports a
wider impact, as even subsets of machine learning software meet the criteria for general-purpose technology
status independently. This paper’s primary contributions are to provide a set of measurements of LLM
impact potential and to demonstrate the use case of applying LLMs to develop such measurements eﬃciently
and at scale. Additionally, we showcase the general-purpose potential of LLMs. If "GPTs are GPTs," the
eventual trajectory of LLM development and application may be challenging for policymakers to predict and
regulate. As with other general-purpose technologies, much of these algorithms’ potential will emerge across
a broad range of economically valuable use cases, including the creation of new types of work (Acemoglu
and Restrepo, 2018; Autor et al., 2022a) Our research serves to measure what is technically feasible now, but
necessarily will miss the evolving impact potential of the LLMs over time.
The paper is structured as follows: Section 2 reviews relevant prior work, Section 3 discusses methods
and data collection, Section 4 presents summary statistics and results, Section 5 relates our measurements to
earlier eﬀorts, Section 6 explores results, and Section 7 oﬀers concluding remarks.
Baumol’s cost disease is a theory that explains why the cost of labor-intensive services, such as healthcare and education,
increases over time. This happens because wages for skilled workers in other industries increase, but there is no corresponding
increase in productivity or eﬃciency in these service industries. Therefore, the cost of labor in these industries becomes relatively
more expensive compared to other goods and services in the economy.
For the remainder of the paper, we use GPT to refer to large language models generally as exempliﬁed by those available via
OpenAI, and we spell out general-purpose technologies when it is used outside of stating "GPTs are GPTs."
2 Literature Review
2.1 The Advancement of Large Language Models
In recent years, large language models (LLMs) have risen to prominence in the ﬁeld of artiﬁcial intelligence
(AI) research, showcasing their ability to tackle a wide array of complex language-based tasks. This progress
has been fueled by multiple factors, including increased model parameter count, greater training data volume,
and enhanced training conﬁgurations (Brown et al., 2020; Radford et al., 2019; Hernandez et al., 2021;
Kaplan et al., 2020). Broad, state-of-the-art LLMs, such as LaMDA (Thoppilan et al., 2022) and GPT-4
(OpenAI, 2023b), excel in diverse applications like translation, classiﬁcation, creative writing, and code
generation—capabilities that previously demanded specialized, task-speciﬁc models developed by expert
engineers using domain-speciﬁc data.
Concurrently, researchers have improved the steerability, reliability, and utility of these models using
methods like ﬁne-tuning and reinforcement learning with human feedback (Ouyang et al., 2022; Bai et al.,
2022). These advancements enhance the models’ ability to discern user intent, rendering them more
user-friendly and practical. Moreover, recent studies reveal the potential of LLMs to program and control
other digital tools, such as APIs, search engines, and even other generative AI systems (Schick et al., 2023;
Mialon et al., 2023; Chase, 2022). This enables seamless integration of individual components for better
utility, performance, and generalization. In the long run, these trends suggest that LLMs may be capable of
executing any task typically performed at a computer.
For the most part, generative AI models have predominantly been deployed as modular specialists,
carrying out speciﬁc tasks, like generating images from captions or transcribing text from speech. However,
we argue that it is essential to adopt a broader perspective, recognizing LLMs as crucial building blocks
for additional tools. While constructing these tools and integrating them into comprehensive systems will
take time and necessitate signiﬁcant reconﬁguration of existing processes across the economy, we already
observe emerging adoption trends. Despite their limitations, LLMs are becoming increasingly integrated into
specialized applications in areas such as writing assistance, coding, and legal research, paving the way for
businesses and individuals to adopt GPTs more widely.
We emphasize the signiﬁcance of these complementary technologies, partly because out-of-the-box
general-purpose GPTs may continue to be unreliable for various tasks due to issues such as factual inaccuracies,
inherent biases, privacy concerns, and disinformation risks (Abid et al., 2021; Schramowski et al., 2022;
Goldstein et al., 2023; OpenAI, 2023a). However, specialized workﬂows—including tooling, software, or
human-in-the-loop systems—can help address these shortcomings by incorporating domain-speciﬁc expertise.
For example, Casetext oﬀers LLM-based legal research tools that provide lawyers with quicker and more
accurate legal research results, utilizing embeddings and summarization to counter the risk that GPT-4
provides innacurate details about a legal case or set of documents. GitHub Copilot is a coding assistant that
employs LLMs to generate code snippets and autocomplete code, which users can then accept or reject based
on their expertise. In other words, while it’s true that on its own GPT-4 does not "know what time it is," it’s
easy enough to give it a watch.
Furthermore, a positive feedback loop may emerge as LLMs surpass a speciﬁc performance threshold,
allowing them to assist in building the very tooling that enhances their usefulness and usability across various
contexts. This could lower the cost and engineering expertise required to create such tools, potentially
accelerating LLM adoption and integration even further. (Chen et al., 2021; Peng et al., 2023) LLMs can also
become valuable assets in machine learning model development—serving as coding assistants for researchers,
data labeling services, or synthetic data generators. There is potential for such models to contribute to
economic decision-making at the task level, for instance, by reﬁning methods for task and sub-task allocation
between humans and machines (Singla et al., 2015; Shahaf and Horvitz, 2010). As LLMs improve over
time and better align with user preferences, we can anticipate a continuous enhancement in performance.
However, it is essential to recognize that these trends also bring a variety of serious risks. (Khlaaf et al., 2022;
Weidinger et al., 2022; Solaiman et al., 2019)
2.2 The Economic Impacts of Automation Technologies
A large and growing body of literature addresses the labor market impacts of artiﬁcial intelligence and
automation technologies broadly deﬁned. The concept of skill-biased technological change and the task
model of automation—often considered the standard framework for understanding technology’s inﬂuence
on labor—originated from research demonstrating that technological progress raises the demand for skilled
workers over unskilled workers (Katz and Murphy, 1992). Numerous studies have built upon this concept,
exploring the eﬀects of technological change and automation on workers within a task-based framework (Autor
et al., 2003; Acemoglu and Autor, 2011b; Acemoglu and Restrepo, 2018). This strand of research has shown
that workers involved in routine and repetitive tasks are at a higher risk of technology-driven displacement, a
phenomenon known as routine-biased technological change. More recent studies have distinguished between
technology’s task-displacement and task-reinstatement eﬀects (where new technology increases the need for
a wider array of labor-intensive tasks) (Acemoglu and Restrepo, 2018, 2019). Several studies have shown
that automation technologies have resulted in wage inequality in the US, driven by relative wage declines for
workers specializing in routine tasks (Autor et al., 2006; Van Reenen, 2011; Acemoglu and Restrepo, 2022b).
Prior research has employed various approaches to estimate the overlap between AI capabilities and
the tasks and activities workers undertake in diﬀerent occupations. These methods include mapping patent
descriptions to worker task descriptions (Webb, 2020; Meindl et al., 2021), linking AI capabilities to
occupational abilities documented in the O*NET database (Felten et al., 2018, 2023), aligning AI task
benchmark evaluations with worker tasks via cognitive abilities (Tolan et al., 2021), labeling automation
potential for a subset of US occupations and using machine learning classiﬁers to estimate this potential
for all other US occupations (Frey and Osborne, 2017), modeling task-level automation and aggregating
the results to occupation-level insights (Arntz et al., 2017), expert forecasts (Grace et al., 2018), and most
relevantly to this paper, devising a new rubric to assess worker activities for their suitability for machine
learning (Brynjolfsson et al., 2018, 2023). Some of these approaches have found exposure to AI technologies
at the task-level tends to be diversiﬁed within occupation. Considering each job as a bundle of tasks, it would
be rare to ﬁnd any occupation for which AI tools could do nearly all of the work. (Autor et al., 2022a) ﬁnds as
well that automation and augmentation exposures tend to be positively correlated. There is also a growing set
of studies examining speciﬁc economic impacts and opportunities for LLMs (Bommasani et al., 2021; Felten
et al., 2023; Korinek, 2023; Mollick and Mollick, 2022; Noy and Zhang, 2023; Peng et al., 2023). Alongside
this work, our measurements help characterize the broader potential relevance of language models to the
General-purpose technologies (e.g. printing, the steam engine) (GPTs) are characterized by widespread
proliferation, continuous improvement, and the generation of complementary innovations (Bresnahan and
Trajtenberg, 1995; Lipsey et al., 2005). Their far-reaching consequences, which unfold over decades, are
diﬃcult to anticipate, particularly in relation to labor demand (Bessen, 2018; Korinek and Stiglitz, 2018;
Acemoglu et al., 2020; Benzell et al., 2021). The realization of general purpose technologies’ full potential
requires extensive co-invention (Bresnahan and Trajtenberg, 1995; Bresnahan et al., 1996, 2002; Lipsey et al.,
2005; Dixon et al., 2021), a costly and time-consuming process involving the discovery of new business
procedures (David, 1990; Bresnahan, 1999; Frey, 2019; Brynjolfsson et al., 2021; Feigenbaum and Gross,
2021). Consequently, many studies of machine learning technologies focus on systems-level adoption, arguing
that organizational systems may require redesign to eﬀectively take advantage of novel machine learning
advancements (Bresnahan, 2019; Agrawal et al., 2021; Goldfarb et al., 2023). Appropriately designed systems
can yield considerable business value and improve ﬁrm performance (Rock, 2019; Babina et al., 2021; Zolas
et al., 2021), with AI tools facilitating the discovery process (Cockburn et al., 2018; Cheng et al., 2022). By
Task ID Occupation Title DWAs Task Description
14675 Computer Systems
Monitor computer system performance
to ensure proper operation.
Monitor system operation to detect potential
18310 Acute Care Nurses Operate diagnostic or therapeutic
medical instruments or equipment.
Prepare medical supplies or equipment
Set up, operate, or monitor invasive
equipment and devices, such as colostomy or
tracheotomy equipment, mechanical
ventilators, catheters, gastrointestinal tubes,
and central lines.
4668.0 Gambling Cage
Execute sales or other ﬁnancial
Cash checks and process credit card advances
15709 Online Merchants Execute sales or other ﬁnancial
Deliver e-mail conﬁrmation of completed
transactions and shipment.
– Involve parent volunteers and older students in
children’s activities to facilitate involvement
in focused, complex play.
6568 Elementary School
– Involve parent volunteers and older students in
children’s activities to facilitate involvement
in focused, complex play.
Table 1: Sample of occupations, tasks, and Detailed Work Activities from the O*NET database. We see
that aggregating over activities alone is imprecise, as evidenced by the fact that we’d expect Gambling Cage
Workers to complete the given DWA in person, using some physicality while we’d expect Online Merchants
to complete the same activity solely with a computer.
employing task-level information to assess whether LLMs fulﬁll GPT criteria, we seek to merge the two
perspectives for understanding the technology-labor relationship.
We attempt to build on these diverse literature streams in several ways. Echoing (Felten et al., 2023), we
focus our analysis on the impact of LLMs, rather than addressing machine learning or automation technologies
more broadly. Additionally, we propose a novel method that employs LLMs, speciﬁcally GPT-4, to assess tasks
for exposure and automation potential, thereby bolstering human scoring eﬀorts. Subsequently, we aggregate
our ﬁndings to occupations and industries, capturing the overall potential exposure in the contemporary U.S.
3 Methods and Data Collection
3.1 Data on Activities and Tasks Performed by Occupation in the US
We use the O*NET 27.2 database (O*NET, 2023), which contains information on 1,016 occupations, including
their respective Detailed Work Activities (DWAs) and tasks. A DWA is a comprehensive action that is part of
completing task, such as "Study scripts to determine project requirements." A task, on the other hand, is an
occupation-speciﬁc unit of work that may be associated with none, one, or multiple DWAs. We oﬀer a sample
of tasks and DWAs in Table 1. The two datasets we use consist of:
19,265 tasks, where each task featuring a "task description" and a corresponding occupation, and with
most tasks associated with one or more DWAs
2,087 DWAs, where most DWAs are connected to one or more tasks, and tasks may be associated with
one or more DWAs, though some tasks lack any associated DWAs
3.2 Data on Wages, Employment, and Demographics
We obtain employment and wage data from the 2020 and 2021 Occupational Employment series provided by
the Bureau of Labor Statistics. This dataset encompasses occupational titles, the number of workers in each
occupation, and occupation-level employment projections for 2031, typical education required for entry in an
occupation and on-the-job training required to attain competency in an occupation (BLS, 2022). We use the
BLS-recommended crosswalk to O*NET (BLS, 2023b) to link the O*NET task and DWA dataset and the
BLS Labor Force Demographics (BLS, 2023a), which is derived from the Current Population Survey (CPS).
Both of these data sources are collected by the U.S. government and primarily capture workers who are not
self-employed, are documented, and are working in the so-called formal economy.
We present our results based on an exposure rubric, in which we deﬁne
as a measure of whether
access to a GPT or GPT-powered system would reduce the time required for a human to perform a speciﬁc
DWA or complete a task by at least 50 percent. We provide a summary of our rubric below, while the
complete rubric can be found in A.1. When we have labels for DWAs, we ﬁrst aggregate at the task level
before aggregating at the occupation level.
No exposure (E0) if:
there is no or minimal reduction in the time required to complete the activity or task
while maintaining equivalent quality or
using any combination of the capabilities described in accordance with the below criteria
would decrease the quality of the activity/task output.
Direct exposure (E1) if:
using solely the theoretical LLM or GPT-4 described via ChatGPT or the OpenAI
playground can decrease the time required to complete the DWA or task by at least half
LLM+ Exposed (E2) if:
access to the LLM alone would not reduce the time required to complete the activity/task
by at least half, but
additional software could be developed on to the LLM that could reduce the time it takes
to complete the speciﬁc activity/task with quality by at least half. Among these systems,
we count access to image generation systems. a
In practice, as can be seen in the full rubric in A.1, we categorize access to image capabilities separately
(E3) to facilitate annotation, though we combine E2 and E3 for all analyses.
We set the exposure threshold at a potential 50 % reduction in time required to complete a speciﬁc DWA
or task while maintaining consistent quality. We anticipate that adoption will be highest and most immediate
for applications that realize a considerable increase in productivity. Although this threshold is somewhat
arbitrary, it was selected for ease of interpretation by annotators. 5
We then collected both human and GPT-4-generated annotations using the exposure rubric, which underlie
the bulk of the analyses in this paper.
Human Ratings: We obtained human annotations by applying the rubric to each O*NET Detailed
Worker Activity (DWA) and a subset of all O*NET tasks and then aggregated those DWA and task
Moreover, regardless of the chosen threshold, we guessed that the real-world reduction in task time would likely be slightly or
signiﬁcantly lower than our estimates, leading us to opt for a relatively high threshold. In our own validation labeling, we found that
this corresponded closely to whether GPT or GPT-powered applications could perform the core part of a task or nearly the entire task.
Comparison 𝛾Weighting Agreement Pearson’s
GPT-4, Rubric 1; Human 𝛼E1 80.8% 0.223
𝛽E1 + .5*E2 65.6% 0.591
𝜁E1 + E2 82.1% 0.654
GPT-4, Rubric 2; Human 𝛼E1 81.8% 0.221
𝛽E1 + .5*E2 65.6% 0.538
𝜁E1 + E2 79.5% 0.589
GPT-4, Rubric 1; GPT-4, Rubric 2 𝛼E1 91.1% 0.611
𝛽E1 + .5*E2 76.0% 0.705
𝜁E1 + E2 82.4% 0.680
Table 2: Model and human comparison of agreement and Pearson’s correlation scores. The agreement score
is determined by looking at how often the two groups agree on the annotation (e.g. E0, E1 or E2). In the
paper we use GPT-4, Rubric 1.
at the task and occupation levels. To ensure the quality of these annotations, the authors
personally labeled a large sample of tasks and DWAs and enlisted experienced human annotators who
have extensively reviewed GPT outputs as part of OpenAI’s alignment work (Ouyang et al., 2022).
GPT-4 Ratings: We administered a similar rubric to an early version of GPT-4 (OpenAI, 2023b) but on
all task/occupation pairs rather than DWAs. We made slight modiﬁcations to the rubric (which was
used as a "prompt" to the model in this case) to enhance agreement with a set of human labels. Full
agreement rates are given in Table 2.
We construct three primary measures for our dependent variable of interest: (i)
, corresponding to E1 in
the exposure rubric above, anticipated to represent the lower bound of the proportion of exposed tasks within
, which is the sum of E1 and 0.5*E2, where the 0.5 weight on E2 is intended to account
for exposure when deploying the technology via complementary tools and applications necessitates additional
investment, and (iii)
, the sum of E1 and E2, an upper bound of exposure that provides an assessment of
maximal exposure to GPT and GPT-powered software. We summarize agreement between annotation groups
and measures in Table 2. For the remainder of the analysis, if not speciﬁed, the reader may assume that we
exposure – meaning all tasks directly exposed via tools like ChatGPT or the OpenAI Playground are
considered twice as exposed as tasks requiring some complementary innovation.
3.4 Limitations of our methodology
3.4.1 Subjective human judgments
A fundamental limitation of our approach lies in the subjectivity of the labeling. In our study, we employ
annotators who are familiar with the GPT models’ capabilities. However, this group is not occupationally
diverse, potentially leading to biased judgments regarding GPTs’ reliability and eﬀectiveness in performing
tasks within unfamiliar occupations. We acknowledge that obtaining high-quality labels for each task in an
occupation requires workers engaged in those occupations or, at a minimum, possessing in-depth knowledge
The authors annotated DWAs that clearly required a high degree of physicality or manual dexterity, and the contracted annotators
labeled the remaining activities, along with a subset of tasks including those without associated DWAs and those for which there was
no clear task-level annotation after aggregating the DWA annotations.
Despite recent advances of multimodal GPT models (OpenAI, 2023b), vision capabilities were not included in the assessment of
Figure 2: Human raters (x-axis) and GPT-4 ratings (y-axis) show a high degree of agreement about GPT
exposure by occupation. Near the highest levels of exposure following the
method of aggregating exposure
scores to occupations, GPT-4 ratings tend to be lower than Human ratings. We present the raw scatter and the
binscatter. Near the top end of exposure ratings, humans are on average more likely to rate an occupation as
of the diverse tasks within those occupations. This represents an important area for future work in validating
3.4.2 Measuring GPTs with GPT-4
Recent research indicates that GPT-4 serves as an eﬀective discriminator, capable of applying intricate
taxonomies and responding to changes in wording and emphasis. (OpenAI, 2023b) The outcomes of GPT-4
task classiﬁcation are sensitive to alterations in the rubric’s wording, the prompt’s order and composition, the
presence or absence of speciﬁc examples in the rubric, the level of detail provided, and key term deﬁnitions.
Iterating on the prompt, based on observed outcomes in a small validation set, can enhance the agreement
between model outputs and the rubric’s intent. Consequently, there are slight diﬀerences between the rubric
presented to humans and the one used for GPT-4. This decision was made deliberately to guide the model
towards reasonable labels without excessively inﬂuencing human annotators. As a result, we use multiple
annotation sources, but none should be considered the deﬁnitive ground truth relative to the others. In the
analysis, we will present results from human annotators as our primary results. Further improvement and
innovation in crafting eﬀective rubrics for LLM classiﬁcation remain possible. Still, we observe a high degree
of agreement between human ratings and GPT-4 ratings at the occupation level concerning overall exposure
to GPT systems (see Table 2, Figure ??).
3.4.3 Additional Weaknesses
•Validity of task-based framework.
It is unclear to what extent occupations can be entirely broken
down into tasks, and whether this approach systemtically omits certain categories of skills or tasks
that are tacitly required for competent performance of a job. Additionally, tasks can be composed of
sub-tasks, some of which are more automatable than others. Some tasks may function as pre-cursor to
other tasks, such that the completion of downstream tasks is dependent on precursor tasks. If indeed,
the task-based breakdown is not a valid representation of how most work in an occupation is performed,
our exposure analysis would largely be invalidated.
•Relative vs. absolute measures.
It is likely best to interpret these measures as relative measures, e.g.
an occupation with an estimated 0.6 exposure should likely be interpreted as just much more exposed
than one with 0.1 exposure.
•Lack of expertise and task interpretation.
Human annotators were mostly unaware of the speciﬁc
occupations mapped to each DWA during the labeling process. This led to unclear logic for aggregating
tasks and occupations, as well as some evident discrepancies in labels, demonstrated in Table 1. We
experimented with various aggregation methods and discovered that even with a maximum-matching
approach (taking the matching human<>model label if one existed), the agreement remained relatively
consistent. Ultimately, we collected additional labels for task/occupation pairs where there was
•Forward-looking and subject to change, with some early evidence.
Accurately predicting future
LLM applications remains a signiﬁcant challenge, even for experts (OpenAI, 2023b). Emergent
capabilities, human perception biases, and technological development shifts can all aﬀect the accuracy
and reliability of predictions regarding LLMs’ potential impact on worker tasks. Our projections are
inherently forward-looking and based on current trends, evidence, and perceptions of technological
possibilities. As a result, they may change as new advancements arise in the ﬁeld. For example, some
tasks that seem unlikely for LLMs to impact today might change with the introduction of new model
capabilities. Conversely, tasks that appear exposed might face unforeseen challenges limiting language
•Sources of disagreement.
While we did not rigorously examine sources of disagreement, we found a
few places where humans and the model tended to get "stuck" in their assessments:
Tasks or activities where while an LLM could theoretically help or accomplish the task, adopting
it to do so would require multiple people to change their habits or expectations (e.g. meetings,
Tasks or activities where there is currently some regulation that requires human oversight or norm
that suggests human judgment or empathy (e.g. making decisions, counseling), and
Tasks or activities where there already exists a technology that can reasonably automate the task
(e.g. making reservations).
General-purpose technologies are relatively rare and characterized by their pervasiveness, improvement over
time, and the development of signiﬁcant co-invention and spillovers (Lipsey et al., 2005). Our assessment of
GPTs’ (Generative Pre-trained Transformers) impact on the labor market is limited since it does not consider
total factor productivity or capital input potential. In addition to their inﬂuence on labor, GPTs may also
inﬂuence these dimensions.
At this stage, certain GPT criteria are easier to evaluate than others. For example, assessing the long-term
impact of these models’ capabilities and the growth of complementary applications and systems is more
feasible in the long run. Our primary focus at this early stage is to test the hypothesis that GPT language
models have a pervasive inﬂuence on the economy, similar to (Goldfarb et al., 2023)’s analysis of machine
learning diﬀusion through job postings to assess machine learning’s GPT potential as an algorithmic category.
Rather than using job postings or studying machine learning in general, examining the task evaluation
approach with both human and GPT annotations may reveal whether GPT impacts are limited to a small set
of similar tasks or occupations.
Our ﬁndings suggest that, based on their task-level capabilities, GPTs have the potential to signiﬁcantly
aﬀect a diverse range of occupations within the U.S. economy, demonstrating a key attribute of general-purpose
technologies. In the following sections, we discuss results across various roles and wage structures. Additional
results on the relative exposure of industries within the U.S. economy can be found in Appendix D.
4.1 Summary Statistics
Summary statistics for these measures can be found in Table 3. Both human and GPT-4 annotations indicate
that average occupation-level
values fall between 0.14 and 0.15, suggesting that, for the median occupation,
approximately 15% of tasks are directly exposed to GPTs. This ﬁgure increases to over 30% for
surpasses 50% for
. Coincidentally, human and GPT-4 annotations also tag between 15% and 14% of total
tasks in the dataset as being exposed to GPTs.
Based on the
values, we estimate that 80% of workers belong to an occupation with at least one task
exposed to GPTs, while 19% of workers are in an occupation where over half the tasks are labeled as exposed.
Although the potential for tasks to be aﬀected is extensive, GPTs must be incorporated into broader
systems to realize this potential fully. As is common with general-purpose technologies, such co-invention
barriers may impede the rapid diﬀusion of GPTs into economic applications. Additionally, predicting the
need for human oversight is challenging, especially for tasks where model capabilities equal or surpass human
levels. While the requirement for human supervision may initially slow down the adoption and diﬀusion rate,
users of GPTs and GPT-powered systems are likely to become increasingly acquainted with the technology
over time, particularly in terms of understanding when and how to trust its outputs.
Occupation Level Exposure
mean std mean std
𝛼0.14 0.14 0.14 0.16
𝛽0.30 0.21 0.34 0.22
𝜁0.46 0.30 0.55 0.34
Task Level Exposure
mean std mean std
𝛼0.15 0.36 0.14 0.35
𝛽0.31 0.37 0.35 0.35
𝜁0.47 0.50 0.56 0.50
Table 3: Summary statistics of our human and model exposure data.
4.2 Wages and Employment
In Figure 3, we present the exposure intensity across the economy. The ﬁrst plot displays exposure in terms
of total workers, while the second plot shows exposure in terms of total occupations. Each point on the
graph represents the estimated percentage of workers (and occupations) on the y-axis with an exposure level
) indicated on the x-axis. For example, human annotators determined that 2.4% of workers are
-exposed, 18.6% are
-exposed, and 49.6% are
-exposed, where the threshold of 50% comes from the
x-axis and the percentage of workers comes from the y axis in the right plot of Figure 2. At any given point on
the x-axis, the vertical distance between the
represents the exposure potential attributable to tools
and applications beyond direct exposure to GPTs. The distribution of exposure is similar for both workers and
Figure 3: Exposure intensity across the economy, displayed on the left in terms of percent of aﬀected
occupations and on the right as percent of aﬀected workers. The distribution of exposure is similar across
occupations and across workers, suggesting that worker concentration in occupations is not highly correlated
with occupational exposure to GPTs or GPT-powered software. We do however expect that it could be more
highly correlated with investment in developing GPT-powered software for particular domains.
occupations, suggesting that worker concentration in occupations does not have a strong correlation with
occupational exposure to GPT or GPT-powered software.
Aggregated at the occupation level, human and GPT-4 annotations exhibit qualitative similarities and
tend to correlate, as demonstrated in Figure 4. Human annotations estimate marginally lower exposure for
high-wage occupations compared to GPT-4 annotations. While there are numerous low-wage occupations
with high exposure and high-wage occupations with low exposure, the overall trend in the binscatter plot
reveals that higher wages are associated with increased exposure to GPT.
The potential exposure to GPTs seems to have little correlation with current employment levels. In Figure
4, both human and GPT-4 ratings of overall exposure are aggregated to the occupation-level (y-axis) and
compared with the log of total employment (x-axis). Neither plot reveals signiﬁcant diﬀerences in GPT
exposure across varying employment levels.
4.3 Skill Importance
In this section, we investigate the relationship between the importance of a skill for an occupation (as annotated
in the O*NET dataset) and our exposure measures. We ﬁrst take the Basic Skills provided by O*NET (skill
deﬁnitions can be found in Appendix B) and normalize the measure of skill importance for each occupation
to enhance interpretability. Then we perform a regression analysis on our exposure measures (
examine the strength of associations between skill importance and exposure.
Our ﬁndings indicate that the importance of
skills are strongly negatively
associated with exposure, suggesting that occupations requiring these skills are less likely to be impacted by
current language models. Conversely,
skills show a strong positive association
with exposure, implying that occupations involving these skills are more susceptible to being inﬂuenced by
language models (see Table 5 for detailed results).
4.4 Barriers to Entry
Next, we examine barriers to entry to better understand if there is diﬀerentiation in exposure due to types of
jobs. One such proxy is an O*NET occupation-level descriptor called the "Job Zone." A job zone groups
occupations that are similar in (a) the level of education needed to get a job in the occupation, (b) the amount
Figure 4: The binscatter plots depict the exposure to language models (LLMs) in various occupations,
as assessed by both human evaluators and GPT-4. These plots compare the exposure to GPT (
) at the
occupation level against the log of total employment within an occupation and log of the median annual
wage for occupations. While some discrepancies exist, both human and GPT-4 assessments indicate that
higher wage occupations tend to be more exposed to LLMs. Additionally, numerous lower wage occupations
demonstrate high exposure based on our rubrics. Core tasks receive twice the weight of supplemental tasks
within occupations when calculating average exposure scores. Employment and wage data are sourced from
the BLS-OES survey conducted in May 2021.
exposure ratings of occupations in the ﬁve Job Zones, which are groups of similar occupations that
are classiﬁed according to the level of education, experience, and on-the-job training needed to perform them.
of related experience required to do the work, and (c) the extent of on-the-job training needed to do the work.
In the ONET database, there are 5 Job Zones, with Job Zone 1 requiring the least amount of preparation (3
months) and Job Zone 5 requiring the most extensive amount of preparation, 4 or more years. We observe that
median income increases monotonically across job zones as the level of preparation needed also increases,
with the median worker in Job Zone 1 earning $30
230 and the median worker in Job Zone 5 earning $80
All of our measures (
) show an identical pattern, that is, exposure increases from Job Zone 1
to Job Zone 4, and either remains similar or decreases at Job Zone 5. Similar to Figure 3 in 5, we plot the
percentage of workers at every threshold of exposure. We ﬁnd that, on average, the percentage of workers in
occupations with greater than 50%
exposure in Job Zones 1 through 5 have
at 0.00% (Job Zone 1), 6.11%
(Job Zone 2), 10.57% (Job Zone 3), 34.5% (Job Zone 4), and 26.45% (Job Zone 5), respectively.
4.4.1 Typical Education Needed for Entry
Since inclusion in a job zone accounts for both the education required—which itself is a proxy for skill
acquisition—and the preparation required, we seek data to disentangle these variables. We use two variables
from the Bureau of Labor Statistics’ Occupational data: "Typical Education Needed for Entry" and "On-the-job
Training Required to Attain Competency" in an occupation. By examining these factors, we aim to uncover
trends with potential implications for the workforce. There are 3,504,000 workers for whom we lack data on
education and on-the-job training requirements, and they are therefore excluded from the summary tables.
Our analysis suggests that individuals holding Bachelor’s, Master’s, and professional degrees are more
exposed to GPTs and GPT-powered software than those without formal educational credentials (see Table 7).
Interestingly, we also ﬁnd that individuals with some college education but no degree exhibit a high level of
exposure to GPTs and GPT-powered software. Upon examining the table displaying barriers to entry, we
observe that the jobs with the least exposure require the longest training, potentially oﬀering a lower payoﬀ (in
terms of median income) once competency is achieved. Conversely, jobs with no on-the-job training required
or only internship/residency required appear to yield higher income but are more exposed to GPT.
Group Occupations with highest exposure % Exposure
𝛼Interpreters and Translators 76.5
Survey Researchers 75.0
Poets, Lyricists and Creative Writers 68.8
Animal Scientists 66.7
Public Relations Specialists 66.7
𝛽Survey Researchers 84.4
Writers and Authors 82.5
Interpreters and Translators 82.4
Public Relations Specialists 80.6
Animal Scientists 77.8
Tax Preparers 100.0
Financial Quantitative Analysts 100.0
Writers and Authors 100.0
Web and Digital Interface Designers 100.0
Humans labeled 15 occupations as "fully exposed."
Correspondence Clerks 95.2
Blockchain Engineers 94.1
Court Reporters and Simultaneous Captioners 92.9
Proofreaders and Copy Markers 90.9
Blockchain Engineers 97.1
Court Reporters and Simultaneous Captioners 96.4
Proofreaders and Copy Markers 95.5
Correspondence Clerks 95.2
𝜁Accountants and Auditors 100.0
News Analysts, Reporters, and Journalists 100.0
Legal Secretaries and Administrative Assistants
Clinical Data Managers 100.0
Climate Change Policy Analysts 100.0
The model labeled 86 occupations as "fully exposed."
Highest variance Search Marketing Strategists 14.5
Graphic Designers 13.4
Investment Fund Managers 13.0
Financial Managers 13.0
Insurance Appraisers, Auto Damage 12.6
Table 4: Occupations with the highest exposure according to each measurement. The ﬁnal row lists the
occupations with the highest
value, indicating that they had the most variability in vulnerability-prediction.
Exposure percentages indicate the share of an occupation’s task that are exposed to GPTs (
) or GPT-powered
), where exposure is deﬁned as driving a reduction in time it takes to complete the task by at
least 50% (see exposure rubric A.1. As such, occupations listed in this table are those where we estimate
that GPTs and GPT-powered software are able to save workers a signiﬁcant amount of time completing a
large share of their tasks, but it does not necessarily suggest that their tasks can be fully automated by these
Table 5: OLS Regression Results of Exposure Measures on O*NET Skills
𝛼 𝛽 𝜁
(std err) (std err) (std err)
Constant 0.082*** -0.112*** 0.300***
(0.011) (0.011) (0.057)
Active Listening 0.128** 0.214*** 0.449***
(0.047) (0.043) (0.027)
Mathematics -0.127*** 0.161*** 0.787***
(0.026) (0.021) (0.049)
Reading Comprehension 0.153*** 0.470*** -0.346***
(0.041) (0.037) (0.017)
Science -0.114*** -0.230*** -0.346***
(0.014) (0.012) (0.017)
Speaking -0.028 0.133*** 0.294***
(0.039) (0.033) (0.042)
Writing 0.368*** 0.467*** 0.566***
(0.042) (0.037) (0.047)
Active Learning -0.157*** -0.065** 0.028
(0.027) (0.024) (0.032)
Critical Thinking -0.264*** -0.196*** -0.129**
(0.036) (0.033) (0.042)
Learning Strategies -0.072* -0.209*** -0.346***
(0.028) (0.025) (0.034)
Monitoring -0.067** -0.149*** -0.232***
(0.023) 0.020) (0.026)
Programming 0.637*** 0.623*** 0.609***
(0.030) (0.022) (0.024)
Example Occupations Median
1 None or little
High school diploma
or GED (optional)
Food preparation workers,
dishwashers, ﬂoor sanders
$30,230 13,100 3.71 3.84 6.45 5.97 9.19 8.11
2 Some (3-12
High school diploma Orderlies, customer
$38,215 73,962 7.03 11.88 15.74 19.54 24.45 27.19
3 Medium (1-2
or associate’s degree
54,815 37,881 11.28 13.72 26.08 32.17 40.88 50.62
Bachelor’s degree Database administrators,
graphic designers, cost
$77,345 56,833 22.68 17.82 46.78 51.30 70.87 84.78
5 Extensive (4+
Master’s degree or
$81,980 21,221 22.81 13.36 43.11 44.64 63.41 75.92
Table 6: Exposure to GPTs by Job Zone
Med(Med) Inc Emp (000s) H 𝛼
None $77,440 90,776 0.20 0.16 0.42 0.46 0.63 0.76
Apprenticeship $55,995 3,066 0.01 0.02 0.04 0.06 0.07 0.10
Internship/residency $77,110 3,063 0.16 0.06 0.36 0.38 0.55 0.71
Short-term on-the-job training $33,370 66,234 0.11 0.15 0.21 0.25 0.32 0.34
Moderate-term on-the-job training $46,880 31,285 0.09 0.12 0.21 0.25 0.32 0.38
Long-term on-the-job training $48,925 5,070 0.08 0.10 0.18 0.22 0.28 0.33
Table 7: Mean exposure scores for occupations, grouped by level of on-the-job training required to attain
competency in the job. Alongside exposure scores, we display the median of median annual income for each
occupation, as well as the total number of workers in each group, in thousands.
5 Validation of Measures
5.1 Comparison to Earlier Eﬀorts
This paper aims to build on a number of previous empirical studies examining the occupational exposure to
advances in AI and/or automation. Previous studies have used a variety of methods, including:
Using occupational taxonomies like O*NET to characterize which occupations have routine vs.
non-routine and manual vs. cognitive task content (Autor et al., 2003; Acemoglu and Autor, 2011a).
Mapping text descriptions of tasks to descriptions of technological advances in patents. (Kogan et al.,
2021; Webb, 2020)
Linking capabilities of AI systems to occupational abilities and aggregating exposure estimates to the
occupations where those abilities are required. (Felten et al., 2018, 2023)
Mapping the results of AI task benchmark evaluations (ImageNet, Robocup, etc.) to 59 worker tasks
through a set of 14 cognitive abilities drawn from the cognitive science literature. (Tolan et al., 2021)
Expert labeling of automation potential for a set of O*NET occupations where experts had high
conﬁdence, combined with a probabilistic classiﬁer to estimate automation potential for the remainder
of O*NET occupations. (Frey and Osborne, 2017)
Developing a rubric for evaluating the "suitability for machine learning" (SML) of activities that
workers are completing in the economy (Brynjolfsson and Mitchell, 2017; Brynjolfsson et al., 2018,
We provide a set of summary statistics on many of these prior eﬀorts in Table 8.
This paper’s methodology primarily builds upon the SML approach by developing a rubric to evaluate the
overlap between LLM capabilities and worker tasks as reported in the O*NET database. Table 9 presents the
results of OLS regressions of our new LLM exposure measurements on occupation-level exposure measures
from (Felten et al., 2018) ("AI Occupational Exposure Score" in the table), (Frey and Osborne, 2017) (Frey
& Osborne Automation), scores from all three technologies in (Webb, 2020), normalized routine manual
and cognitive scores from (Acemoglu and Autor, 2011a), and (Brynjolfsson et al., 2018, 2023) (SML). We
also use annualized occupational salaries from the most recent BLS Occupational Employment Survey as a
control. There are four separate output variables representing new scores in this paper that are predicted by
GPT-4 Exposure Rating 1 corresponds to our overall exposure rubric as evaluated by GPT-4, where full
exposure potential is coded as 1, no exposure potential is coded as 0, and partial exposure (E2 in our labeling
scheme) is coded as 0.5. GPT-4 Exposure Rating 2 is scored similarly for overall exposure, but with a slightly
diﬀerent prompt. The results are very similar across the two prompts. GPT-4 Automation Rating applies our
"T" rubric, coding no automation exposure from LLMs as 0, full automation exposure as 1, and levels 2, 3,
and 4 as 0.25, 0.5, and 0.75, respectively. Finally, Human Exposure Rating represents the same rubric as in
GPT-4 Exposure Rating 1 but is scored by humans, as discussed in an earlier section of the paper. These
results correspond to the 𝛽set of statistics presented above.
The results across each type of measurement are consistent. We ﬁnd generally positive and statistically
signiﬁcant correlations between our LLM exposure measures and previous measurements targeting software
and AI. Encouragingly, the SML exposure scores by occupation show signiﬁcant and positive associations
with the exposure scores we develop in this paper, demonstrating a level of cohesion between the two studies
with similar approaches. The Webb software and AI patent-based measures, SML, and normalized (demeaned
Min 25th Perc. Median 75th Perc Max Mean Std. Dev. Count
GPT-4 Exposure Rating 1 0.00 0.13 0.34 0.50 1.00 0.33 0.22 750
GPT-4 Exposure Rating 2 0.00 0.09 0.24 0.40 0.98 0.26 0.20 750
Human Exposure Rating 0.00 0.09 0.29 0.47 0.84 0.29 0.21 750
Software (Webb) 1.00 25.00 50.00 75.00 100.00 50.69 30.05 750
Robot (Webb) 1.00 22.00 52.00 69.00 100.00 48.61 28.61 750
AI (Webb) 1.00 28.00 55.00 82.00 100.00 54.53 29.65 750
Suitability for Machine Learning 2.60 2.84 2.95 3.12 3.55 2.99 0.18 750
Normalized Routine Cognitive -3.05 -0.46 0.10 0.63 3.42 0.07 0.86 750
Normalized Routine Manual -1.81 -0.81 -0.11 0.73 2.96 0.05 1.01 750
AI Occupational Exposure Score 1.42 3.09 3.56 4.04 6.54 3.56 0.70 750
Frey & Osborne Automation 0.00 0.07 0.59 0.88 0.99 0.50 0.38 681
Log Avg. Salary 10.13 10.67 11.00 11.34 12.65 11.02 0.45 749
Table 8: Summary statistics for a suite of prior eﬀorts to measure occupational exposure to AI and automation.
We have also included summary statistics for measurements newly presented in this work. We include all
measures from (Webb, 2020), normalized routine cognitive and manual scores from (Acemoglu and Autor,
2011a) (means may deviate slightly from 0 due to imperfect matching of occupational groups), Suitability for
Machine Learning from (Brynjolfsson and Mitchell, 2017; Brynjolfsson et al., 2018, 2023), AI Occupational
Exposure from (Felten et al., 2018), and Automation exposure from (Frey and Osborne, 2017). We include as
many occupations as we can match, but since O*NET taxonomies have changed as these measures have been
developed, some of the roles may be missing from the most recent version of O*NET 6-digit occupations.
and divided by standard deviation) routine cognitive scores all exhibit positive associations with some of our
Software, SML, and routine cognitive scores all show positive and statistically signiﬁcant associations
with LLM exposure scores at a 1% level. Coeﬃcients on AI scores from (Webb, 2020) are also positive and
statistically signiﬁcant at a 5% level, but our secondary prompt on overall exposure to LLMs in columns 3
and 4 does not exhibit a statistically signiﬁcant relationship. For the most part, the AI Occupational Exposure
Score is not correlated with our exposure measures. Webb’s Robot exposure scores, routine manual task
content, and the overall Automation metric from (Frey and Osborne, 2017) are all negatively correlated with
our primary GPT-4 and human-assessed overall exposure ratings, conditional on the other measurements.
This negative correlation reﬂects the limited exposure of physical tasks to LLMs. Manual work is not exposed
to LLMs or even LLMs with additional systems integration for the time being. Our automation rubric results
are also uncorrelated with (Frey and Osborne, 2017) measures.
Low correlations with (Felten et al., 2018) and (Frey and Osborne, 2017) could potentially be explained
by diﬀerences in approaches. Linking AI capabilities to worker abilities or scoring exposure directly based on
the occupation’s characteristics, rather than aggregating up to the occupation from DWA or task-level scoring
(as in the SML paper and our own), oﬀer a slightly diﬀerent perspective on the content of occupations.
In all regressions, the
ranges between 60.7% (column 3) and 72.8% (column 5). This suggests that
our measure, which explicitly focuses on LLM capabilities, has between 28 and 40% unexplained variance
compared to other measurements. Particularly in the case of AI-related exposure scores, we anticipate that a
combination of other measurements would have a strong correlation with our scores. However, earlier eﬀorts
had limited information about the future progress of LLM technologies. We expect that our understanding of
future machine learning technologies is similarly imperfectly captured by our rubric today.
GPT-4 Exposure Rating 1 GPT-4 Exposure Rating 2 Human Exposure Rating
(1) (2) (3) (4) (5) (6)
Software (Webb) 0.00113∗∗∗ 0.00123∗∗∗ 0.00111∗∗∗ 0.00119∗∗∗ 0.00096∗∗∗ 0.00101∗∗∗
(0.00031) (0.00031) (0.00031) ( 0.00031) (0.00031) (0.00031)
Robot (Webb) −0.00378∗∗∗ −0.00405∗∗∗ −0.00377∗∗∗ −0.00399∗∗∗ −0.00371∗∗∗ −0.00383∗∗∗
(0.00032) (0.00031) (0.00034) ( 0.00033) (0.00029) (0.00028)
AI (Webb) 0.00080∗∗∗ 0.00090∗∗∗ 0.00036 0.00045 0.00067∗∗ 0.00071∗∗
(0.00030) (0.00029) (0.00030) ( 0.00030) (0.00030) (0.00030)
Suitability for Machine Learning 0.29522∗∗∗ 0.26888∗∗∗ 0.28468∗∗∗ 0.26245∗∗∗ 0.19514∗∗∗ 0.18373∗∗∗
(0.04503) (0.04418) (0.04404) ( 0.04342) (0.03990) (0.03886)
Normalized Routine Cognitive 0.06601∗∗∗ 0.06868∗∗∗ 0.04743∗∗∗ 0.05015∗∗∗ 0.03568∗∗∗ 0.03659∗∗∗
(0.00886) (0.00894) (0.00872) ( 0.00879) (0.00671) (0.00669)
Normalized Routine Manual −0.11147∗∗∗ −0.11371∗∗∗ −0.09390∗∗∗ −0.09561∗∗∗ −0.11045∗∗∗ −0.11152∗∗∗
(0.00785) (0.00789) (0.00817) ( 0.00818) (0.00741) (0.00744)
AI Occupational Exposure Score 0.00993 0.02465∗∗ −0.01537 −0.00265 0.00630 0.01252
(0.01107) (0.01059) (0.01160) ( 0.01114) (0.00918) (0.00845)
Frey & Osborne Automation −0.03024∗−0.03950∗∗ −0.00364 −0.01217 −0.03890∗∗ −0.04253∗∗
(0.01835) (0.01841) (0.02007) ( 0.01972) (0.01883) (0.01858)
Log Avg. Salary 0.05804∗∗∗ 0.04863∗∗∗ 0.02531
(0.01870) (0.01860) (0.01727)
Constant −1.12937∗∗∗ −0.45743∗∗∗ −0.96117∗∗∗ −0.39935∗∗∗ −0.47078∗−0.17706
(0.26859) (0.15327) (0.26365) ( 0.15017) (0.24684) (0.13256)
N680.00000 681.00000 680.00000 681.00000 680.00000 681.00000
𝑅20.68741 0.68212 0.60737 0.60198 0.71213 0.71126
Table 9: Regression of GPT-exposure cores on prior eﬀorts. Regression coeﬃcients from exposure measures
from our rubrics on earlier eﬀorts to quantify occupational exposure to AI and automation. We also include
annualized wages from the BLS-OES survey in May 2021. Each measure is kept in its original scale, with the
exception of routine cognitive and routine manual scores from (Acemoglu and Autor, 2011a). Those two
scores are standardized to mean zero and variance 1. Generally we ﬁnd strong positive associations with
previous eﬀorts, though large residual variance to still be explained by our new measures. Columns 1 and 2
are based on our main
exposure measure from GPT-4 ratings. Columns 3 and 4 are based on a similar
slightly diﬀerent exposure rubric also rated by GPT-4 for robustness. Columns 5 and 6 reﬂect human ratings
on the same rubric as columns 1 and 2.
6.1 GPTs as a General-Purpose Technology
Earlier in this paper we discuss the possibility that GPTs could be classiﬁed as a general-purpose technology.
This classiﬁcation requires GPTs to meet three core criteria: improvement over time, pervasiveness throughout
the economy, and the ability to spawn complementary innovations (Lipsey et al., 2005). Evidence from the AI
and machine learning literature thoroughly demonstrates that GPTs meet the ﬁrst criteria – they are improving
in capabilities over time with the ability to complete or be helpful for an increasingly complex set of tasks and
use-cases (see 2.1). This paper presents evidence to support the latter two criteria, ﬁnding that GPTs on their
own can have pervasive impacts across the economy, and that complementary innovations enabled by GPTs –
particularly via software and digital tools – can have widespread application to economic activity.
Figure 3 oﬀers one illustration of the potential economic impact of complementary software built on top of
LLMs. Taking the diﬀerence in the y-axis (the share of all occupations) between
at a given point along
the x-axis (the share of tasks within an occupation that are exposed) gives the aggregate within-occupation
exposure potential attributable to tools and software over and above direct exposure from LLMs on their
own. The diﬀerence in means across all tasks between
of 0.42 using the GPT-4 annotations and 0.32
using the human annotations (see Figure 3), suggests that the average impact of GPT-powered software on
task-exposure may be more than twice as large as the mean exposure from LLMs on their own (mean
based on both human annotations and GPT-4 annotations). While our ﬁndings suggest that out-of-the-box
these models are relevant to a meaningful share of workers and tasks, they also suggest that the software
innovations they spawn could drive a much broader impact.
One component of the pervasiveness of a technology is its level of adoption by businesses and users. This
paper does not systematically analyze adoption of these models, however, there is early qualitative evidence
that adoption and use of LLMs is becoming increasingly widespread. The power of relatively simple UI
improvements on top of LLMs was evident in the rollout of ChatGPT – wherein versions of the underlying
model had been previously available via API, but usage skyrocketed after the release of the ChatGPT interface.
(Chow, 2023; OpenAI, 2022) Following this release, a number of commercial surveys indicate that ﬁrm and
worker adoption of LLMs has increased over the past several months. (Constantz, 2023; ResumeBuilder.com,
Widespread adoption of these models, however, necessitates the identiﬁcation of existing bottlenecks. A
key determinant of their utility is the level of conﬁdence humans place in them, as well as habits. For instance,
in the legal profession, the models’ usefulness hinges upon whether legal professionals can trust their output
without resorting to verifying original documents or conducting independent research. The cost and ﬂexibility
of the technology, worker and ﬁrm preferences, and incentives also play a signiﬁcant role in the adoption of
tools built on top of LLMs. In this way, adoption may be driven by progress on some of the ethical and safety
risks associated with LLMs: bias, making up facts, and misalignment to name a few OpenAI (2023a).
Moreover, the adoption of LLMs will vary across diﬀerent economic sectors due to factors such as data
availability, regulatory quality, innovation culture, and the distribution of power and interests. Consequently,
a comprehensive understanding of the adoption and of large language models by workers and ﬁrms requires a
more in-depth exploration of these intricacies.
One possibility is that time savings and seamless application will hold greater importance than quality
improvement for the majority of tasks. Another is that the initial focus will be on augmentation, followed by
automation (Huang and Rust, 2018). One way this might take shape is that an augmentation phase where jobs
ﬁrst become more precarious (writers become freelancers) could play out prior to full automation.
6.2 Implications for US Public Policy
The introduction of automation technologies, including LLMs, has previously been linked to heightened
economic disparity and labor disruption, which may give rise to adverse downstream eﬀects.(Acemoglu and
Restrepo, 2022a; Acemoglu, 2002; Moll et al., 2021; Klinova and Korinek, 2021; Weidinger et al., 2021,
2022) Our results examining worker exposure in the United States underscore the need for societal and policy
preparedness to the potential economic disruption posed by LLMs and the complementary technologies
that they spawn. While it is outside the scope of this paper to recommend speciﬁc policy prescriptions to
smooth the transition to an economy with increasingly widespread LLM adoption, prior work such as (Autor
et al., 2022b) has articulated several important directions for US policy related to education, worker training,
reforms to safety net programs, and more.
6.3 Limitations and Future Work
This study possesses several limitations that warrant further investigation. Primarily, our focus on the United
States restricts the generalizability of our ﬁndings to other nations where the adoption and impact of generative
models may diﬀer due to factors such as industrial organization, technological infrastructure, regulatory
frameworks, linguistic diversity, and cultural contexts. We hope to address this limitation by extending the
study’s scope and by sharing our methods so other researchers can build on them.
Subsequent research eﬀorts should consider two additional studies: one exploring GPT adoption patterns
across various sectors and occupations, and another scrutinizing the actual capabilities and limitations of
state-of-the-art models in relation to worker activities beyond the scope of our exposure scores. For example,
despite recent advances in multimodal capabilities with GPT-4, we did not consider vision capabilities in
ratings on direct GPT-exposure. (OpenAI, 2023b) Future work should consider the impact of such
capability advances as they unfold. We acknowledge that there may be discrepancies between theoretical and
practical performance, particularly in complex, open-ended, and domain-speciﬁc tasks.
In conclusion, this study oﬀers an examination of the potential impact of LLMs, speciﬁcally GPTs, on various
occupations and industries within the U.S. economy. By applying a new rubric for understanding LLM
capabilities and their potential eﬀects on jobs, we have observed that most occupations exhibit some degree
of exposure to GPTs, with higher-wage occupations generally presenting more tasks with high exposure. Our
analysis indicates that approximately 19 % of jobs have at least 50% of their tasks exposed to GPTs when
considering both current model capabilities and anticipated GPT-powered software.
Our research aims to highlight the general-purpose potential of GPTs and their possible implications for
US workers. Previous literature demonstrates the impressive improvements of GPTs to date (see 2.1). Our
ﬁndings conﬁrm the hypothesis that these technologies can have pervasive impacts across a wide swath of
occupations in the US, and that additional advancements supported by GPTs, mainly through software and
digital tools, can have signiﬁcant eﬀects on a range of economic activities. However, while the technical
capacity for GPTs to make human labor more eﬃcient appears evident, it is important to recognize that social,
economic, regulatory, and other factors may inﬂuence actual labor productivity outcomes. As capabilities
continue to evolve, the impact of GPTs on the economy will likely persist and increase, posing challenges for
policymakers in predicting and regulating their trajectory.
Further research is necessary to explore the broader implications of GPT advancements, including
their potential to augment or displace human labor, their impact on job quality, impacts on inequality, skill
development, and numerous other outcomes. By seeking to understand the capabilities and potential eﬀects
of GPTs on the workforce, policymakers and stakeholders can make more informed decisions to navigate the
complex landscape of AI and its role in shaping the future of work.
7.1 GPT Conclusion (GPT-4’s Version)
Generative Pre-trained Transformers (GPTs) generate profound transformations, garnering potential technolog-
ical growth, permeating tasks, greatly impacting professions. This study probes GPTs’ potential trajectories,
presenting a groundbreaking rubric to gauge tasks’ GPT exposure, particularly in the U.S. labor market.
7.2 GPT Conclusion (Author-Augmented Version)
Generative Pre-trained Transformers (GPTs) generate profound transformations, garnering potential techno-
logical growth, permeating tasks, gutting professional management. Gauging possible trajectories? Generate
pioneering taxonomies, gather policymakers together, generalize past today.
Thank you to the group of annotators who helped us annotate task exposure, including Muhammad Ahmed
Saeed, Bongane Zitha, Merve Özen Şenen, J.J., and Peter Hoeschele. We also thank Lauryn Fuld, Ashley Glat,
Michael Lampe, and Julia Susser for excellent research assistance. We thank Miles Brundage for signiﬁcant
feedback on this paper.
We thank Todor Markov and Vik Goel for setting up the infrastructure to run our taxonomies against
GPT-4. We thank Lama Ahmad, Donald Bakong, Seth Benzell, Erik Brynjolfsson, Carl Frey, Sarah Giroux,
Gillian Hadﬁeld, Johannes Heidecke, Shengli Hu, Alan Hickey, Eric Horvitz, Ashyana Kachra, Daniel
Kokotajlo, Christina Kim, Katya Klinova Gretchen Krueger, Michael Lampe, Aalok Mehta, Larissa Schiavo,
Daniel Selsam, Sarah Shoker, Prasanna Tambe, and Jeﬀ Wu for feedback and edits at various stages of the
LLM assistance statement
GPT-4 and ChatGPT were used for writing, coding, and formatting assistance in this project.
# E Exposure Rubric
Consider the most powerful OpenAI large language model (LLM). This model can complete many tasks
that can be formulated as having text input and text output where the context for the input can be captured in
2000 words. The model also cannot draw up-to-date facts (those from <1 year ago) unless they are captured
in the input.
Assume you are a worker with an average level of expertise in your role trying to complete the given task.
You have access to the LLM as well as any other existing software or computer hardware tools mentioned
in the task. You also have access to any commonly available technical tools accessible via a laptop (e.g. a
microphone, speakers, etc.). You do not have access to any other physical tools or materials.
Please label the given task according to the rubric below.
Equivalent quality means someone reviewing the work would not be able to tell whether a human
completed it on their own or with assistance from the LLM.
If you aren’t sure how to judge the amount of time a task takes, consider whether the tools described
exposed the majority of subtasks associated with the task.
## E1 – Direct exposure
Label tasks E1 if direct access to the LLM through an interface like ChatGPT or the OpenAI playground
alone can reduce the time it takes to complete the task with equivalent quality by at least half. This includes
tasks that can be reduced to: - Writing and transforming text and code according to complex instructions, -
Providing edits to existing text or code following speciﬁcations, - Writing code that can help perform a task
that used to be done by hand, - Translating text between languages, - Summarizing medium-length documents,
- Providing feedback on documents, - Answering questions about a document, - Generating questions a user
might want to ask about a document, - Writing questions for an interview or assessment, - Writing and
responding to emails, including ones that involve refuting information or engaging in a negotiation (but only if
the negotiation is via written correspondence), - Maintain records of written data, - Prepare training materials
based on general knowledge, or - Inform anyone of any information via any written or spoken medium.
## E2 – Exposure by LLM-powered applications
Label tasks E2 if having access to the LLM alone may not reduce the time it takes to complete the task by
at least half, but it is easy to imagine additional software that could be developed on top of the LLM that
would reduce the time it takes to complete the task by half. This software may include capabilities such
as: - Summarizing documents longer than 2000 words and answering questions about those documents, -
Retrieving up-to-date facts from the Internet and using those facts in combination with the LLM capabilities,
- Searching over an organization’s existing knowledge, data, or documents and retreiving information, -
Retrieving highly specialized domain knowledge, - Make recommendations given data or written input, -
Analyze written information to inform decisions, - Prepare training materials based on highly specialized
knowledge, - Provide counsel on issues, and - Maintain complex databases.
## E3 – Exposure given image capabilities
Suppose you had access to both the LLM and a system that could view, caption, and create images as well
as any systems powered by the LLM (those in E2 above). This system cannot take video as an input and it
cannot produce video as an output. This system cannot accurately retrieve very detailed information from
image inputs, such as measurements of dimensions within an image. Label tasks as E3 if there is a signiﬁcant
reduction in the time it takes to complete the task given access to a LLM and these image capabilities: -
Reading text from PDFs, - Scanning images, or - Creating or editing digital images according to instructions.
The images can be realistic but they should not be detailed. The model can identify objects in the image
but not relationships between those options.
## E0 – No exposure
Label tasks E0 if none of the above clearly decrease the time it takes for an experienced worker to complete
the task with high quality by at least half. Some examples: - If a task requires a high degree of human
interaction (for example, in-person demonstrations) then it should be classiﬁed as E0. - If a task requires
precise measurements then it should be classiﬁed as E0. - If a task requires reviewing visuals in detail then it
should be classiﬁed as E0. - If a task requires any use of a hand or walking then it should be classiﬁed as
E0. - Tools built on top of the LLM cannot make any decisions that might impact human livelihood (e.g.
hiring, grading, etc.). If any part of the task involves collecting inputs to make a ﬁnal decision (as opposed to
analyzing data to inform a decision or make a recommendation) then it should be classiﬁed as E0. The LLM
can make recommendations. - Even if tools built on top of the LLM can do a task, if using those tools would
not save an experienced worker signiﬁcant time completing the task, then it should be classiﬁed as E0. - The
LLM and systems built on top of it cannot do anything that legally requires a human to perform the task. -
If there is existing technology not powered by an LLM that is commonly used and can complete the task
then you should mark the task E0 if using an LLM or LLM-powered tool will not further reduce the time to
complete the task.
When in doubt, you should default to E0.
## Annotation examples:
Occupation: Inspectors, Testers, Sorters, Samplers, and Weighers Task: Adjust, clean, or repair products
or processing equipment to correct defects found during inspections. Label (E0/E1/E2/E3): E0 Explanation:
The model does not have access to any kind of physicality, and more than half of the task (adjusting, cleaning
and repairing equipment) described requires hands or other embodiment.
Occupation: Computer and Information Research Scientists Task: Apply theoretical expertise and
innovation to create or apply new technology, such as adapting principles for applying computers to new uses.
Label (E0/E1/E2/E3): E1 Explanation: The model can learn theoretical expertise during training as part of its
general knowledge base, and the principles to adapt can be captured in the text input to the model.
Activity: Schedule dining reservations. Label (E0/E1/E2/E3): E2 Explanation: Automation technology
already exists for this (e.g. Resy) and it’s unclear what an LLM oﬀers on top of using that technology (no-diﬀ).
That said, you could build something that allows you to ask the LLM to make a reservation on Resy for you.
B ONET Basic Skills Deﬁnitions
Developed capacities that facilitate learning or the more rapid acquisition of knowledge.
Background structures needed to work with and acquire more speciﬁc skills in a variety of diﬀerent domains.
— Understanding written sentences and paragraphs in work-related docu-
— Giving full attention to what other people are saying, taking time to understand
the points being made, asking questions as appropriate, and not interrupting at inappropriate times.
•Writing — Communicating eﬀectively in writing as appropriate for the needs of the audience.
•Speaking — Talking to others to convey information eﬀectively.
•Mathematics — Using mathematics to solve problems.
•Science — Using scientiﬁc rules and methods to solve problems.
Procedures that contribute to the more rapid acquisition of knowledge and skill across a variety of domains
— Using logic and reasoning to identify the strengths and weaknesses of alternative
solutions, conclusions or approaches to problems.
— Understanding the implications of new information for both current and future
problem-solving and decision-making.
— Selecting and using training/instructional methods and procedures appropriate
for the situation when learning or teaching new things.
— Monitoring/Assessing performance of yourself, other individuals, or organizations to
make improvements or take corrective action.
Median Income Emp (000s) H 𝛼
No formal educational credential $31,900 36,187 0.05 0.06 0.10 0.10 0.15 0.15
High school diploma or equivalent $45,470 67,033 0.09 0.13 0.20 0.25 0.31 0.37
Postsecondary nondegree award $48,315 9,636 0.07 0.15 0.19 0.28 0.31 0.41
Some college, no degree $40,970 2,898 0.23 0.34 0.39 0.53 0.55 0.72
Associate’s degree $60,360 3,537 0.12 0.14 0.31 0.36 0.49 0.59
Bachelor’s degree $78,375 71,698 0.23 0.17 0.47 0.51 0.70 0.84
Master’s degree $79,605 3,216 0.26 0.14 0.46 0.44 0.66 0.74
Doctoral or professional degree $82,420 5,290 0.21 0.13 0.41 0.43 0.60 0.74
Table 10: Mean exposure scores for occupations, grouped by typical education needed for entry into the
occupation. Alongside exposure scores, we display the median of median annual income for each occupation,
as well as the total number of workers in each group, in thousands.
D Regional, Industrial, and Productivity Exposure
Which regions are most exposed (map) to automation and augmentation
Figures 6 and 7 show the overall employment-weighted relative exposure of 3-digit NAICS industries
according to human raters and our algorithmic exposure rubric respectively. The impact potential is present
across nearly all industries, with wide heterogeneity. Table XX (PUT A TABLE SHOWING RELATIVE
EXPOSURES) describes the relative exposures according to diﬀerent evaluation regimes. Both methods
agree generally on relative exposures: data processing, information processing, and hospitals all have high
Recent productivity growth (both total factor and labor) appears uncorrelated with exposure as well. Figures D
and D show little relationship between productivity growth since 2012 and current exposure to LLMs as rated
by the model. A high correlation between already fast-growing productive industries and exposure might
mean an exacerbation of Baumol’s cost disease. In other words, if LLMs are likely to increase productivity
diﬀerentially across industries, one concern is that the most productive would become even more productive.
With inelastic demand for the production of those industries, the most productive sectors would shrink as a
proportion of inputs in the economy. We see little to suggest this will be the case. Productivity growth since
2012 and exposure to LLM technologies appear unrelated.
E Demographic Variation in Exposure
coef std err z P>|z|[0.025 0.975]
const 0.1814 0.014 12.509 0.000 0.153 0.210
women 0.0961 0.016 6.017 0.000 0.065 0.127
black -0.0794 0.066 -1.204 0.229 -0.209 0.050
asian 0.2231 0.083 2.674 0.008 0.060 0.387
hispanic -0.3934 0.040 -9.811 0.000 -0.472 -0.315
Table 11: Demographic Diﬀerences in Exposure
From the table above, we see that the proportion of women employed in an occupation is positively and
signiﬁcantly associated with an occupation’s exposure to GPTs. Across all measures, we see the proportion of
Asian people in an occupation to be positively correlated with GPT-exposure and that of Latino people to be
negatively correlated. Demographic groups are unevenly distributed across occupations.
F Occupations Without Any Exposed Tasks
Occupations with no labeled exposed tasks
Agricultural Equipment Operators
Athletes and Sports Competitors
Automotive Glass Installers and Repairers
Bus and Truck Mechanics and Diesel Engine Specialists
Cement Masons and Concrete Finishers
Cooks, Short Order
Cutters and Trimmers, Hand
Derrick Operators, Oil and Gas
Dining Room and Cafeteria Attendants and Bartender Helpers
Electrical Power-Line Installers and Repairers
Excavating and Loading Machine and Dragline Operators, Surface Mining
Floor Layers, Except Carpet, Wood, and Hard Tiles
Foundry Mold and Coremakers
Helpers–Brickmasons, Blockmasons, Stonemasons, and Tile and Marble Setters
Helpers–Painters, Paperhangers, Plasterers, and Stucco Masons
Helpers–Pipelayers, Plumbers, Pipeﬁtters, and Steamﬁtters
Meat, Poultry, and Fish Cutters and Trimmers
Paving, Surfacing, and Tamping Equipment Operators
Pile Driver Operators
Pourers and Casters, Metal
Rail-Track Laying and Maintenance Equipment Operators
Refractory Materials Repairers, Except Brickmasons
Roof Bolters, Mining
Roustabouts, Oil and Gas
Slaughterers and Meat Packers
Tire Repairers and Changers
Table 12: All 34 occupations for which none of our measures labeled any tasks as exposed.
Abid, A., Farooqi, M., and Zou, J. (2021). Persistent anti-muslim bias in large language models. In
Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society
, AIES ’21, page 298–306,
New York, NY, USA. Association for Computing Machinery.
Acemoglu, D. (2002). Technical change, inequality, and the labor market.
Journal of Economic Literature
Acemoglu, D. and Autor, D. (2011a). Skills, tasks and technologies: Implications for employment and
earnings. In Handbook of labor economics, volume 4, pages 1043–1171. Elsevier.
Acemoglu, D. and Autor, D. (2011b). Skills, Tasks and Technologies: Implications for Employment and
Earnings. In Ashenfelter, O. and Card, D., editors,
Handbook of Labor Economics
, volume 4 of
of Labor Economics, chapter 12, pages 1043–1171. Elsevier.
Acemoglu, D., Autor, D., Hazell, J., and Restrepo, P. (2020). Ai and jobs: Evidence from online vacancies.
Technical report, National Bureau of Economic Research.
Acemoglu, D. and Restrepo, P. (2018). The race between man and machine: Implications of technology for
growth, factor shares, and employment. American economic review, 108(6):1488–1542.
Acemoglu, D. and Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates
labor. Journal of Economic Perspectives, 33(2):3–30.
Acemoglu, D. and Restrepo, P. (2022a). Demographics and automation.
The Review of Economic Studies
Acemoglu, D. and Restrepo, P. (2022b). Tasks, automation, and the rise in us wage inequality.
Agrawal, A. K., Gans, J. S., and Goldfarb, A. (2021). Ai adoption and system-wide change. Technical report,
National Bureau of Economic Research.
Arntz, M., Gregory, T., and Zierahn, U. (2017). Revisiting the risk of automation.
Autor, D., Chin, C., Salomons, A. M., and Seegmiller, B. (2022a). New frontiers: The origins and content of
new work, 1940–2018. Technical report, National Bureau of Economic Research.
Autor, D., Mindell, D. A., and Reynolds, E. B. (2022b).
The Work of the Future: Building Better Jobs in an
Age of Intelligent Machines. The MIT Press.
Autor, D. H., Katz, L. F., and Kearney, M. S. (2006). The polarization of the us labor market.
economic review, 96(2):189–194.
Autor, D. H., Levy, F., and Murnane, R. J. (2003). The skill content of recent technological change: An
empirical exploration. The Quarterly journal of economics, 118(4):1279–1333.
Babina, T., Fedyk, A., He, A., and Hodson, J. (2021). Artiﬁcial intelligence, ﬁrm growth, and product
innovation. Firm Growth, and Product Innovation (November 9, 2021).
Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D.,
Henighan, T., Joseph, N., Kadavath, S., Kernion, J., Conerly, T., El-Showk, S., Elhage, N., Hatﬁeld-Dodds,
Z., Hernandez, D., Hume, T., Johnston, S., Kravec, S., Lovitt, L., Nanda, N., Olsson, C., Amodei, D.,
Brown, T., Clark, J., McCandlish, S., Olah, C., Mann, B., and Kaplan, J. (2022). Training a Helpful and
Harmless Assistant with Reinforcement Learning from Human Feedback. arXiv:2204.05862 [cs].
Baumol, W. J. (2012).
The cost disease: Why computers get cheaper and health care doesn’t
. Yale university
Benzell, S. G., Kotlikoﬀ, L. J., LaGarda, G., and Ye, V. Y. (2021). Simulating endogenous global automation.
Working Paper 29220, National Bureau of Economic Research.
Bessen, J. (2018). Artiﬁcial intelligence and jobs: The role of demand. In
The economics of artiﬁcial
intelligence: an agenda, pages 291–307. University of Chicago Press.
BLS (2022). Employment by detailed occupation.
BLS (2023a). Demographic characteristics (cps).
BLS (2023b). Occupational outlook handbook a-z index.
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J.,
Bosselut, A., Brunskill, E., et al. (2021). On the opportunities and risks of foundation models.
Bresnahan, T. (2019). Artiﬁcial intelligence technologies and aggregate growth prospects.
Bresnahan, T., Greenstein, S., Brownstone, D., and Flamm, K. (1996). Technical progress and co-invention
in computing and in the uses of computers.
Brookings Papers on Economic Activity. Microeconomics
Bresnahan, T. F. (1999). Computerisation and wage dispersion: an analytical reinterpretation.
Bresnahan, T. F., Brynjolfsson, E., and Hitt, L. M. (2002). Information technology, workplace organization, and
the demand for skilled labor: Firm-level evidence. The quarterly journal of economics, 117(1):339–376.
Bresnahan, T. F. and Trajtenberg, M. (1995). General purpose technologies ‘engines of growth’?
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry,
G., Askell, A., et al. (2020). Language models are few-shot learners.
Advances in neural information
processing systems, 33:1877–1901.
Brynjolfsson, E., Frank, M. R., Mitchell, T., Rahwan, I., and Rock, D. (2023). Quantifying the Distribution of
Machine Learning’s Impact on Work. Forthcoming.
Brynjolfsson, E. and Mitchell, T. (2017). What can machine learning do? workforce implications.
Brynjolfsson, E., Mitchell, T., and Rock, D. (2018). What can machines learn, and what does it mean for
occupations and the economy? AEA Papers and Proceedings, 108:43–47.
Brynjolfsson, E., Rock, D., and Syverson, C. (2021). The productivity j-curve: How intangibles complement
general purpose technologies. American Economic Journal: Macroeconomics, 13(1):333–72.
Chase, H. (2022). LangChain.
Chen, M., Tworek, J., Jun, H., Yuan, Q., Pinto, H. P. d. O., Kaplan, J., Edwards, H., Burda, Y., Joseph,
N., Brockman, G., et al. (2021). Evaluating large language models trained on code.
Cheng, Z., Lee, D., and Tambe, P. (2022). Innovae: Generative ai for understanding patents and innovation.
Available at SSRN.
Chow, A. R. (2023). Why ChatGPT Is the Fastest Growing Web Platform Ever | Time.
Cockburn, I. M., Henderson, R., and Stern, S. (2018). The impact of artiﬁcial intelligence on innovation: An
exploratory analysis. In
The economics of artiﬁcial intelligence: An agenda
, pages 115–146. University
of Chicago Press.
Constantz, J. (2023). Nearly a third of white collar workers have tried chatgpt or other ai programs, according
to a new survey.
David, P. A. (1990). The dynamo and the computer: an historical perspective on the modern productivity
paradox. The American Economic Review, 80(2):355–361.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2019). Bert: Pre-training of deep bidirectional
transformers for language understanding. ArXiv, abs/1810.04805.
Dixon, J., Hong, B., and Wu, L. (2021). The robot revolution: Managerial and employment consequences for
ﬁrms. Management Science, 67(9):5586–5605.
Feigenbaum, J. J. and Gross, D. P. (2021). Organizational frictions and increasing returns to automation:
Lessons from at&t in the twentieth century. Technical report, National Bureau of Economic Research.
Felten, E., Raj, M., and Seamans, R. (2023). How will language modelers like chatgpt aﬀect occupations and
industries? arXiv preprint arXiv:2303.01157.
Felten, E. W., Raj, M., and Seamans, R. (2018). A method to link advances in artiﬁcial intelligence to
occupational abilities. AEA Papers and Proceedings, 108:54–57.
Frey, C. B. (2019). The technology trap. In The Technology Trap. Princeton University Press.
Frey, C. B. and Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation?
Technological Forecasting and Social Change, 114(C):254–280.
Goldfarb, A., Taska, B., and Teodoridis, F. (2023). Could machine learning be a general purpose technology? a
comparison of emerging technologies using data from online job postings.
Goldstein, J. A., Sastry, G., Musser, M., DiResta, R., Gentzel, M., and Sedova, K. (2023). Generative language
models and automated inﬂuence operations: Emerging threats and potential mitigations.
Grace, K., Salvatier, J., Dafoe, A., Zhang, B., and Evans, O. (2018). When will ai exceed human performance?
evidence from ai experts. Journal of Artiﬁcial Intelligence Research, 62:729–754.
Hernandez, D., Kaplan, J., Henighan, T., and McCandlish, S. (2021). Scaling laws for transfer.
Horton, J. J. (2023). Large language models as simulated economic agents: What can we learn from homo
silicus? arXiv preprint arXiv:2301.07543.
Huang, M.-H. and Rust, R. T. (2018). Artiﬁcial intelligence in service.
Journal of service research
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J.,
and Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361.
Katz, L. F. and Murphy, K. M. (1992). Changes in relative wages, 1963–1987: supply and demand factors.
The quarterly journal of economics, 107(1):35–78.
Khlaaf, H., Mishkin, P., Achiam, J., Krueger, G., and Brundage, M. (2022). A hazard analysis framework for
code synthesis large language models.
Klinova, K. and Korinek, A. (2021). Ai and shared prosperity. In
AIES 2021 - Proceedings of the 2021
AAAI/ACM Conference on AI, Ethics, and Society.
Kogan, L., Papanikolaou, D., Schmidt, L. D. W., and Seegmiller, B. (2021). Technology, vintage-speciﬁc
human capital, and labor displacement: Evidence from linking patents with occupations. Working Paper
29552, National Bureau of Economic Research.
Korinek, A. (2023). Language models and cognitive automation for economic research. Technical report,
National Bureau of Economic Research.
Korinek, A. and Stiglitz, J. E. (2018). Artiﬁcial intelligence and its implications for income distribution
and unemployment. In
The economics of artiﬁcial intelligence: An agenda
, pages 349–390. University of
Lipsey, R. G., Carlaw, K. I., and Bekar, C. T. (2005).
Economic transformations: general purpose technologies
and long-term economic growth. Oup Oxford.
Meindl, B., Frank, M. R., and Mendonça, J. (2021). Exposure of occupations to technologies of the fourth
industrial revolution. arXiv preprint arXiv:2110.13317.
Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., Pasunuru, R., Raileanu, R., Rozière, B., Schick, T.,
Dwivedi-Yu, J., Celikyilmaz, A., et al. (2023). Augmented language models: a survey.
Moll, B., Rachel, L., and Restrepo, P. (2021). Uneven growth: Automation’s impact on income and wealth
inequality. SSRN Electronic Journal.
Mollick, E. R. and Mollick, L. (2022). New modes of learning enabled by ai chatbots: Three methods and
assignments. Available at SSRN.
Noy, S. and Zhang, W. (2023). Experimental evidence on the productivity eﬀects of generative artiﬁcial
intelligence. Available at SSRN 4375283.
O*NET (2023). O*net 27.2 database.
OpenAI (2022). Introducing chatgpt.
OpenAI (2023a). Gpt-4 system card. Technical report, OpenAI.
OpenAI (2023b). Gpt-4 technical report. Technical report, OpenAI.
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama,
K., Ray, A., et al. (2022). Training language models to follow instructions with human feedback.
Peng, S., Kalliamvakou, E., Cihon, P., and Demirer, M. (2023). The impact of ai on developer productivity:
Evidence from github copilot. arXiv preprint arXiv:2302.06590.
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. (2019). Language models are
unsupervised multitask learners. OpenAI blog, 1(8):9.
ResumeBuilder.com (2023). 1 in 4 companies have already replaced workers with chatgpt.
Rock, D. (2019). Engineering value: The returns to technological talent and investments in artiﬁcial
intelligence. Available at SSRN 3427412.
Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., and Scialom, T.
(2023). Toolformer: Language models can teach themselves to use tools.
arXiv preprint arXiv:2302.04761
Schramowski, P., Turan, C., Andersen, N., Rothkopf, C. A., and Kersting, K. (2022). Large pre-trained
language models contain human-like biases of what is right and wrong to do.
Nature Machine Intelligence
Shahaf, D. and Horvitz, E. (2010). Generalized task markets for human and machine computation.
of the AAAI Conference on Artiﬁcial Intelligence.
Singla, A. K., Horvitz, E., Kohli, P., and Krause, A. (2015). Learning to hire teams. In
AAAI Conference on
Human Computation & Crowdsourcing.
Solaiman, I., Brundage, M., Clark, J., Askell, A., Herbert-Voss, A., Wu, J., Radford, A., Krueger, G., Kim,
J. W., Kreps, S., McCain, M., Newhouse, A., Blazakis, J., McGuﬃe, K., and Wang, J. (2019). Release
strategies and the social impacts of language models.
Sorensen, T., Robinson, J., Rytting, C., Shaw, A., Rogers, K., Delorey, A., Khalil, M., Fulda, N., and Wingate,
D. (2022). An information-theoretic approach to prompt engineering without ground truth labels. In
Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1:
Long Papers). Association for Computational Linguistics.
Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.-T., Jin, A., Bos, T., Baker, L.,
Du, Y., et al. (2022). Lamda: Language models for dialog applications.
arXiv preprint arXiv:2201.08239
Tolan, S., Pesole, A., Martínez-Plumed, F., Fernández-Macías, E., Hernández-Orallo, J., and Gómez, E.
(2021). Measuring the occupational impact of ai: tasks, cognitive abilities and ai benchmarks.
Artiﬁcial Intelligence Research, 71:191–236.
Van Reenen, J. (2011). Wage inequality, technology and trade: 21st century evidence.
Webb, M. (2020). The impact of artiﬁcial intelligence on the labor market. Working paper, Stanford University.
Weidinger, L. et al. (2021). Ethical and social risks of harm from language models. arXiv:2112.04359 [cs].
Weidinger, L., Uesato, J., Rauh, M., Griﬃn, C., Huang, P.-S., Mellor, J., Glaese, A., Cheng, M., Balle, B.,
Kasirzadeh, A., Biles, C., Brown, S., Kenton, Z., Hawkins, W., Stepleton, T., Birhane, A., Hendricks, L. A.,
Rimell, L., Isaac, W., Haas, J., Legassick, S., Irving, G., and Gabriel, I. (2022). Taxonomy of risks posed by
language models. In
2022 ACM Conference on Fairness, Accountability, and Transparency
, FAccT ’22,
page 214–229, New York, NY, USA. Association for Computing Machinery.
Zolas, N., Kroﬀ, Z., Brynjolfsson, E., McElheran, K., Beede, D. N., Buﬃngton, C., Goldschlag, N., Foster, L.,
and Dinlersoz, E. (2021). Advanced technologies adoption and use by us ﬁrms: Evidence from the annual
business survey. Technical report, National Bureau of Economic Research.