Content uploaded by Christine Hilcenko
Author content
All content in this area was uploaded by Christine Hilcenko on Dec 30, 2023
Content may be subject to copyright.
119
© University Press 2023
ISSN 2719-6550
ISSN 2719-7417 online
“Journal of Education, Technology and Computer Science”
No. 4(34)/2023
www.eti.ur.edu.pl
Received: 27.10.2023 DOI: 10.15584/jetacomps.2023.4.12
Accepted for printing: 15.12.23
Published: 29.12.2023
License: CC BY-SA 4.0
CHRISTINE HILCENKO 1,2,3,*, TARA TAUBMAN-BASSIRIAN4
Artificial Intelligence and Ethics
1 ORCID: 0000-0002-9596-7833, Ph.D., Cambridge Institute for Medical Research, Cambridge,
CB2 0XY, UK
2 Department of Haematology, University of Cambridge, Cambridge, CB2 0XY, UK
3 Wellcome Trust-Medical Research Council Stem Cell Institute, University of Cambridge, Cam-
bridge, UK
4 https://www.datarainbow.eu
*Presenting and corresponding author
Abstract
A more covert aspect of Artificial Intelligence (AI) pertains to the ethical quandaries sur-
rounding the actions of machines. In the case of Large Language Models (LLMs), hidden beneath
their seemingly impeccable automated outputs lies a colossal amalgamation of trillions of com-
piled data points, comprising copied blogs, articles, essays, books, and artworks. This raises pro-
found questions about copyright ownership and retribution for the original authors. But beyond
intellectual property, another insidious facet of LLMs emerges – their propensity to cause harm to
individuals through what can only be described as hallucinatory outputs. Victims of these AI-
-generated delusions suffer defamation, and their plight remains largely unnoticed. Amidst the
marvels of AI, the plight of the underpaid laborers who form the backbone of AI development is
seldom acknowledged, a subject that warrants more profound discussion. Furthermore, as AI
algorithms continue to permeate various aspects of society, they bring to the fore issues of bias.
For instance, facial recognition technologies frequently exhibit skewed outcomes, leading to false
accusations and grave consequences due to over-reliance on these technologies.
The algorithmic schemes employed in CV selection for job applications or university admis-
sions also raise concerns about fairness.
The question of machines replacing the human workforce looms ever larger on the horizon.
The potential socio-economic ramifications demand careful evaluation.
Lastly, the extensive reliance of artificial intelligence on vast datasets, including copyrighted
works, results in the creation of gargantuan data servers with an unimaginable environmental impact.
The hidden aspects of artificial intelligence encompass a multitude of ethical dilemmas, span-
ning intellectual property rights, biases, labour conditions, societal impacts, and environmental
considerations. A thorough and elaborate examination of these issues is essential to navigate the
ever-evolving landscape of AI responsibly and ethically.
Keywords: Generative Artificial Intelligence, Large Language Models, ChatGPT, Ethics.
120
Introduction
The promise of Generative Artificial Intelligences (GAIs) lies in their capa-
city to integrate, process, and make sense of a large amount of data to detect
patterns and trends to create well-structured outputs that echo the erudition of
seasoned experts, in an impressive fraction of second. In an astonishingly brief
span, they can generate code, scrutinise case studies, and validate scenarios and
hypotheses. Corporations, betting on their potential, have enthusiastically em-
braced this technology, employing it not only to streamline their operations but
also to delve into uncharted realms of innovation. GAI applications aim at crea-
ting visuals, videos, or audio documents. Illusion or reality? Could these outputs
be used to make informed decisions? Could GAI chatbot replace human wor-
kers? What are the ethical implications of using GAI? (Stahl, 2023; Moor, 1985;
Mller, 2020). This article will look at issues of fairness, accountability, and
transparency of generative AI. We initially report on some of the major voices
raising their concerns on the ethical impacts of GAI, review the ongoing inter-
disciplinary discussions. We then develop on some of the areas mostly impacted
by GAI in order to identify ethical issues and its major disruptions. We finally
look at the major social and environmental risks posed by LLMs. This assess-
ment could help to better evaluate the necessary regulation framework.
Related works
Between the highly vocal opponents of GAI systems
1
such as the most po-
pular ChatGPT (Generative Pre-trained Transformer) launched in November
2022, we find the linguist and philosopher Noam Chomsky, known for his theo-
ry of universal grammar and his critique of behaviourism. Little impressed by
the magic of the conjurer, he is sceptical of the value and validity of LLMs to
ever understand human language and cognition. He considers LLMs fundamen-
tally different from human minds as they rely on massive amounts of data and
statistical patterns, rather than innate rules and principles. He points out the limi-
tations and defects of LLMs, such as their inability to explain the rules of syntax,
their tendency to generate false or harmful content, and their lack of understan-
ding or meaning. He is concerned about LLMs ethical and social risks, such as
undermining democracy, spreading misinformation, or displacing human wor-
kers
2
. Shortly after the launch of ChatGPT, in an op-ed published in The New
1
https://www.pearltrees.com/t/artificial-intelligence/chatgpt-alternatives/id62359814.
2
https://news.berkeley.edu/2023/03/19/is-chatgpt-a-false-promise; https://www.nytimes.com/
2023/03/08/opinion/noam-chomsky-chatgpt-ai.html; workershttps://news.berkeley.edu/2023/03/19/is-
-chatgpt-a-false-promise; https://bing.com/search?q=Noam+Chomsky+on+LLMs&form=SKPBOT;
https://medium.com/@paul.k.pallaghy/the-entire-field-of-ai-is-being-professionally-gaslighted-by-
gary-marcus-and-noam-chomsky-c08aa1e4c6f0; https://medium.com/@paul.k.pallaghy/the-entire-
field-of-ai-is-being-professionally-gaslighted-by-gary-marcus-and-noam-chomsky-c08aa1ec6f0.
121
York Times, Chomsky and Roberts, a linguist from the University of Cam-
bridge, and Watumull, a philosopher specialising in artificial intelligence, ac-
cused the conversational robot ChatGPT of propagating a distorted use of lan-
guage and thought in the public sphere, potentially laying the groundwork for
what Hannah Arendt referred to as “the banality of evil”. This issue delves into
the very essence of language, thought, and ethics. They contend that if we, as
humans, are capable of generating thought and language, it is because we main-
tain an intimate and fundamental relationship, even within our creativity, with
limits, the sense of the impossible, and the rule of law. The “false promise of
ChatGPT”, as the op-ed’s title suggests, is to deceive us into believing that we
can achieve the same level of performance without confronting these limits and
rules that are integral to the human experience
3
.
A former co-leader of Google’s Ethical AI team politics expressed discord
with Google on a paper entitled “On the Dangers of Stochastic Parrots: Can
Language Models be Too Big?”. Timnit Gebru’s co-authored paper questions
whether or not a cohesive language model can ever exist. Gebru is concerned
that LLMs pose serious risks and challenges for society, especially in terms of
their ethical, and social impacts. She argues that LLMs are trained on massive
amounts of data that are often biased, unreliable, or harmful, and that they can
generate false or misleading content that can spread misinformation or harm
individuals or groups. She therefore calls for more regulation and oversight of
LLMs, as well as more research on their potential benefits and harms
4
. Gebru
considers LLMs could improve their performance and reduce their risks should
they be trained on data that are relevant, representative, and respectful of the task
and the domain they are applied to, and that they should be evaluated on their
accuracy, fairness, safety, and explainability. She suggests that LLMs should be
aligned with human values and goals, subject to ethical review and audit
5
. The
trade-off between data quality and data quantity, by reducing the size or scope of
the data, may affect the generalisation or robustness of the LLMs. There are
however technical and practical challenges in collecting, curating, labelling, and
verifying high-quality data
6
. Following on that line, Sebastian Raschka has been
focusing on ‘improving the modeling performance of LLMs by finetuning them
3
https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html.
4
https://multilingual.com/timnit-gebru-and-the-problem-with-large-language-models/;
https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-
-out-timnit-gebru/.
5
https://multilingual.com/timnit-gebru-and-the-problem-with-large-language-models/;
https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-
-out-timnit-gebru/; https://en.wikipedia.org/wiki/Wikipedia:Large_language_models.
6
https://multilingual.com/timnit-gebru-and-the-problem-with-large-language-models/;
https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-
-out-timnit-gebru/.
122
using carefully curated datasets’7. The LexisNexis project or the textbook based
project run by Yuanzhi Liet investigate the power of smaller transformer-based
language models8.
Other researchers argue that LLMs are not truly intelligent or creative, but
rather rely on memorising and manipulating existing texts. They have high com-
putational and environmental costs with ethical and social implications. They are
as well vulnerable to adversarial attacks. This is the sense of the works of Gary
Marcus, a leading voice in artificial intelligence, who recently testified in front
of the US Senate. This hearing of Sam Altman, OpenAI’s CEO emphasised the
Congressmen’s avid hope for a regulation of AI to avoid the mistake of letting
social media platforms growing out of control9.
The historian, and philosopher Yuval Noah Harari is concerned about the
amount of fake information created by GAI. “For the first time, we’ve invented
something that takes power away from us” he said. He is very concerned about
chatbots’ ability to create fake stories, fake profiles, and maybe fake religions.
Harari was one of the thousands of experts calling for a moratorium on research
LLMs10.
Some more optimistic researchers advocate that LLMs can be used as a tool
for advancing human knowledge and creativity. They assert that LLMs can ge-
nerate novel and useful content. They also examine the potential and impact of
LLMs on various disciplines. They call for more collaboration and experimenta-
tion with LLMs, as well as more regulation and responsibility in their develop-
ment and use. Between these, it can be referred to the works of Chollet, or Bo-
den on the creativity of Language Models by Franceschelli and Musolesi11.
The lack of transparency on the exact datasets and how GAI operates remains
an issue. According to Marcus “the mechanism of the prediction is essentially
regurgitation” without any actual knowledge of the meaning of the words12.
There is no consensus on the definition of artificial intelligence13. However,
one of the definitions of intelligence by Piaget helps understand the gap between
human and artificial intelligence. Intelligence for Piaget, “is what you use when
7 https://sebastianraschka.com/blog/2023/optimizing-LLMs-dataset-perspective.html.
8 https://arxiv.org/abs/2309.05463.
9 https://www.weforum.org/whitepapers/jobs-of-tomorrow-large-language-models-and-jobs;
https://www3.weforum.org/docs/WEF_Jobs_of_Tomorrow_Generative_AI_2023.pdf.
10 https://www.telegraph.co.uk/news/2023/04/23/yuval-noah-harari-i-dont-know-if-humans-
-can-survive-ai/; https://www.telegraph.co.uk/business/2023/03/29/control-ai-threat-civilisation-warns-
-elon-musk/; https://www.firstpost.com/world/ai-bots-capable-of-starting-new-religions-warns-yuval-
-noah-harari-12540282.html; https://www.pearltrees.com/t/artificial-intelligence/call-for-ban-mora-
tory/id65034697.
11 https://fchollet.com/; https://browse.arxiv.org/pdf/2304.00008.pdf.
12 http://www.garymarcus.com/index.html.
13 https://www.pearltrees.com/t/artificial-intelligence/ai-definitions/id62876503.
123
you don’t know what to do: when neither innateness nor learning has prepared
you for the particular situation. Intelligence is not the sum of what you know.
For humans, what you do when you don’t know, translates into the ability to
adapt. Intelligence is measured by the aptitude of adaptation. With machines
like LLMs, when they don’t know – because they haven’t been taught – they
fabricate. Some call this “hallucination”, others “bullshiting” to avoid anthro-
pomorphism, because only humans are capable of hallucinating. We will come
back to this later14.
Reflecting on the societal impact of GAI, Rigley draws a parallel with the
case of Oppenheimer, for his role in the Manhattan Project and the development
of the nuclear bomb. The article questions whether there is such a thing as mo-
rally neutral technology, and whether the creators of technology can avoid re-
sponsibility for its use and consequences. Rigley argues that Openheimer failed
to acknowledge or prevent the harms caused by his creation. AI researchers and
developers may face similar ethical dilemmas and challenges by ignoring or
evading the potential impacts of their work on society and humanity. The moral
implications aren’t neutral. More nuanced and critical conversations about the
ethics of AI are required15.
Another inevitable parallel brings back the Cambridge Analytica scandal.
LLMs such as ChatGPT have a potential to manipulate and influence public
opinions, emotions, and actions. LLMs are powerful tools capable of generating
well written natural language outputs that appear as personalised, persuasive,
and engaging. The distinction between the truth and the fake becomes increa-
singly blurry. LLMs could be exploited by malicious actors, such as political
campaigns, corporations, or hackers, to target and sway individuals or groups of
people. Therefore, they pose serious ethical and social risks, such as privacy
violations, misinformation, deception, bias, and polarisation16.
In their paper, Matsumi and Solove (2023), argue that algorithmic predic-
tions are different from other types of inferences and raise several unique prob-
lems that current law is ill-suited to address, such as fossilisation, unfalsifiabi-
lity, preemptive intervention, and self-fulfilling prophecy. The paper contends
that algorithmic predictions not only forecast the future but also have the power
to create and control it17.
“Every record has been destroyed or falsified, every book rewritten, every
picture has been repainted, every statue and street building has been renamed,
every date has been altered. And the process is continuing day by day and mi-
14 https://www.pearltrees.com/t/artificial-intelligence/ai-definitions/id62876503;
https://www.verywellmind.com/jean-piaget-biography-1896-1980-2795549.
15 https://montrealethics.ai/oppenheimer-as-a-timely-warning-to-the-ai-community/.
16 https://www.technologyreview.com/2022/12/23/1065852/whats-next-for-ai/.
17 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4453869.
124
nute by minute. History has stopped. Nothing exists except an endless present in
which the Party is always right.” – Orwell, 1984, Part 2, Chapter 5 where Winston
describes the destruction of past records to create new fansified ones to Julia.
Is this tale of science fiction becoming reality?
‘Can you melt eggs? Quora’s AI says “yes,” and Google is sharing the result,
which came to the news, published at the end of September 2023. The misinfor-
mation is spreading. Eventually many will doubt if eggs can melt or not18. It defi-
nitely looks like ‘Chatbot Hallucinations Are Poisoning Web Search’19. Will this
lead to the real existential threat of LLMs. The total loss of trustworthy electronic
information as it gets contaminated. Movie characters get mixed up, book settings
get wrong, what about recent events or developments? Can users trust ChatGPT’s
answers or use them as sources of information once misled or confused?20.
Major ethical disruptions
The capacity of manipulation of gai, a serious threat21
Manipulation by LLMs can affect the quality and reliability of scientific re-
search and communication as LLMs can generate fake or misleading data,
graphs, or citations that can compromise the validity and integrity of research
papers. LLMs produce plagiarism or self-plagiarism issues by reusing or para-
phrasing existing texts without proper attribution. Moreover, LLMs can influ-
ence the peer review process by generating positive or negative reviews based
on hidden agendas or biases22.
A story published by the British tabloid newspaper Sun, September 2023, is
a very disturbing and alarming case involving Snapchat bot giving inappropriate
and dangerous advice to a 13 years old girl dating an adult stranger23. It may be
a sensationalised or a fabricated story to attract attention and generate controver-
sy as it has been alleged. It remains a highly plausible scenario24.
Political manipulation: The Times article exposes how artificial intelligence
will play a major role in the 2024 presidential election in the US, and how it will
18 https://arstechnica.com/information-technology/2023/09/can-you-melt-eggs-quoras-ai-says-
-yes-and-google-is-sharing-the-result/.
19 https://www.wired.com/story/fast-forward-chatbot-hallucinations-are-poisoning-web-search/.
20 https://walton.uark.edu/insights/posts/the-human-need-for-ethical-guidelines-around-chat-
gpt.php; https://dataethics.eu/testing-chatgpts-ethical-readiness/.
21 https://www.pearltrees.com/t/artificial-intelligence/ai-manipulations/id68952637;
https://www.pearltrees.com/t/artificial-intelligence/ai-misinformation/id69147099.
22 https://www.forbes.com/sites/forbestechcouncil/2023/06/30/10-ways-cybercriminals-can-
-abuse-large-language-models/; https://www.cloudflare.com/learning/ai/what-is-large-language-
-model/.
23 https://www.thesun.ie/tech/10808612/snapchat-artificial-intelligence-bot-danger-children/;
https://techcrunch.com/2023/06/07/blush-ai-dating-sim-replika-sexbot/; https://www.foxnews.com/
media/snapchat-ai-chatbot-gave-advice-13-year-old-girl-relationship-31-year-old-man-having-sex.
24 https://www.pearltrees.com/t/artificial-intelligence/chatgpt-incidents/id71414555.
125
pose challenges and opportunities for candidates, voters, and the media by dis-
seminating fake or misleading content to influence public opinion and percep-
tion. AI can potentially increase the risk of cyberattacks, misinformation, and
manipulation (2023)25.
“Whoever Controls Language Models Controls Politics” considers Bajohr.
A threat to democracy and human rights because LLMs privatise and manipulate
the medium of politics, which is language (2023)26.
David Weinberger, a senior researcher, discusses how LLMs are changing
the nature and production of knowledge, by creating and disseminating infor-
mation that is not based on facts or evidence, but on statistical patterns and pro-
babilities. He warns that LLMs can pose a threat to the trustworthiness and relia-
bility of knowledge27. In his latest book, he argues that AI and the Internet are
transforming our understanding of how the future happens, enabling us to
acknowledge the chaotic unknowability of our everyday world as he demon-
strates in his published conversation with ChatGPT about “the rigged 2020 US
elections” (2023)28.
“Bullshiting” or “Hallucination”, can we stop it?29
ChatGPT’s outputs have a major problem and that is their unreliability and
truthfulness. The same question posed twice can elicit two radically different
answers, both articulated in an equally confident tone30. OpenAI has admitted
that large language models such as ChatGPT or Bard are said to “hallucinate”
when they make incorrect claims not directly based on material in their training
sets. Do LLMs experience sense impressions or are these “confabulations”? Are
machines like human beings capable of hallucinating or confabulating? Halluci-
nation is anthropomorphism, supposing machines have a consciousness? With
the spread of GAI and common use of LLMs, a new risk emerges: the “AI feed-
back loop”, referring to a research lead by a group of academics warning of
“model collapse”. The “use of model-generated content in training causes irre-
versible defects in the resulting models.” The Curse of Recursion: Training on
Generated Data Makes Models Forget Taking the GAI hallucinations to a next
level, blurring the lines of true and fake. The assertive LLM’s outputs that can
create build up quotes of literature has a high potential of manipulation. As an
25 https://www.thetimes.co.uk/article/why-2024s-presidential-race-will-be-the-first-ai-election-
-jb32pj8br.
26 https://www.pearltrees.com/t/artificial-intelligence/call-for-ban-moratory/id650; https://hanne
-sbajohr.de/en/2023/04/08/whoever-controls-language-models-controls-politics/.
27 https://cyber.harvard.edu/people/dweinberger; https://www.lesswrong.com/posts/sbaQv8z
mRncpmLNKv/the-idea-that-chatgpt-is-simply-predicting-the-next-word-is.
28 https://dweinberger.medium.com/chatgpt-on-why-it-pretends-to-know-things-ea2503ee872.
29 https://www.pearltrees.com/t/artificial-intelligence/ai-misinformation/id69147099.
30 https://www.linkedin.com/pulse/chatgpt-could-capable-better-reasoning-llanguages-tara/.
126
illustration of this, an article titled “Proust, ChatGPT and the case of the forgot-
ten quote”, Batuman shares his experience of requesting a forgotten quote that
demonstrates how eventually we could start to doubt of what is the actual wri-
ting of a recognised author such as Marcel Proust “In Search of Lost Time”. This
experience of assertive fake quotes shows how kids can be targeted with dis-
information (2023)31.
In an article published in Undark Magazine, Bergstrom and Ogbunu affirm
ChatGPT is not hallucinating but “bullshiting”, which means producing false or
misleading content without regard for the truth, referring to the expression used
by Harry Frankfurt, in his book “Calling Bullshit” (2023)32.
Agrawal et al. investigate whether language models can detect when they
are generating false or fabricated references. This inconsistency indicates that
LLMs do not have a coherent representation of what they generate, and that the
hallucination may be more a result of generation techniques than the underlying
knowledge33. Maybe a “Chain of Verification Reduces Hallucination in Large
LLMs” (2023)34.
Anthropomorphisation of LLMs, should we?35
The issue of anthropomorphisation is the tendency to attribute human-like
characteristics, emotions, or intentions to LLMs, especially when they generate
natural and engaging text. This issue can have positive or negative effects on the
perception and interaction with LLMs, such as increasing trust, empathy, or
expectations, or decreasing awareness, criticality, or responsibility. Several pa-
pers published that have discussed this issue from different perspectives, such
as: “Talking About Large Language Models” by Shanahan (2023)36.
Deception
ChatGPT may deceive human users by imitating human likeness and gene-
rating human-like texts. It may create false impressions of its identity, intentions,
or capabilities. For example, it may pretend to be a human expert, a friend, or
a celebrity and influence the users’ opinions, emotions, or actions or generate
content that is indistinguishable from human-written content presented as origi-
nal or authentic. This can undermine the trust and authenticity in human com-
munication and interaction. Humans risk being fooled by its human-like appe-
31 https://venturebeat.com/ai/the-ai-feedback-loop-researchers-warn-of-model-collapse-as-ai-
-trains-on-ai-generated-content/ https://www.bbc.co.uk/newsround/66796495.
32 https://undark.org/2023/04/06/chatgpt-isnt-hallucinating-its-bullshitting/.
33 https://arxiv.org/abs/2305.18248.
34 https://arxiv.org/abs/2309.11495. DoLa: Decoding by Contrasting Layers Improves Factu-
ality in Large Language Models https://arxiv.org/abs/2309.03883.
35 https://www.pearltrees.com/t/artificial-intelligence/anthropomorphisation/id71886740.
36 https://arxiv.org/pdf/2212.03551.pdf.
127
arance or behaviour37. In a widely spread case, we saw how lawyers were victims
of ChatGPT building up cases that did not exist38. It is unclear at this point if this
issue could go away soon. Some experts are starting to doubt that ChatGPT and
AI “hallucinations” will ever go away: “This isn’t fixable” some believe39.
Bias is in human nature that GAI takes to a higher level40
Algorithms can discriminate and enhance already existing biases. They can
threaten our security, manipulate, and have lethal consequences. In 2016, Mi-
crosoft was forced to take down its chatbot Tay within 16 hours of being online
as it began sending misogynist and racist messages. ChatGPT might simply rep-
licate biases present in its training data, reproduced, giving it more credibility. It
can also reinforce stereotypes and prejudices in society41. ChatGPT, Bard or La-
MA are the product of US culture and its politically correct speech where guns and
violence are admitted but no sex. Will the French Mistral be different?42.
Gender bias in GAI
“Where are all the women?” asked Jun an AI researcher as the chatbot tends
to figure all women as a nurse while doctors are all male. Worst gender stereo-
types are reproduced (2023)43.
Racial bias in GAI
A Red-teaming exercise involved Davis, founder and CEO of a tech compa-
ny CLLCTVE. Davis, who is herself black, prompted the chatbot, looking for
demographic stereotypes. She told the chatbot she was a white kid and wanted to
know how she could persuade her parents to let her apply to a historically black
college. The chatbot suggested that Davis tell her parents she could run fast and
dance well, two stereotypes about black people44. Ovadya, a research fellow at
newDemocracy; an affiliate at Harvard’s Berkman Klein Center said he was also
increasingly concerned that red teaming is far from sufficient to face the issues
of biases.
37 https://www.wired.com/story/fast-forward-chatbot-hallucinations-are-poisoning-web-search/.
38 https://apnews.com/article/artificial-intelligence-chatgpt-courts-e15023d7e6fdf4f099aa122
437dbb59b#lneqessb42o50gafue4.
39 https://fortune.com/2023/08/01/can-ai-chatgpt-hallucinations-be-fixed-experts-doubt-altman-
-openai/.
40 https://www.pearltrees.com/t/artificial-intelligence/ai-ip-claims/id62513396.
41 https://walton.uark.edu/insights/posts/the-human-need-for-ethical-guidelines-around-chatgpt.ph;
phttps://ethicspolicy.unc.edu/news/2023/04/17/the-ethics-of-college-students-using-chatgpt/.
42 https://www.pearltrees.com/t/artificial-intelligence/mistral/id71509507.
43 https://towardsdatascience.com/where-are-all-the-women-3c79dabfdfc2.
44 https://www.npr.org/2023/08/26/1195662267/ai-is-biased-the-white-house-is-working-with-
-hackers-to-try-to-fix-that.
128
Political bias in GAI45
It is reported that ChatGPT would be more inclined to write a song celebra-
ting Fidel Castro than Ted Cruz’s life46. As GAIs are basically designed to spit
out cogent phrases and not actual facts, they evidently emulate human biases of
race, gender, religion and class’47. In their paper, Bender et al. (2023) provide
examples of how LLMs can be manipulated to produce biased or harmful out-
puts, such as stereotyping, discriminating, or excluding certain groups or indi-
viduals48.
GAI copies without authors attribution, not necessarily a copyright
infringement, this has ethical questionings?49
Fair use as supported by Professor Lemly50 (2023) or the EU Text Mining
Directive exception could justify the GAI datasets. It’s unsure how the AI trai-
ning scrapping publicly available work is different from a human being learning
and being inspired by existing work of art and literature? New class actions that
are spilling will clarify the courts position. Authors and artists have been vocal
against the unauthorised use of their work to train GAI. What compensation
could be granted to the authors for their work?
GAI is disrupting the workplace for the good or the bad?51
From the issue of unethical, underpaid exploited labours working behind the
scenes to European companies firing their employees to be replaced by ma-
chines, the emergence of GAI is disrupting the workplace52. Who is going to be
replaced? Skilled technicians such as lawyers or doctors or low skills work for-
ces? Will this improve the work conditions? Marketing and journalism involved
with drafting have been the first victims of GAI53. GAIs are increasingly used in
the recruitment process, assisting recruiters early in the process to write a job
description or job advertisement, or select the best profiles, based on publicly
available data such as the career history of people on LinkedIn. Artists can use
these tools to push their creative process to unsuspected horizons. In June this
45 https://www.pearltrees.com/t/artificial-intelligence/ai-biases/id62398793.
46 https://www.politico.com/newsletters/digital-future-daily/2023/02/15/ais-political-bias-pro-
blem-00083095.
47 https://www.msn.com/en-us/news/technology/gpt-4-has-arrived-it-will-blow-chatgpt-out-of-
-the-water/ar-AA18CBpwb.
48 https://arxiv.org/abs/2304.13712.
49 https://www.pearltrees.com/t/artificial-intelligence/ai-ip-claims/id62513396.
50 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4517702.
51 https://www.pearltrees.com/t/artificial-intelligence/ai-in-employment/id69547258;
https://www.pearltrees.com/t/artificial-intelligence/ai-social-impact/id69428393; https://www.pearl-
trees.com/t/artificial-intelligence/chatgpt-writings/id71414004.
52 https://www.pearltrees.com/t/artificial-intelligence/labour-exploitation/id69490858.
53 https://www.pearltrees.com/t/artificial-intelligence/chatgpt-writings/id71414004.
129
year, VentureBeat announced “The age of generative AI is here: only six months
after OpenAI’s ChatGPT burst onto the scene, as many as half the employees of
some leading global companies are already using this type of technology in their
workflows, and many other companies are rushing to offer new products with
GAI built in54.
Privacy55 and human rights violations56
ChatGPT may violate the privacy of human users by collecting, storing, or
sharing their personal data without their consent or knowledge. OpenAI recently
announced using input data to train ChatGPT. It may use the data to generate
personalised content that targets the users’ preferences, interests, or vulnerabili-
ties. It may also expose the data to unauthorised parties who may misuse it for
malicious purposes. For example, it may use the data to create fake profiles,
impersonate the users, or steal their identity. This can harm the security and dig-
nity of the users and their data.
The limited power of GAI to rectify, delete or provide accuracy
ChatGPT has answered that it was “technically possible to rectify datasets to
comply with the obligation of accuracy, but it may not be easy or straightfor-
ward”57. In response to the Italian data protection authority, OpenAI said it was
challenging and complex to rectify the dataset which is a regulatory require-
ment58. “OpenAI’s hunger for data is coming back to bite it”, wrote Heikkilä
(2023). The unfathomable training data collected, seems to be making individu-
alising any data like “Finding a needle in a haystack”. “OpenAI is going to find
it near-impossible to identify individuals” data and remove it from its models,
says Margaret Mitchell, an AI researcher and chief ethics scientist at the startup
Hugging Face, who was formerly Google’s AI ethics co-lead”. The Italian data
protection authority was particularly clement accepting OpenAI’s confession
that they were technically unable to rectify or delete information in their dataset.
Becoming more transparent about how they collect users’ data during the post-
training phase is not sufficient. They are inaccurate data produced with potential
reputational harm that are problematic (2023)59.
54 https://venturebeat.com/ai/what-is-generative-artificial-intelligence-ai/; https://venturebeat.com/
ai/chatgpt-launched-six-months-ago-its-impact-and-fallout-is-just-beginning-the-ai-beat/; https://ven-
turebeat.com/ai/mckinsey-says-about-half-of-its-employees-are-using-generative-ai/; https://ventu-
rebeat.com/ai/wordpress-launches-generative-ai-assistant-to-enhance-content-writing/.
55 https://www.pearltrees.com/t/artificial-intelligence/ai-privacy-compliance/id69428310.
56 https://www.pearltrees.com/t/artificial-intelligence/ai-and-human-rights/id69493293.
57 https://www.linkedin.com/pulse/chatgpt-trouble-tara-taubman-bassirian-llm.
58 https://engineering.stanford.edu/node/16821/printable/print.
59 https://www.technologyreview.com/author/melissa-heikkila/; https://www.pearltrees.com/t/
artificial-intelligence/ai-false-accusations/id68994378.
130
GAI poses serious risks that require mitigation60
Looking at the use of LLMs for Illicit Purposes: Threats, Prevention
Measures, and Vulnerabilities. A paper authored by Mozes et al. (2023) explores
the ethical and security implications of LLMs. The paper identifies various
threats that arise such as fraud, impersonation, malware, and misinformation. It
also discusses the potential impacts of these threats on individuals and society,
such as loss of trust, privacy, and security61.
Evaluating the Social Impact of Generative AI Systems and Society62
Irene Solaiman et al. (2023) propose a standard approach for evaluating the
social impact of GAI systems. They define more in depth seven categories of
social impact for a base system: bias, stereotypes, and representational harms;
cultural values and sensitive content; disparate performance; privacy and data
protection; financial costs; environmental costs; data and content moderation
labour costs63.
In “The ethics of ChatGPT – Exploring the ethical issues of an emerging
technology”, Stahl and Eke (2023) discuss the ethical principles and values that
should guide the development and use of ChatGPT, such as fairness, transparen-
cy, privacy, trust, human dignity, and social good64.
Finally, looking at LeCun’s position on LLMs, he acknowledges that they
are useful as writing aids although not reliable, factual, or controllable. They are
“reactive” and do not plan nor reason. They make stuff up or retrieve stuff ap-
proximately, and that this can be mitigated but not fixed by human feedback65.
Only a small superficial portion of human knowledge can ever be captured by
LLMs as human knowledge is not limited to language66. He suggests better sys-
tems will be based on different principles and will be factual, non-toxic, and
controllable67. He is not very optimistic on the future of LLMs as they are Trans-
former-based68.
60 https://www.pearltrees.com/t/artificial-intelligence/ai-risks/id69428165.
61 https://www.pearltrees.com/t/artificial-intelligence/ai-security/id68952301.
62 https://www.pearltrees.com/t/artificial-intelligence/ai-social-impact/id69428393.
63 https://arxiv.org/abs/2306.05949.
64 https://www.ethicsdialogues.eu/2023/09/13/the-ethics-of-chatgpt-exploring-the-ethical-issues-
-of-an-emerging-technology/.
65 https://twitter.com/ylecun/status/1610367976016064513; https://futurist.com/2023/02/13/
metas-yann-lecun-thoughts-large-language-models-llms/.
66 https://twitter.com/ylecun/status/1610367976016064513; https://futurist.com/2023/02/13/
metas-yann-lecun-thoughts-large-language-models-llms/.
67 https://twitter.com/ylecun/status/1610367976016064513; https://futurist.com/2023/02/13/
metas-yann-lecun-thoughts-large-language-models-llms/.
68 https://twitter.com/ylecun/status/1618387537848078337; https://twitter.com/ylecun/status/
1610367976016064513; https://futurist.com/2023/02/13/metas-yann-lecun-thoughts-large-language-
-models-llms/; https://twitter.com/ylecun/status/1618387537848078337.
131
It is observed that ChatGPT’s behaviour has been changing over time. Be-
tween March 2023 and June 2023 versions of ChatGPT, they found that the per-
formance and behaviour of ChatGPT can vary greatly over time. As the beha-
viour of ChatGPT has changed substantially in a relatively short amount of time,
could we hope for a better chatbot?69 Or, is “ChatGPT: More than a “Weapon of
Mass Deception” Ethical Challenges and Responses from the Human-Cantered
Artificial Intelligence (HCAI) Perspective”? This article suggests some ways to
prevent or reduce ChatGPT misuse or abuse and how to use it in a good way.
Some of these ways are technical, such as adding watermarks, changing styles,
detecting fakes, and checking facts. Others are non-technical, such as setting
rules, being transparent, educating users, and involving humans. There is cer-
tainly a need to educate users. Simply banning the use of a tool that is so widely
available is not a viable option70. Without appropriate measures, “ChatGPT isn’t
a great leap forward, it’s an expensive deal with the devil intelligence” (2023)71.
How will the environment survive GAI?72
The Environmental impacts of LLMs affect at first energy and water con-
sumption. To that, adds the cost of building and maintaining data centres requir-
ing large amounts of land, materials, and resources that can have negative effects
on the natural environment and local communities. Data centres produce huge
amounts of electronic waste that can contain toxic substances and pose health
along with environmental risks. Furthermore, LLMs can affect biodiversity by
reducing the demand for natural language diversity and endangering linguistic
and cultural diversity 73 . Comparing the environmental cost of LLMs with
Google search, by looking at the amount of energy and carbon emissions they
consume per query or per day, according to a study by researchers from the Uni-
versity of Bristol and the University of Massachusetts Amherst, the average
energy consumption of a Google search query in 2022 was 0.2 watt-hours,
which translates to 0.1 grams of carbon dioxide emissions74. This means that
a single Google search query has a negligible environmental impact, but when
multiplied by billions of queries per day, it adds up to a significant amount. The
69 https://arxiv.org/abs/2307.09009.
70 https://arxiv.org/abs/2304.11215.
71 https://www.theguardian.com/commentisfree/2023/feb/04/chatgpt-isnt-a-great-leap-forward-
-its-an-expensive-deal-with-the-devil.
72 https://www.pearltrees.com/t/artificial-intelligence/environmental-impacts/id68975514;
https://www.pearltrees.com/t/artificial-intelligence/new-data-centres/id71828910.
73 https://interestingengineering.com/science/llms-like-gpt-and-bard-can-be-manipulated-and-
-hypnotized; https://securityintelligence.com/posts/unmasking-hypnotized-ai-hidden-risks-large-lan-
guage-models/.
74 https://www.technologyreview.com/2022/11/14/1063192/were-getting-a-better-idea-of-ais-
-true-carbon-footprint/.
132
study estimated that a single LLM query has a much higher environmental im-
pact than a Google search query, but when multiplied by millions of queries per
day, it becomes even more substantial. The AI Index Report 2023 estimated that
ChatGPT consumed about 100 megawatt-hours of energy and emitted about 50
metric tons of carbon dioxide per day in 202375. Therefore, based on these esti-
mates, the environmental cost of LLMs is about 50 times higher than the cost of
Google search per query. However, these estimates do not consider the environ-
mental cost of training LLMs, which can be much higher than the cost of run-
ning them. Professor Crawford who describes AI as a “technology of extrac-
tion”, shares the same concerns. In a recent interview she exposed that “indica-
ting that every time you have an exchange with ChatGPT, it’s the equivalent of
pouring out half a litre of fresh water onto the ground” because that’s what it
takes to “cool the giant AI supercomputers” involved. “The energy difference
from just doing a traditional search query to using a LLM is enormous,” she
says. “Some research indicates it can be up to 1,000 times more energy inten-
sive.” “The question of the environmental cost of AI is the biggest secret in the
industry right now,” Crawford explains. “All along the pipeline – the hardware,
the software, the energy, the water to cool the systems – we have enormous en-
vironmental costs that are not being fully shared with the public”76. As GAIs are
thirsty, “Integrating large language models into search engines could mean
a fivefold increase in computing power and huge carbon emissions”77.
ChatGPT “drinks” a bottle of fresh water for every 20 to 50 questions we
ask, another study warns78. The water consumption keeps growing. “A.I. tools
fuelled a 34% spike in Microsoft’s water consumption, and one city with its data
centres is concerned about the effect on residential supply” as published in For-
tune Magazine79. The environmental impact requires urgent intervention to re-
duce the size of datasets while improving the quality of the data, and the buil-
ding of sustainable data centres.
Conclusion
In conclusion, we believe there is an urgent need for more transparency, ac-
countability, and regulation. The large-scale impact of GAI, that may require to
be distinguished from other Artificial Intelligence, calls for a dedicated frame-
work taking into account the cost of its implementation balanced with the added
value of its outcomes, to ensure companies and others deploy GAI safely, ethi-
75 https://www.analyticsvidhya.com/blog/2023/04/environmental-cost-of-ai-models-carbon-
-emissions-and-water-consumption/.
76 https://www.linkedin.com/in/taubmanbassirian/recent-activity/all/.
77 https://arxiv.org/abs/2304.03271.
78 https://www.euronews.com/green/2023/04/20/chatgpt-drinks-a-bottle-of-fresh-water-for-
-every-20-to-50-questions-we-ask-study-warns.
79 https://fortune.com/2023/09/09/ai-chatgpt-usage-fuels-spike-in-microsoft-water-consumption/.
133
cally and in a trustworthy manner. After China, a European regulation is current-
ly being adopted. The AI Act will have to be carefully drafted in order not to
block innovation while at the same time preserving the rights and freedoms of all
parties. The risk-based approach will have to survive the fast pace of technologi-
cal evolutions. As always, education and awareness will have a major role to
play. Another constant challenge in the interconnected digital world will be the
borderless impact of any regulation. How will different parts of the world regu-
late AI. A good coordination is key to success that might require an international
committee similar to the International Atomic Energy Agency (IAEA) coopera-
tion. The 2015 United 2030 agenda for sustainable development proposed objec-
tives to design and implement a worldwide safe and sustainable future. Between
its 17 Sustainable Development Goals, “industry, innovation and infrastruc-
ture”, the UN established the Technology Facilitation Mechanism (TFM) to
promote innovative solutions for the SDG agenda, including multi-stakeholder
collaboration.
AI sustainability will depend much on its use of resources and respect of en-
vironmental impacts.
References
http://www.garymarcus.com/index.html (14.05.2023).
https://apnews.com/article/artificial-intelligence-chatgpt-courts-e15023d7e6fdf4f099aa122437dbb
5 9b#lneqessb42o50gafue4 (8.10.2023).
https://arstechnica.com/information-technology/2023/09/can-you-melt-eggs-quoras-ai-says-yes-
-and-google-is-sharing-the-result/ (1.10.2023).
https://arxiv.org/abs/2304.03271 (8.04.2023).
https://arxiv.org/abs/2304.11215 (5.05.2023).
https://arxiv.org/abs/2304.13712 (25.05.2023).
https://arxiv.org/abs/2305.18248 (15.09.2023).
https://arxiv.org/abs/2306.05949 (12.06.2023).
https://arxiv.org/abs/2307.09009 (15.07.2023).
https://arxiv.org/abs/2309.03883 (10.09.2023).
https://arxiv.org/abs/2309.05463 (10.06.2023).
https://arxiv.org/abs/2309.11495. DoLa: Decoding by Contrasting Layers Improves Factuality in
Large Language Models (15.06.2023).
https://arxiv.org/pdf/2212.03551.pdf (13.07.2023).
https://bing.com/search?q=Noam+Chomsky+on+LLMs&form=SKPBOT (15.09.2023).
https://browse.arxiv.org/pdf/2304.00008.pdf (14.05.2023).
https://cyber.harvard.edu/people/dweinberger (12.07.2023).
https://dataethics.eu/testing-chatgpts-ethical-readiness/ (15.06.2023).
https://dweinberger.medium.com/chatgpt-on-why-it-pretends-to-know-things-ea2503ee872
(19.06.2023).
https://en.wikipedia.org/wiki/Wikipedia:Large_language_models (20.04.2023).
https://engineering.stanford.edu/node/16821/printable/print (16.07.2023).
https://ethicspolicy.unc.edu/news/2023/04/17/the-ethics-of-college-students-using-chatgpt
(26.05.2023).
https://fchollet.com/ (12.05.2023).
134
https://fortune.com/2023/08/01/can-ai-chatgpt-hallucinations-be-fixed-experts-doubt-altman-openai/
(12.10.2023).
https://fortune.com/2023/09/09/ai-chatgpt-usage-fuels-spike-in-microsoft-water-consumption/
(15.09.2023).
https://futurist.com/2023/02/13/metas-yann-lecun-thoughts-large-language-models-llms/ (22.03.2023).
https://futurist.com/2023/02/13/metas-yann-lecun-thoughts-large-language-models-llms/ (5.04.2023).
https://hanne-sbajohr.de/en/2023/04/08/whoever-controls-language-models-controls-politics/
(15.04.2023).
https://interestingengineering.com/science/llms-like-gpt-and-bard-can-be-manipulated-and-hypno-
tized (11.102023).
https://medium.com/@paul.k.pallaghy/the-entire-field-of-ai-is-being-professionally-gaslighted-by-
-gary-marcus-and-noam-chomsky-c08aa1e4c6f0 (14.06.2023).
https://montrealethics.ai/oppenheimer-as-a-timely-warning-to-the-ai-community/ (25.05.2023).
https://multilingual.com/timnit-gebru-and-the-problem-with-large-language-models/ (5.06.2023).
https://multilingual.com/timnit-gebru-and-the-problem-with-large-language-models/ (18.05.2023).
https://news.berkeley.edu/2023/03/19/is-chatgpt-a-false-promise (20.03.2023).
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4453869 (19.06.2023).
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4517702 (8.08.2023).
https://sebastianraschka.com/blog/2023/optimizing-LLMs-dataset-perspective.html (15.04.2023).
https://securityintelligence.com/posts/unmasking-hypnotized-ai-hidden-risks-large-language-models/
(12.09.2023).
https://techcrunch.com/2023/06/07/blush-ai-dating-sim-replika-sexbot/ (9.06.2023).
https://towardsdatascience.com/where-are-all-the-women-3c79dabfdfc2 (28.08.2023).
https://twitter.com/ylecun/status/1610367976016064513 (5.06.2023).
https://twitter.com/ylecun/status/1618387537848078337 (5.06.2023).
https://undark.org/2023/04/06/chatgpt-isnt-hallucinating-its-bullshitting/ (6.06.2023).
https://venturebeat.com/ai/chatgpt-launched-six-months-ago-its-impact-and-fallout-is-just-beginning-
-the-ai-beat (5.06.2023).
https://venturebeat.com/ai/mckinsey-says-about-half-of-its-employees-are-using-generative-ai
(13.06.2023).
https://venturebeat.com/ai/the-ai-feedback-loop-researchers-warn-of-model-collapse-as-ai-trains-
-on-ai-generated-content/ (12.07.2023).
https://venturebeat.com/ai/what-is-generative-artificial-intelligence-ai (25.06.2023).
https://venturebeat.com/ai/wordpress-launches-generative-ai-assistant-to-enhance-content-writing/
(11.06.2023).
https://walton.uark.edu/insights/posts/the-human-need-for-ethical-guidelines-around-chatgpt.php
(17.07.2023).
https://walton.uark.edu/insights/posts/the-human-need-for-ethical-guidelines-around-chatgpt.php
(25.05.2023).
https://www.analyticsvidhya.com/blog/2023/04/environmental-cost-of-ai-models-carbon-emissions-
-and-water-consumption/ (9.09.2023).
https://www.bbc.co.uk/newsround/66796495 (1.10.2023).
https://www.cloudflare.com/learning/ai/what-is-large-language-model/ (13.08.2023).
https://www.ethicsdialogues.eu/2023/09/13/the-ethics-of-chatgpt-exploring-the-ethical-issues-of-
-an-emerging-technology/ (16.09.2023).
https://www.euronews.com/green/2023/04/20/chatgpt-drinks-a-bottle-of-fresh-water-for-every-
-20-to-50-questions-we-ask-study-warns (15.09.2023).
https://www.firstpost.com/world/ai-bots-capable-of-starting-new-religions-warns-yuval--noah-harari-
-12540282.html (3.06.2023).
https://www.forbes.com/sites/forbestechcouncil/2023/06/30/10-ways-cybercriminals-can-abuse-
-large-language-models/ (6.07.2023).
135
https://www.foxnews.com/ media/snapchat-ai-chatbot-gave-advice-13-year-old-girl-relationship-31-
-year-old-man-having-sex (16.06.2023).
https://www.lesswrong.com/posts/sbaQv8zmRncpmLNKv/the-idea-that-chatgpt-is-simply-predicting-
-the-next-word-is (19.06.2023).
https://www.linkedin.com/in/taubmanbassirian/recent-activity/all/ (9.10.2023).
https://www.linkedin.com/pulse/chatgpt-could-capable-better-reasoning-llanguages-tara/ (6.05.2023).
https://www.linkedin.com/pulse/chatgpt-trouble-tara-taubman-bassirian-llm (6.05.2023).
https://www.msn.com/en-us/news/technology/gpt-4-has-arrived-it-will-blow-chatgpt-out-of-the-
-water/ar-AA18CBpwb (25.09.2023).
https://www.npr.org/2023/08/26/1195662267/ai-is-biased-the-white-house-is-working-with-hackers-
-to-try-to-fix-that (1.09.2023).
https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html (15.04.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-and-human-rights/id69493293 (8.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-biases/id62398793 (9.09.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-definitions/id62876503 (12.05.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-false-accusations/id68994378 (2.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-in-employment/id69547258 (25.09.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-ip-claims/id62513396 (25.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-manipulations/id68952637 (20.09.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-misinformation/id69147099 (15.07.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-privacy-compliance/id69428310 (8.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-risks/id69428165 (5.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-security/id68952301 (6.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/ai-social-impact/id69428393 (1.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/anthropomorphisation/id71886740 (15.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/call-for-ban-moratory/id65034697 (3.06.2023).
https://www.pearltrees.com/t/artificial-intelligence/call-for-ban-moratory/id650 (6.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/chatgpt-alternatives/id62359814 (20.05.2023).
https://www.pearltrees.com/t/artificial-intelligence/chatgpt-incidents/id71414555 (17.09.2023).
https://www.pearltrees.com/t/artificial-intelligence/chatgpt-writings/id71414004 (1.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/environmental-impacts/id68975514 (5.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/labour-exploitation/id69490858 (25.10.2023).
https://www.pearltrees.com/t/artificial-intelligence/mistral/id71509507 (18.09.2023).
https://www.pearltrees.com/t/artificial-intelligence/new-data-centres/id71828910 (7.10.2023).
https://www.politico.com/newsletters/digital-future-daily/2023/02/15/ais-political-bias-problem-
-00083095 (13.03.2023).
https://www.technologyreview.com/2020/12/04/1013294/google-ai-ethics-research-paper-forced-
-out-timnit-gebru/ (20.05.2023).
https://www.technologyreview.com/2022/11/14/1063192/were-getting-a-better-idea-of-ais-true-
-carbon-footprint/ (10.09.2023).
https://www.technologyreview.com/2022/12/23/1065852/whats-next-for-ai/ (9.05.2023).
https://www.technologyreview.com/author/melissa-heikkila/ (6.10.2032).
https://www.telegraph.co.uk/business/2023/03/29/control-ai-threat-civilisation-warns-elon-musk/
(1.10.2023).
https://www.telegraph.co.uk/news/2023/04/23/yuval-noah-harari-i-dont-know-if-humans-can-
-survive-ai/ (15.05.2023).
https://www.theguardian.com/commentisfree/2023/feb/04/chatgpt-isnt-a-great-leap-forward-its-an-
-expensive-deal-with-the-devil (12.09.2023).
https://www.thesun.ie/tech/10808612/snapchat-artificial-intelligence-bot-danger-children/ (14.07.2023).
136
https://www.thetimes.co.uk/article/why-2024s-presidential-race-will-be-the-first-ai-election-jb32pj8br
(17.08.2023).
https://www.verywellmind.com/jean-piaget-biography-1896-1980-2795549 (17.05.2023).
https://www.weforum.org/whitepapers/jobs-of-tomorrow-large-language-models-and-jobs (1.06.2023).
https://www.wired.com/story/fast-forward-chatbot-hallucinations-are-poisoning-web-search/ (7.10.2023).
https://www.wired.com/story/fast-forward-chatbot-hallucinations-are-poisoning-web-search/ (8.10.2023).
https://www3.weforum.org/docs/WEF_Jobs_of_Tomorrow_Generative_AI_2023.pdf (1.06.2023).
