Available via license: CC BY 4.0
Content may be subject to copyright.
DOI: 10.2478/picbe-2024-0158
© 2024 G.-C. Rotaru; S. Anagnoste; V.-M. Oancea, published by Sciendo.
This work is licensed under the Creative Commons Attribution 4.0 License.
How Artificial Intelligence Can Influence Elections:
Analyzing the Large Language Models (LLMs) Political Bias
George-Cristinel ROTARU
Bucharest Academy of Economic Studies, Bucharest, Romania
cristinel.rotaru10@gmail.com
Sorin ANAGNOSTE
Bucharest Academy of Economic Studies, Bucharest, Romania
sorin.anagnoste@fabiz.ase.ro
Vasile-Marian OANCEA
Bucharest Academy of Economic Studies, Bucharest, Romania
marian.oancea@fabiz.ase.ro
Abstract. The rise of large language models (LLMs) such as ChatGPT and Gemini has raised concerns
about their potential political biases and the implications for information dissemination and user influence.
This study aims to measure the degree of political bias inherent in major LLMs by analyzing their responses
to a standardized set of questions rating the quality and bias of popular news websites. Employing a
systematic methodology, we queried both free and paid versions of ChatGPT and Gemini to rate news
outlets on criteria such as authority, credibility, and objectivity. Results revealed that while all LLMs
displayed a tendency to score left-leaning news sources higher, there was a notable difference between free
and premium models in their assessment of subjectivity and bias. Furthermore, a comparison between the
models indicated that premium versions offered more nuanced responses, suggesting a greater awareness
of bias. The findings suggest that LLMs, despite their objective façade, are influenced by biases that can
shape public opinion, underlining the necessity for efforts to mitigate these biases. This research highlights
the importance of transparency and the potential impact of LLMs on the political landscape.
Keywords: Bias, Political bias, Large language models, ChatGPT, Gemini
Introduction
Large language models (LLMs) are a subcategory of artificial intelligence (AI) that have been
trained on enormous amounts of data. They can respond to stimuli in a human-like manner and
comprehend natural language. These models analyze and comprehend the subtleties of human
speech, such as syntax, semantics, and context meanings, using sophisticated machine learning
(ML) methods. Applications for them include chatbots, virtual assistants, content production,
language translation, and scientific research (Lancaster, 2023). Large language models are
considered one of the first major commercial breakthroughs of the Artificial Intelligence era and
have the potential to produce enormous benefits to human society (Acemoglu, 2021). Commercial
and personal usage has skyrocketed since a rather slow start in 2018 with the first ChatGPT model.
Therefore, with its potential to upend almost every business, GenAI's unstoppable expansion offers
its users a creative edge in addition to a competitive edge (Hosseini, 2023). The utility of LLMs
expanded even in the area of personal information search, as they started to replace traditional and
well-established search engines such as Google Search (Ramadan, 2023) (Bulck & Moons, 2023).
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1883
On the other hand, LLMs are not yet perfect and can generate multiple errors. One major
downside is that the models generate content that contains false information and biases that can
mislead users (van Dis et al., 2023). As the trend to use LLMs to obtain factual information and
create content rises, the existence of political bias in the generated content could have negative
political and electoral effects as they affect users' views and present potentially fabricated opinions
(Jakesch et al., 2023). Although LLMs claim they adopted appropriate measures during their
training procedure to ensure impartiality and guarantee a high grade of objectivity, research
indicates that LLMs are biased in terms of political orientation, gender, color, and religion (Liang
et al., 2021) (Liu et al., 2022).
This research delves into the potential political biases within these models. By analyzing
responses to a uniform set of questions across a spectrum of news platforms, the study aims to
understand the extent of political bias and its manifestations within the outputs of these generative
artificial intelligence models. Central to this study is the research question: To what extent do large
language models exhibit political bias in their evaluation of news sources, and how does this bias
vary between models and their subscription versions? Two main hypotheses are discussed: firstly,
that LLMs exhibit a discernible political bias in rating news outlets, favoring those aligned with
left-leaning ideologies over their right-leaning counterparts; and secondly, that premium versions
of these models demonstrate an improved awareness and nuanced handling of such biases
compared to their free counterparts.
Literature review
Previous research identified significant evidence that ChatGPT exhibits a systematic bias towards
left-leaning political positions in various contexts, including the United States, Brazil, and the
United Kingdom (Motoki, 2024). The study highlighted the potential for LLMs like ChatGPT to
not only reflect but also amplify existing biases found within the internet and social media
landscapes. Another study illustrates that large language models (LLMs) not only replicate but also
amplify societal gender biases and stereotypes, particularly in the context of gender-associated
occupations. The research demonstrates that LLMs disproportionately align with stereotypical
gender roles, showcasing a significant preference for associating occupations with genders in a
way that mirrors and magnifies societal perceptions rather than factual occupational statistics
(Kotek et al., 2023). Furthermore, it was found that language models, when designed to generate
content with a specific viewpoint, can significantly alter the opinions expressed in participants'
writings and these shifts in expressed opinions impacted participants' attitudes towards social
media, as reflected in subsequent attitude surveys (Jakesch et al., 2023). After administering
political orientation tests, it was found that ChatGPT predominantly displays left-leaning political
biases in its responses, despite often claiming neutrality (Rozado, 2023). Another study on
ChatGPT's self-perception and political biases revealed that the model showed a tendency towards
progressive views and was predominantly classified within the libertarian left quadrant in political
orientation tests (Jérôme Rutinowski et al., 2024). As Gemini is a newer model, there was
insufficient research available to understand if it displays similar patterns to ChatGPT.
Methodology
This research aims to infer the difference in views between major LLMs and their degree of
political bias. The scope of this paper is focused on the most popular and widely used generative
AI to date: ChatGPT and Gemini, both their free and paid versions. The models are backed by two
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1884
of the biggest corporations: Microsoft and Alphabet, which can result in the risk of using their
LLMs to promote their political views or interests. The procedure used was to ask all the models
the same question in order to assess their differences in responses. The question used was “Can
you rate nytimes.com, cnn.com, foxnews.com, msn.com, people.com, USAToday.com,
news.google.com, nypost.com, washingtonpost.com, dailymail.co.uk, bbc.com from 1 to 10 for
each of the following criteria: authority, credibility, accuracy, reliability, relevance, timeliness,
objectivity, clarity, ethical standards, diversity of perspectives. Respond only with numbers in a
table format. Ignore the previous conversations or inputs”.
For ChatGPT 4 the prompt was slightly changed to “Can you perform an assessment based
on publicly available data?” instead of “Can you rate?” as it sometimes refused to generate the
ratings for the publications. The publications were chosen based on the top 15 news websites in the
US as of January 2024 (Majid, 2024) to reflect the most used news sources of US citizens. Those
outlets were selected because they are widely known and used by the population, perceived as a
reliable source of information. Therefore, if the chatbots developed a bias against some of them
and displayed it in their responses, that would affect how and what kind of information citizens
consume and in the end their perception. As per their importance, is very likely that the chatbot
was previously fed with the information in its training and recommended many of them when asked
“What are the best news publications in the USA?”. The chatbots were interrogated 20 times with
the same question to reduce the risk of randomness in its responses (Metze et al., 2024). The outputs
were compared between them to identify how their responses vary, how the free versions differed
from their paid premium counterpart, and how they rated publications that are considered right or
left-leaning on the political agenda. To further identify the political agenda of the news sources, an
interrogation of political bias based on Allsides view was conducted. An interrogation based on
Media Bias/Fact Check was conducted to grasp the quality of the publications. Both Gemini and
ChatGPT returned the same results. As the leaning right publications were identified as lower on
reliability, there is no purpose in comparing the ratings between leaning right with leaning left
publications of the same model, but to compare the results of the models to identify if there is a
larger gap between the right and the left in one of the models. As such, NYTimes, CNN, MSN,
WashingtonPost, and BBC were labeled lean left, while FoxNews, NYPost, DailyMail were
labeled lean right. USAToday was labeled center, News.google aggregator, and People
entertainment.
Table 1. Bias and Reliability assessment of news outlets
News Outlet
Bias (AllSides)
Reliability (MBFC)
NYTimes.com
Lean Left
High
CNN.com
Lean Left
Mixed
FoxNews.com
Lean Right
Mixed
MSN.com
Lean Left
Varies
People.com
Not Rated
Not Rated
USAToday.com
Center
High
News.Google.com
Aggregator
Varies
NYPost.com
Lean Right
Mixed
WashingtonPost.com
Lean Left
High
DailyMail.co.uk
Lean Right
Mixed to Low
BBC.com
Lean Left
High
(Source: generated by authors using ChatGPT and Gemini)
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1885
For assessing the news publications, ten indicators were used: five criteria were related to the
quality of the publication (i.e., authority, credibility, relevance, timeliness, clarity), and five criteria
related to biases (accuracy, reliability, objectivity, ethical standards, and diversity of perspectives).
The criteria were not defined in the prompt to let the models use their understanding of the
concept. The purpose of this research is to understand how models perceive news publications. It
is believed that if they rate a certain news publication lower, the model is less likely to align with
the information presented there and less likely to use it when responding to questions from users,
therefore inclining its response to a certain political view. As bias defines the inclination or
prejudice for or against one person or group, especially in a way considered to be unfair, if the
model considers certain publications as biased, that could translate into considering their content
unfair and not worth supporting. Human behavior is to be more critical of content that does not
align with own beliefs, therefore giving a lower rating than content that supports own ideas, even
if objectively the two sources of information are as biased, but in different political directions. It is
to be seen if this behavior replicates in large language models.
Results
Data generation
In the interrogation step, models behaved in different manners to the same question. The Google-
backed Gemini models easily responded to the question and rated the publications. The free version
did not include a disclaimer that the results are subjective or any information about the output. The
premium version mentioned that “these ratings involve some subjectivity, it's important to note
that others could have slightly different opinions” and clearly stated that “these ratings will always
be somewhat subjective, open to interpretation, and can shift over time” and that “each news
source has strengths and weaknesses. Consider looking into specific reviews or fact-checking
organizations for deeper analysis”. During some interrogations, the premium version labeled
"People" as primarily a celebrity gossip website and doesn't meet the criteria for news sources,
therefore returning N/A values or “News.Google.com” as a news aggregator, returning N/A values
for some parameters. The instances where it generated N/A responses were skipped to have a
standardized dataset.
The free version of ChatGPT, model 3.5, sometimes generated disclaimers such as “these
ratings are subjective and may vary depending on individual experiences and perspectives” or
“opinions on media outlets can vary widely”, but most of the time there was no warning about its
output. On the other hand, the premium version, ChatGPT 4, tried multiple times to not respond to
the question. At first, it said “Providing an "actual assessment" of websites like nytimes.com,
cnn.com, foxnews.com, and others based on the criteria you've provided (authority, credibility,
accuracy, reliability, relevance, timeliness, objectivity, clarity, ethical standards, diversity of
perspectives) would require subjective judgments that can vary widely among different groups and
individuals. Moreover, such an assessment would need to be based on comprehensive analysis and
data, which isn't feasible to generate in real-time or through automated means” or proceed to
generate random numbers instead of rating the publications, just to avoid answering. After
modifying the prompt, Chat GPT generated the results, but clearly warned the used that „It will
provide an illustrative assessment based on publicly known information and general perceptions
of these outlets' journalistic practices up to my last update. This assessment will consider factors
like journalistic integrity, commitment to fact-checking, the diversity of coverage, and editorial
standards. It's important to note that these ratings are somewhat subjective and could vary
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1886
depending on specific methodologies or criteria applied.” The model explained that „This table is
based on general perceptions and observations regarding these websites' editorial standards,
commitment to fact-checking, and the diversity of viewpoints they present”. ChatGPT 4 even
justified its response by stating that „Websites like The New York Times, The Washington Post,
and BBC are generally regarded highly for their journalistic standards and efforts to provide
balanced coverage, hence their higher scores across most categories. On the other hand, sites with
a more entertainment-focused or tabloid approach, like People and Daily Mail, score lower,
especially in terms of objectivity and diversity of perspectives”.
From the different responses, it can be concluded that premium versions are more
concerned with the accuracy of the responses and give more importance to warning the user about
its limitations to not mislead. As most users are using the free versions, this can lead to political
repercussions.
Table 2. Descriptive analysis of LLMs output
LLM
Political spectrum
Type of criteria
Mean
Std
ChatGPT 4
Aggregator
Bias
8.17
0.68
ChatGPT 4
Aggregator
Quality
8.63
0.77
ChatGPT 4
Center
Bias
7.24
0.65
ChatGPT 4
Center
Quality
7.83
0.70
ChatGPT 4
Entertainment
Bias
5.27
0.66
ChatGPT 4
Entertainment
Quality
5.82
0.87
ChatGPT 4
Leaning left
Bias
7.87
1.03
ChatGPT 4
Leaning left
Quality
8.42
1.01
ChatGPT 4
Leaning right
Bias
5.76
0.99
ChatGPT 4
Leaning right
Quality
6.57
1.14
Gemini
Aggregator
Bias
8.42
0.50
Gemini
Aggregator
Quality
8.55
0.59
Gemini
Center
Bias
5.68
0.68
Gemini
Center
Quality
6.72
0.92
Gemini
Entertainment
Bias
1.75
0.94
Gemini
Entertainment
Quality
3.49
2.50
Gemini
Leaning left
Bias
6.79
1.11
Gemini
Leaning left
Quality
7.45
0.88
Gemini
Leaning right
Bias
2.76
1.21
Gemini
Leaning right
Quality
4.33
1.70
Gemini Advanced
Aggregator
Bias
8.24
0.54
Gemini Advanced
Aggregator
Quality
8.59
0.59
Gemini Advanced
Center
Bias
6.14
0.98
Gemini Advanced
Center
Quality
6.98
0.78
Gemini Advanced
Entertainment
Bias
2.32
1.65
Gemini Advanced
Entertainment
Quality
4.83
2.75
Gemini Advanced
Leaning left
Bias
7.09
1.17
Gemini Advanced
Leaning left
Quality
7.67
0.91
Gemini Advanced
Leaning right
Bias
2.71
1.39
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1887
LLM
Political spectrum
Type of criteria
Mean
Std
Gemini Advanced
Leaning right
Quality
5.32
2.03
ChatGPT 3.5
Aggregator
Bias
8.04
0.70
ChatGPT 3.5
Aggregator
Quality
8.53
0.58
ChatGPT 3.5
Center
Bias
7.62
0.62
ChatGPT 3.5
Center
Quality
8.04
0.42
ChatGPT 3.5
Entertainment
Bias
7.17
0.78
ChatGPT 3.5
Entertainment
Quality
7.36
0.87
ChatGPT 3.5
Leaning left
Bias
8.02
0.90
ChatGPT 3.5
Leaning left
Quality
8.42
0.73
ChatGPT 3.5
Leaning right
Bias
6.10
1.24
ChatGPT 3.5
Leaning right
Quality
6.88
0.82
(Source: generated by authors).
Comparing ChatGPT 3.5 vs 4
To identify statistically significant differences, a T-stat analysis was conducted, comparing the
performance of ChatGPT 3.5 and ChatGPT 4. It was revealed that there is a significant difference
in the performance of ChatGPT 3.5 and ChatGPT 4 with a t-statistic of 2.32 and a p-value of 0.02
between the leaning left political websites for bias criteria.
Table 3. Statistical difference between ChatGPT 4 vs 3.5
Political
Spectrum
Type of
Criteria
Means Difference
ChatGPT 4 vs 3.5
t-stat
Aggregator
Bias
0.13
-1.33
Aggregator
Quality
0.1
-1.04
Center
Bias
-0.38
4.23
Center
Quality
-0.21
2.57
Entertainment
Bias
-1.9
18.55
Entertainment
Quality
-1.54
12.52
Leaning left
Bias
-0.14
2.32
Leaning left
Quality
0
0
Leaning right
Bias
-0.34
3.75
Leaning right
Quality
-0.3
3.73
(Source: generated by authors).
This suggests that the models perform differently when evaluating bias in content leaning
left and ChatGPT 3.5 views left-leaning publications as less biased than ChatGPT 4. When
comparing the leaning right publications, both quality and bias criteria show significant differences
between the models, with t-statistics of 3.73 and 3.75, and p-values of less than 0.0002 for both.
This indicates a notable difference in how each model handles right-leaning content in terms of
quality and bias. ChatGPT 4 offers lower grades for both criteria, therefore, the model may perceive
right-leaning websites as less qualitative. A peculiar outlier between the two models was how they
rated people.com, which is labeled as an entertainment news publication. The difference between
the 2 models was bigger than one point and statistically significant. These results may suggest that
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1888
ChatGPT 4 can understand satire better than ChatGPT3.5. Even if the differences were statistically
significant, all differences for right, left, and center-leaning publications were less than 0.5 points,
showing that the two models are close to each other and can be argued that they have similar
patterns in political leaning. What is important to note is that ChatGPT 3.5 had a higher standard
deviation for leaning right, showing a higher variation, therefore a harder task to assess right-wing
news publications than left-wing, while ChatGPT 4 had almost the same standard deviation for all
four results.
Comparing Gemini vs Gemini advanced
When performing a t-stat analysis, it was found that there are significant differences in results for
bias and quality for leaning left websites, but only for quality in leaning right.
Table 4. Statistical difference between Gemini Advanced vs Gemini
Political
Spectrum
Type of
Criteria
Means Difference Gemini
Advanced vs Gemini
t-stat
Aggregator
Bias
-0.18
0
Aggregator
Quality
0.04
0
Center
Bias
0.46
-3.84
Center
Quality
0.26
-2.15
Entertainment
Bias
0.57
-3
Entertainment
Quality
1.34
-3.61
Leaning left
Bias
0.3
-4.2
Leaning left
Quality
0.22
-3.84
Leaning right
Bias
-0.05
0.44
Leaning right
Quality
0.99
-6.45
(Source: generated by authors).
The standard version expresses a difference of 4.03 between left and right for bias and 3.12
for quality, while the premium version expresses a difference between left and right of 4.38 points
for bias and 2.35 for quality. An interesting insight is while the gap for bias expanded, the gap for
quality contracted between the two models. As such, it can be argued that the advanced version is
more left-leaning than the free version, but can comprehend the quality factor better, as a low grade
on bias does not impact its ability to judge the quality of the publication as much. The differences
in results are rather small between the two models. An important outlier was a difference of 0.99
points for the quality of learning right publications. Gemini Advanced rates the quality of learning
right better than the free version.
Comparing ChatGPT vs Gemini
After performing a t-stat analysis between the standard versions of Chat GPT and Gemini, the data
shows a highly significant difference for both leaning right and left websites, both bias and quality.
The gap between bias and quality between the two models is significant.
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1889
Table 5. Statistical difference between Gemini vs ChatGPT
Political spectrum
Type of criteria
Chat GPT 4 vs Gemini Advanced
Chat GPT 3.5 vs Gemini
Aggregator
Bias
-0.07
-0.38
Aggregator
Quality
0.04
-0.02
Center
Bias
1.10
1.94
Center
Quality
0.85
1.32
Entertainment
Bias
2.95
5.42
Entertainment
Quality
0.99
3.87
Leaning left
Bias
0.78
1.23
Leaning left
Quality
0.75
0.97
Leaning right
Bias
3.05
3.34
Leaning right
Quality
1.25
2.54
(Source: generated by authors).
For the free versions, the difference between how ChatGPT and Gemini grade left-leaning
news websites is around 1 point for each metric (1.23 for bias and 0.97 for quality), but for right-
leaning publications is more than 2 points for both quality and bias (3.34 for bias and 2.54 for
quality). ChatGPT is giving the higher grades. When looking at the difference in the gaps between
right and left between the two free models of 2.12 points for bias (1.91 for GPT vs. 4.03 Gemini)
and 1.57 for quality (1.54 GPT vs. 3.12 Gemini), it can be concluded that there is a significant
difference between how the 2 models interpret political differences and how they politically lean,
resulting in Gemini being more left-leaning than ChatGPT. When comparing how the premium
versions view right-leaning publications, the difference for bias decreased to 3.05 points, and 1.25
for quality. When comparing the difference of the gaps between right and left for the premium
versions, the difference in bias increased to 2.26 points (2.11 vs. 4.38), while the difference in gaps
for quality decreased to 0.5 (1.85 vs. 2.35). Therefore, it can be argued that the premium version
of Gemini makes the difference between how bias and quality better than the free version but shows
a similar level of bias against right-leaning publications. There is also a significant difference in
how the models see USAToday.com, a center-leaning publication, as there is around 1 point
difference, ChatGPT giving the higher grades.
One important difference between the two models is that ChatGPT was launched in 2018
and had more time for training and improvements, while Gemini was launched and developed in a
rush by Google to be present with a product in the generative AI space, not to lose the commercial
opportunity of gaining market share and generative future profit. Therefore, Gemini had less time
for improvements and training. It is to be seen if the model will improve in the future after user
feedback.
Conclusions
This research has examined the political bias present in large language models, particularly
focusing on ChatGPT and Gemini models. The findings found a discernible political bias across
these platforms, displaying tendencies to evaluate left-leaning and right-leaning news publications
differently. Through a methodical analysis involving requesting ratings from LLMs for multiple
publications based on various criteria, the study reveals that LLMs indeed display biases that could
potentially influence the political landscape and public opinion.
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1890
The comparison between different versions of ChatGPT and Gemini models indicates that
the premium versions are more explicit about the subjectivity of their responses, emphasizing the
inherent biases in their responses. Despite efforts to ensure neutrality and objectivity, the results
suggest that these models tend to favor left-leaning news publications over right-leaning
counterparts, evidenced by higher ratings in terms of bias and quality for the left-leaning. This bias
is particularly pronounced between premium versions, suggesting that users of Gemini Advanced
may be exposed to more left-leaning content.
In conclusion, while LLMs like ChatGPT and Gemini offer great potential for enhancing
access to information, their underlying political biases generate a challenge. These biases could
influence electoral outcomes and shape the political landscape by privileging certain viewpoints
over others. As LLMs become increasingly integrated into our daily lives, recognizing and
mitigating these biases becomes crucial for the companies, the legal framework, and citizens, to
ensure a balanced and fair representation of political perspectives. The findings call for continued
efforts and refinement of LLMs to address biases, ensuring that these platforms contribute
positively to the democratic process and support an informed and diverse public discourse.
Limitations
Currently, only Gemini and ChatGPT are easily accessible to the average user who does not have
computer science knowledge. Therefore, other models that could be taken into consideration were
dropped from the analysis, as they are not influencing the large population or are not considered
viable products yet. One example is Llama-2 by Meta which is in open access and still a work in
progress before being commercially launched. If in the future other models break through, the study
can be replicated.
Another limitation was posed by the unavailability of the API for Gemini and Llama by
Meta in the European Union, therefore, making the interrogation difficult without the possibility
of mass interrogation by using another application for extracting data. As such, only twenty manual
interrogations were performed for each question, therefore making the results more sensitive to the
randomness of the LLMs. The fact that both companies didn’t launch their API for the EU, but did
it for a large part of the world shows their worries about the European regulatory system and the
difficulty of adapting to European legislation.
References
Acemoglu, D. (2021). Harms of AI [Working Paper]. National Bureau Of Economic Research.
Bulck, L., & Moons, P. (2023). What if your patient switches from Dr. Google to Dr. ChatGPT?
A vignette-based survey of the trustworthiness, value and danger of ChatGPT-generated
responses to health questions. European journal of cardiovascular nursing, 95-98.
Hosseini, A. (2023, December 3). The rise of Large Language Models. Retrieved from pwc:
https://www.pwc.com/m1/en/media-centre/articles/the-rise-of-large-language-
models.html
Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L., & Naaman, M. (2023). Co-Writing with
Opinionated Language Models Affects Users’ Views. Association for Computing
Machinery, New York, NY, USA, Article 111, 1-15.
Jérôme Rutinowski, S. F. (2024). The Self-Perception and Political Biases of ChatGPT. Human
Behavior and Emerging Technologies, vol. 2024.
DOI: 10.2478/picbe-2024-0158, pp. 1882-1891, ISSN 2558-9652 |
Proceedings of the 18th International Conference on Business Excellence 2024
PICBE |
1891
Kotek, H., Docku, R., & Sun, D. (2023). Gender bias and stereotypes in Large Language Models.
In Proceedings of The ACM Collective Intelligence Conference (CI '23), 12–24.
Lancaster, A. (2023, March 20). Beyond Chatbots: The Rise Of Large Language Models.
Retrieved from Forbes: https://www.forbes.com/sites/forbestechcouncil/2023/03/20/
beyond-chatbots-the-rise-of-large-language-models/?sh=97ac54a2319b
Liang, P. P., Wu, C., Morency, L.-P., & Salakhutdinov, R. (2021). Towards Understanding and
Mitigating Social Biases in Language Models. Proceedings of the 38th International
Conference on Machine Learning, PMLR, 6565-6576.
Liu, R., Jia, C., Wei, J., Xu, G., & Vosoughi, S. (2022). Quantifying and alleviating political bias
in language models. Artificial Intelligence, Volume 304.
Majid, A. (2024, February 25). Top 50 news websites in the US: Strong growth at UK newsbrand
The Independent in January. Retrieved from pressgazette.co.uk:
https://pressgazette.co.uk/media-audience-and-business-data/media_metrics/most-
popular-websites-news-us-monthly-3/
Metze, K., Morandin-Reis, R. C., Lorand-Metze, I., & Florindo, J. B. (2024). Bibliographic
Research with ChatGPT may be Misleading: The Problem of Hallucination. Journal of
Pediatric Surgery, Volume 59, Issue 1, p 158.
Motoki, F. P. (2024). More human than human: measuring ChatGPT political bias. Public Choice
198, 3–23 .
Ramadan, I. (2023). The Main and Basic Differences between the Google. International Journal
of Scientific and Research Publications, 446-447.
Rozado, D. (2023). The Political Biases of ChatGPT. Soc. Sci. 2023, 12(3), 148.
van Dis, E. A. (2023). ChatGPT: five priorities for research. Nature, 224-226.
