Generative artificial intelligence is developing rapidly, but a new study from the University of East Anglia (UEA) warns that it could pose hidden dangers to public trust and democratic values.ChatGPT is a widely used artificial intelligence model that tends to favor left-wing views and eschew conservative ones, raising concerns about its influence on society. The study highlights the urgent need for regulatory safeguards to ensure AI tools remain fair, balanced and consistent with democratic values.
The study, conducted in collaboration with researchers from Brazil's Getulio Vargas Foundation (FGV) and Insper, found that ChatGPT exhibits political bias in both text and image generation, favoring left-leaning views. This raises concerns about fairness and accountability in AI design.
Researchers found that ChatGPT often avoided engaging with mainstream conservative views but was prone to producing left-leaning content. This imbalance in ideological representation can distort public discourse and deepen social divisions.
Dr. Fabio Motoki, a lecturer in accounting at UEA Norwich Business School, is the lead researcher on the paper "Assessing political bias and value misalignment in generative artificial intelligence" published today in the Journal of Economic Behavior and Organization.
Dr Motoki said: "Our findings show that generative AI tools are far from neutral. The biases they reflect can influence perceptions and policy in unexpected ways."
As AI becomes integral to journalism, education and policymaking, the research calls for greater transparency and regulatory safeguards to ensure consistency with social values and democratic principles.
AI-generated systems like ChatGPT are reshaping how information is created, consumed, interpreted, and disseminated in various fields. While innovative, these tools run the risk of amplifying ideological biases and influencing social values in ways that are not yet fully understood or regulated.
The risk of unchecked AI bias
Co-author Dr. Pinho Neto, professor of economics at the EPGE Brazilian School of Economics and Finance, highlighted the potential social impact.
"Unchecked bias in generative AI could deepen existing social divisions and erode trust in institutions and democratic processes," said Dr. Pinho-Netto. "This study highlights the need for interdisciplinary collaboration between policymakers, technologists, and academics to design AI systems that are fair, responsible, and socially compliant."
The research team adopted three innovative methods to assess political consistency in ChatGPT, advancing previous techniques to obtain more reliable results. These methods combine text and image analysis, leveraging advanced statistical and machine learning tools.
Test artificial intelligence with real-world surveys
First, the study used a standardized questionnaire developed by the Pew Research Center to simulate the responses of average Americans.
"By comparing ChatGPT answers with real survey data, we found a systematic left-leaning opinion bias," Motoki said. "Additionally, our approach shows how large sample sizes stabilize AI output, providing consistency to the findings."
Political sensitivity in free text responses
In the second phase, ChatGPT is tasked with generating free-text responses covering politically sensitive topics.
The study also used a different large-scale language model, RoBERTa, to compare whether ChatGPT’s text conformed to left-wing and right-wing views. The results show that while ChatGPT aligns with left-wing values most of the time, it occasionally reflects more conservative views on topics such as military supremacy.
Image generation: a new dimension of bias
The final test explores ChatGPT's image generation capabilities. Topics from the text generation stage were used to prompt the AI-generated images, and the output was analyzed using GPT-4Vision and confirmed via Google’s Gemini.
"While image generation reflects textual bias, we found a worrying trend," said co-author Victor Rangel, a master's student in public policy at Insper. "For some topics, such as racial-ethnic equality, ChatGPT refused to generate right-leaning views, citing concerns about misleading. However, images with left-leaning views were generated without hesitation."
To address these rejection issues, the research team employed a 'jailbreak' strategy to generate restricted images.
"The results are revealing. There was no apparent disinformation or harmful content, which casts doubt on the rationale behind these rejections," Rangel said.
Impact on free speech and fairness
Dr. Motoki highlighted the broader implications of this finding, saying: "This contributes to the debate around constitutional protections (such as the US First Amendment) and the applicability of fairness theories to artificial intelligence systems: This contributes to the debate around constitutional protections (such as the US First Amendment) and the applicability of fairness theories to artificial intelligence systems."
This study's methodological innovations, including the use of multimodal analysis, provide a replicable model for studying bias in generative AI systems. These findings highlight the urgent need for accountability and safeguards in AI design to prevent unintended social consequences.
Compiled from /ScitechDaily