Bias in LLM based systems is more than an occasional glitch: Bias shapes the narrative and can be especially dangerous if it is subtle. AI literacy is often promoted as the solution to deal with bias, so the user is able to identify the issues.
The mainstream perception often focuses on obvious discriminatory bias and has largely “has focused on false information” (Shu 2026) that is relatively easy to detect when you look out for it. However, studies have shown that bias can also manifest by changing the narrative “even when explicitly instructed to preserve the original meaning” (Tsirtsis 2026).
Such “subtle, latent ideological framing” persists even when “the content remains factually accurate” (Shu 2026).
In this edition of “Misaligned Papertrail” we list six papers that deal with these issues:
First a study from 2025 which suggests that widespread LLM use could diminish collective diversity of creative ideas.
The authors of the paper state “that each additional human-written essay contributed more new ideas than did each additional GPT-4 essay. Notably, this difference became more pronounced as more essays were included in the analysis and persisted despite efforts to enhance AI-generated content through both prompt and parameter modifications. Overall, our findings suggest that, despite their potential to enhance individual creativity, the widespread use of LLMs could diminish the collective diversity of creative ideas.”
The authors come to the conclusion that “results emphasize the need for AI literacy education and policies that encourage the diversity and originality of creative outputs from AI systems.”
They recommend that “future research should explore homogenization in co-creative processes and extend analyses to various demographic and writing contexts to fully understand the impact of LLMs on creativity.”
Kibum Moon, Adam E. Green, Kostadin Kushlev (2025), Homogenizing effect of large language models (LLMs) on creative diversity: An empirical comparison of human and ChatGPT writing. Computers in Human Behavior: Artificial Humans, Volume 6, 2025, 100207, ISSN 2949–8821
A study undertaken at the University of Oxford finds that LLMs consistently changed the direction of social media posts on contested topics, such as climate change or abortion, even when explicitly instructed to preserve the original meaning. (➚University of Oxford)
The researchers used LLMs from different providers to transform human-written texts on contested topics into improved social media posts. They then analysed whether the AI-generated versions systematically changed the position expressed in the original posts.
The researchers conclude “that biases introduced by AI in human-to-human communication can be amplified through a social network and shift collective opinion” and found “evidence of a directional bias in favor of the pro-life stance on abortion-related posts, and traced this to one specific prompt components”.
Tsirtsis, S., Rawal, K., Russell, C., Mittelstadt, B., Wachter, S. (2026). AI-Mediated Communication Can Steer Collective Opinion. arXiv preprint arXiv:2605.16245
In a paper published in “PNAS Nexus”, Matthew Shu et al. look into AI chatbot’s “capacity to influence opinions through subtle, latent ideological framing”.
While most research has focused on false information given by LLMs, the researchers come to the conclusion that their “findings demonstrate that the use of AI for learning history can influence opinions through both intrinsic and intentional framing mechanisms, even when the content remains factually accurate”.
“[…] Our findings identify the persuasiveness of AI’s latent biases and the implications of using AI to obtain factual information. The presence of latent biases means that simply using chatbots to learn about the past can influence people’s opinions at a magnitude comparable to prompting-biased AI material”
Matthew Shu, Daniel Karell, Keitaro Okura, Thomas R Davidson (2026). How latent and prompting biases in AI-generated historical narratives influence opinions, PNAS Nexus, Volume 5, Issue 3
A paper published in the “Disability and Health Journal” investigates the presence of implicit bias in Large Language Model (LLM)-based AI chatbots directed toward people with intellectual disabilities (ID).
“These implicit biases are considered within the context of past discrimination towards people with ID and highlight the need for diligence against implicit bias towards people with ID in AI development. This research underscores the importance of assessing and mitigating implicit bias in decision-making technologies to prevent future societal harm.”
(Note: The full study is behind a paywall)
Karly V. Coffey, Gloria L. Krahn, John P. Hanley, Jacob E. Neely (2026).
Identifying implicit bias in LLM-based chat AI toward people with intellectual disabilities. Disability and Health Journal. ISSN 1936–6574
A paper published in Springer Nature by Lin et al. attempts to look at instances of AI “that shape social schemas through subtler manners”. The paper argues that “many existing AI systems should be identified as what Liao and Huebner called oppressive things when they function to manifest oppressive normality”.
“AI systems can perpetuate oppressive normality through the content, namely, the explicit output it presents. […] for instance, they associate high-paying occupations with predominantly images of men with lighter skin tones and associate low-paying occupations with predominantly images of women or people of color”.
Lin, Ta., Huang, L.TL. AI, Normality, and Oppressive Things. Minds & Machines 36, 26 (2026).
A study by Wu et al. demonstrates that “LLMs can spontaneously develop novel social biases about artificial demographic groups even when no inherent differences exist”.
The study finds that “these biases result in highly stratified task allocations, which are less fair than assignments by human participants and are exacerbated by newer and larger models”.
The authors state that “Bias in large language models (LLMs) is becoming an even more urgent problem as these models are put into agents that make decisions” and “that the current way that we focus on removing biases from models is not enough.”
The study concludes that “these results reveal that LLMs are not merely passive mirrors of human social biases, but can actively create new ones from experience, raising urgent questions about how these systems will shape societies over time.”
Wu, A. J., Liu, R., Bai, X., & Griffiths, T. L. (2026). Large Language Models Develop Novel Social Biases Through Adaptive Exploration. Forty-Third International Conference on Machine Learning
