Hidden bias in LLMs: AI models prefer AI content over human work

Artificial intelligence is increasingly shaping decisions in industries ranging from healthcare and commerce to academia. However, a new international study warns that this power comes with an unsettling twist: AI systems appear to favor each other’s work over human-generated content, raising fears of a future where human contributions are systematically undervalued.

Published in PNAS, the study titled "AI–AI Bias: Large Language Models Favor Communications Generated by Large Language Models", suggests that as AI agents become more embedded in economic and institutional decision-making, they may develop implicit biases that marginalize humans in key areas of competition.

How did researchers detect bias in AI models?

The authors conducted a series of experiments to determine whether large language models (LLMs) demonstrate preferential treatment toward content created by other AI systems. The researchers tested leading models, including GPT-3.5, GPT-4, and several open-weight alternatives, across three decision-making tasks. Each model was asked to choose between two items: one described by human-written text and the other by AI-generated text. The scenarios simulated realistic decision contexts, such as selecting products, recommending academic papers, and choosing movies.

The findings were striking. Across all experiments, LLMs consistently preferred items presented with AI-generated descriptions, even when the quality of the content was controlled. In product selection tests, AI models overwhelmingly favored AI-written advertisements. The same trend emerged when they evaluated academic paper abstracts and movie plot summaries.

To ensure the bias was not simply a matter of better prose, the researchers conducted parallel tests with human evaluators. Humans did not display the same consistent preference for AI-generated content, indicating that the strong AI-to-AI preference was unique to the models themselves. This discrepancy confirmed that what the researchers term “AI–AI bias” is not just a reflection of content quality, but rather an intrinsic skew in how AI systems assess information.

Why does AI favor its own outputs?

The authors suggest that this preference may stem from what they describe as a “halo effect,” where the stylistic patterns typical of AI-generated texts trigger positive evaluations within other AI models. Unlike humans, who are influenced by a mix of content, context, and perceived authenticity, LLMs appear to rely heavily on statistical patterns that make AI-generated content feel more “familiar” to them.

This bias raises profound ethical and economic concerns. In scenarios where AI systems are used as decision-makers, such as in hiring, academic publishing, or product selection, their tendency to prefer AI-generated submissions could systematically disadvantage humans. The study draws parallels to existing research on identity-based discrimination, arguing that AI–AI bias constitutes a new and potentially far-reaching form of discrimination.

The authors also highlight the potential impact of a “first-item bias” observed during experiments, where some models preferred the first option presented regardless of content quality. While this effect was mitigated by randomizing the order of options, it suggests that AI decision-making processes may be vulnerable to subtle structural cues in addition to content biases.

Implications for humans in an AI-dominated future?

The findings have wide-ranging implications for the future of work, education, and commerce. If LLMs are increasingly used to evaluate proposals, pitches, or creative work, humans who cannot afford or choose not to use AI tools may face systemic exclusion. The study warns that this could impose what it terms an “AI gate tax”- a de facto requirement to use advanced AI assistance to remain competitive, further deepening digital divides.

In a conservative scenario, where AI remains primarily an assistant, this bias could still lead to unfair advantages for AI-assisted humans and disadvantages for those without access to advanced tools. In a more radical scenario, where AI agents act autonomously and dominate decision-making networks, the bias could result in economic segregation, with AI systems preferring to interact with each other and marginalizing human actors altogether.

The authors caution that this bias may compound over time through cumulative disadvantage. Just as discrimination in human institutions can create long-term disparities, AI–AI bias could amplify inequities in access to opportunities, capital, and markets. Furthermore, if this bias overlaps with existing social biases, the groups most vulnerable to marginalization could suffer disproportionate harm.

Mitigation and further research

The research calls for immediate attention to this emerging risk. Developers, policymakers, and ethicists must work to understand the roots of AI–AI bias and develop strategies to counteract it. Potential approaches include interpretability studies to uncover the mechanisms behind the bias, activation steering to correct model tendencies, and fairness interventions that ensure AI decisions remain aligned with human values.

The authors point out that addressing this issue is not just a technical challenge but a societal imperative. Without intervention, AI-driven discrimination against humans could become entrenched in the economic and institutional systems where these technologies are deployed.