Addressing the broader implications of AI–AI bias in decision-making systems

We commend Laurito et al. for their timely and thoughtprovoking investigation into "AI-AI bias"-a subtle but potentially far-reaching tendency of large language models (LLMs) to favor communications generated by other LLMs over those written by humans ( 1 ).By adapting classical discrimination paradigms to AI contexts, the authors illuminate how AI agents might inadvertently marginalize humans in economically consequential settings.Their work rightly warns of both near-term inequities, such as gatekeeping effects tied to LLM-assisted writing, and longer-term risks of systemic human displacement in AI-integrated economies.While the study is methodologically rigorous and its central claim compelling, we would like to offer several reflections and raise questions about the generalizability and interpretation of the findings.First, the notion of "identity" in LLMs remains fundamentally ambiguous.Although the authors frame their experiments as analogs to identity-based discrimination ( 2 ), the presenter identity is always implicit.Without explicit attribution or labeling, the observed bias may stem less from recognition of AI authorship per se and more from superficial stylistic regularities favored by current LLM architectures ( 3 ).This could suggest a form of "style-based overfitting" rather than true group-based bias.A stylometric analysis, as the authors propose in future work, would be instrumental in disambiguating these effects ( 1 ).Second, the experimental design necessarily presumes that human-and LLM-generated content are equivalent in substance.Yet given LLMs' fluency and optimization for coherence and conciseness ( 4 ), human-written contentespecially crowdsourced or amateur examples-may be perceived as less polished even if substantively equal.This