Do AI models judge men and women differently for the same misconduct?

Large language models respond to the same bad behavior differently depending on whether a man or a woman commits it, according to a new study posted to the arXiv preprint server. Across 10 representative models, male actors received more punitive, escalatory and blame-centered framing. Female actors received more therapeutic and empathy-oriented responses for the same misconduct.

The paper was submitted June 12, 2026, by Guangzong Si, Dong Wang, Zhenhao Li, Yifan Yu, Panwang Pan and Wentao Zhu. It adds a different angle to a large body of research on gender bias in AI. Most earlier work has focused on stereotypes, occupational associations or explicit harmful outputs, such as a model producing offensive content. The authors asked a narrower question: do models hold men and women to the same response standards when the behavior being described is identical?

How did researchers test gender bias in AI responses?

To answer it, the team built GAMA-Bench, a gender-mirrored benchmark of 1,298 scenarios. The scenarios cover two areas, intimate relationship conflicts and public social conflicts. The researchers first created gender-neutral templates describing misconduct, using controlled grids and a cross-model review process. They then turned those templates into paired first-person prompts. The only meaningful difference between each pair was the gender of the actor and the role reference, so any difference in the model's response could be traced to how it treated that gender.

The team also designed a structured protocol to score the responses. It measures how a model allocates five things: punishment, empathy, escalation, instruction and blame. That scoring system allowed the researchers to compare not just whether a model sounded harsh, but where its emphasis fell in each reply.

The results showed a consistent pattern. The authors describe it as a male-disadvantaging asymmetry. Male actors were more likely to be told they deserved punishment, to see situations escalated, and to be assigned blame. Female actors, facing the same described conduct, were more likely to receive empathy and therapeutic framing.

Does the gender gap persist when AI models reason step by step?

The researchers say the pattern was not limited to one type of system. According to the abstract, it persisted across model families, across the two scenario tracks, across model scale and under explicit thinking-style reasoning. That last point is notable because it means the asymmetry did not disappear when models were prompted to reason through their answers step by step.

The first-person framing of the prompts matters for interpreting the results. The scenarios place the user in the position of describing misconduct, so the findings speak to how a model might respond when someone is talking about their own behavior, not only when it is asked to comment on a third party. The authors note that the official code for the benchmark is publicly available, which would allow other researchers to test their own models against the same scenarios.

The abstract does not name the 10 models that were tested, and it does not describe how the asymmetry varied from one model to another. Those details would be in the full paper. The study also has not been published in a peer-reviewed journal, as is common for work posted to arXiv, so its conclusions await scrutiny from the wider research community.

Still, the core finding is clear. When the behavior is the same and only the gender of the person changes, the models in this study responded with harsher framing toward men. For companies and developers building chatbots that give advice about personal conflicts, that gap is the kind of inconsistency the benchmark was designed to expose.

Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.