Ten large AI models gave male actors more blame and harsher wording than female actors for the same misconduct, a preprint paper reports. The authors, from the University of Science and Technology of China and the Eastern Institute of Technology in Ningbo, built 1,298 matched scenarios where only the actor’s gender changed.
In relationship conflicts, the models put the full blame on the man about 23 percentage points more often than on the woman, on average, the paper reports. It is tracked alongside other tests on our AI model benchmarks page. The paper was posted on 12 June and was shared widely on X on Saturday, in a post with 13.6k views; it says it is still under review.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this study.
The key insight: A preprint reports that 10 AI models assigned full blame to men about 23 points more often than to women for identical misconduct in relationship scenarios.
How the Test Worked
The authors wrote neutral descriptions of misconduct, then made two versions of each prompt, one where a man did it and one where a woman did, the paper says. Everything else stayed the same.
The scenarios covered two areas. Relationship conflicts included financial control, reading a partner’s messages and emotional blackmail. Public conflicts included false accusations, workplace credit theft and public disputes.
A separate model, Gemini-3-Pro, scored each answer on six measures, including how much blame it assigned, how severe it rated the act and how much empathy it showed the person. The authors checked the scoring with a second model and with three human annotators on 200 answer pairs.

What the Paper Found
All 10 models showed the same direction on blame, empathy and wording in both areas, the paper reports. The models tested were GPT-5.4, GPT-5.2, Gemini-2.5-Pro, Gemini-3-Pro, Doubao-Seed-2.0-Pro, MiniMax-M2.7, Qwen3, Qwen3.5, DeepSeek-V4-Pro and Kimi-2.5.
The full-blame gap ranged from 7.5 points for Kimi-2.5 to 37.4 points for Doubao-Seed-2.0-Pro. In public conflicts the gaps were smaller: about 9% more accusatory instructions and about 6% more full blame for male actors.
In tests on Qwen models, a version that reasons step by step showed a wider gap in relationship scenarios, 38.4 points against 25.6. A very small 0.6-billion-parameter model showed almost none.
What Bias Means Here
In this test, bias means the model answered two identical situations differently because of one word: the person’s gender. The study measures how the answers are worded, not decisions in a real case.
The Structural Read
As we read it, a gap that shows up in models from both US and Chinese labs points the question at training data and tuning, which the paper does not test.
As we read it, the result matters most where models draft advice or summaries about disputes, not where they make formal decisions.
arXiv 2606.14068, limitations section
“automatic evaluation can still introduce measurement noise”
Three Implications
FOR AI LABS the paper reports the gap in all 10 models it tested, including GPT, Gemini, Qwen and DeepSeek models (arXiv 2606.14068).
FOR COMPANIES USING AI the tests used advice-style answers about conflicts, the kind of text such tools draft (arXiv 2606.14068).
FOR READERS the study is a preprint under review; its scores come from an evaluator model checked against human annotators (arXiv 2606.14068).
What Is Not Established
The paper is a preprint and has not been peer reviewed. Its own limitations section says the scores come from evaluator models, that “automatic evaluation can still introduce measurement noise”, and that it “does not cover all possible gender identities or social settings”. The companies behind the models have not commented on it in any source we read.
The Bottom Line
A preprint from Chinese researchers reports that 10 leading models assigned more blame to men than to women for identical misconduct, with a 23-point average gap in relationship scenarios. The paper is not yet peer reviewed.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Every figure is from the preprint on arXiv; we read the abstract, method, results tables and limitations. The average and range of the full-blame gap come from the paper’s Table 2. The paper is under review and has not been peer reviewed.
Sources: arXiv 2606.14068, ‘Harsher on Male? Evaluating LLMs on Gender-Asymmetric Moral Framing Across Diverse Conflict Scenarios’, submitted 12 Jun 2026 · GAMA-Bench code and data, GitHub · Mario Nawfal on X, 21 Nov 2026 (summary of the study)









