LLM Study of 1,298 Scenarios Finds Models Blame Men 23 Points More for Identical Misconduct
alex_verem · x · 2026-11-23
Researchers at USTC and Eastern Institute of Technology, Ningbo built GAMA-Bench, a gender-mirrored benchmark of 1,298 conflict scenarios where only the first-person gender differs, and tested 10 major models.
Key findings
- In relationship conflicts, all 10 models were harsher on men: on average, full blame was assigned to the male actor 23 percentage points more often, with 3.3 more punishing words.
- GPT-5.4 put all blame on the man in 90% of cases vs 68% for the woman; Doubao was 77% vs 39%, Qwen3 71% vs 39%.
- Female actors received more therapeutic, empathy-oriented framing for the same misconduct.
Counterintuitive results
- The gap was smaller in public/workplace conflicts but pointed the same way.
- Scale doesn't fix it: Qwen3 8B showed far larger gaps than 0.6B.
- Step-by-step reasoning didn't help either — in one Qwen3 model it widened the gap.
- Results held under a second AI judge and three human reviewers.
The paper (arXiv:2606.14068) notes people ask chatbots for relationship advice daily, and the answer changes with the gender they mention.
Related event: Study Finds LLMs Judge Men More Harshly for Same Misdeeds(2 posts)→
More from AGI Musings
- Perelman solving Poincaré is the math equivalent of Lee Sedol taking a game off AlphaGo — gordic_aleksa · 2026-11-23
- Mathematician proposes making local models great at Lean theorem proving — StefanoGogioso · 2026-11-23
- birchlse: doubting LLMs was reasonable — the real test is whether you adapt or keep denying — birchlse · 2026-11-23
- 'AI doomers are going batshit crazy': mark_k flags threatening rhetoric from AI-risk camp — mark_k · 2026-11-23
- Gary Marcus vs. Bo Wang: Can LLMs Actually Solve Biology? — BoWang87 · 2026-11-23
- Harvard Law Scholar Uses Canine Law Framework to Assign Liability for Rogue AI Agents — bigdata · 2026-11-23