Signal Gender Bias Across LLMs Is Common but Highly Inconsistent Across Vendors

Summary

A new arXiv paper released September 29, 2026 compared gender bias across ten LLMs from nine vendors released between April 2025 and June 2026. The study used two paradigms: attributing stereotyped phrases to a gender, and moral judgments about protecting a woman or man from harm. In the stereotype-attribution task, two of ten models attributed masculine-stereotyped phrases to female writers more often, while three showed the opposite pattern. In the moral-judgment task, several models showed a male-disadvantaging asymmetry consistent with a known human tendency to protect women, but the conditions varied by model and three models showed no variation at all. The authors conclude that gender bias is common across LLMs but its direction and magnitude are highly heterogeneous, sometimes diametrically opposite between models. They argue bias auditing must be an ongoing, multi-vendor process rather than a one-time check.