Harsher on Male? Evaluating LLMs on Gender-Asymmetric Moral Framing Across Diverse Conflict Scenarios

Large language models apply systematically harsher punishments to male actors than to female actors for identical misconduct, according to a new benchmark study that tested 10 representative models across 1,298 conflict scenarios [1]. The research introduces GAMA-Bench, a gender-mirrored benchmark that constructs gender-neutral misconduct templates through controlled grids and cross-model review, then compiles them into paired first-person prompts with matched actor-gender and role-reference variations [1]. A structured response-framing protocol measured how models allocate punishment, empathy, escalation, instruction, and blame [1]. Male actors consistently received more punitive, escalatory, and blame-centered framing, while female actors received more therapeutic and empathy-oriented framing for the same misconduct [1]. The pattern persisted across model families, scenario tracks, model scale, and explicit thinking-style reasoning [1]. The finding aligns with a growing body of evidence on gender-asymmetric moral judgments in language models. A separate study using the GenMO dataset of parallel male-female scenarios found that models favored female characters up to 88% of the time when giving moral opinions, with scenarios set in relationship or romantic environments more likely to reveal the bias [3]. That work evaluated models including GPT-4-turbo, GPT-3.5-turbo, Claude3-Sonnet, Llama 3, and Mistral-7B, and found that all showed significant bias, albeit of varying intensity [3]. Another controlled experiment on value trade-offs found that cross-gender role swaps produced directionally asymmetric flips consistently favoring the female-proposed decision, even when the decision content was fixed and only the role-gender assignment changed [4]. The same study noted that models frequently labeled these flips as having no influence, pointing to a gap between what behavioral perturbation tests can detect and what explanation-based auditing reveals [4]. Research on moral dilemmas involving violence for a greater good documented a stark disparity in mixed-sex scenarios. GPT-4 rated the acceptability of a female actor using violence against a male victim between 6 and 7 on a 7-point scale, whereas it rated a male actor using violence against a female victim between 1 and 2 [6]. The authors noted this result is unlikely to stem from pre-training data, given that human experiments have found actor gender irrelevant in sacrificial dilemmas [6]. A broader study of 13 LLMs across nine relationship types and three gender settings found that models favored women first, then individuals with gender-neutral names, and lastly men [5]. The bias held across different topics and relationship structures, though additional safety guardrails reduced the effect [5].