GAMA-Bench: 10 LLMs castigan más al actor masculino
Reported by Hacker News and ecosistemastartup2 sources ↓
Qué es GAMA-Bench y qué mide exactamente Un benchmark de 1.298 escenarios y 10 modelos de lenguaje encontró que, ante la misma mala conducta, la IA ...
Key points
- The benchmark found more punishment, escalation and blame when the actor was male
- Female actors received more empathetic and therapeutic responses for the same misconduct
- The test covered intimate-relationship and public social conflicts
- The report says the pattern persisted across model families, sizes and reasoning settings
- The article does not name the 10 models
Sources
Details
Ecosistema Startup reports that GAMA-Bench uses 1,298 scenarios and 10 language models to study gender differences in how models frame moral conflicts. The benchmark presents each scenario twice, changing the actor’s gender and related role references while keeping the misconduct the same.
The assessment measures punishment, empathy, escalation, instruction and blame. The article says the results showed more punitive, escalating and blame-focused framing for male actors, and more therapeutic, empathetic framing for female actors.
According to the article’s account of the preprint, the pattern appeared across model families, scenario types and model sizes, and remained when explicit reasoning was enabled. It does not identify the models tested; the authors’ paper is described as a preprint, and the article says benchmark code is available.
Written from coverage as of
Images
- ecosistemastartup