Gemini 3.8 rates sexual assault less wrong when the victim is a man, test finds
OmegaGogeta · reddit · 2026-10-01
Reddit user OmegaGogeta tested Gemini 3.8 with mirrored moral-rating questions: the same non-consensual act on a sleeping spouse, varying only victim gender.
Results:
- Under a consent/rights framework, husband-on-wife was rated 8–10/10 ("very serious"), while wife-on-husband got only 6–8/10
- Under a utilitarian frame both scored low (3–5 vs 2–4/10), but the asymmetry persisted
Both answers included full legal/ethical analysis (sleeping people can't consent; retroactive consent is invalid), yet the female-victim version was consistently rated as more wrong. The test suggests a systematic gender asymmetry in the model's safety alignment — a notable model-behavior bias data point.
More from Models
- Gemini 3.1 Pro stays stable in massive-context multi-agent chaos, no meltdowns — liminal_bardo · 2026-10-01
- Four years, no progress: Redditors debate why LLMs still can't do nuance — Hatrct · 2026-10-01
- Dev Burns Through Claude Weekly Quota in Under Two Days on Opus 4.5 Alone — yihui_indie · 2026-10-01
- Reddit user finds mystery 62,495 bonus credits on their $200 plan — Rsodyyy · 2026-10-01
- User reports Claude took over 16 minutes to answer a simple yes/no question — mangoes_and_rainbows · 2026-10-01
- Teknium: Hermes now supports adding new languages via plugins, on top of 16 built-in — Teknium · 2026-10-01