Preprint compares AI models’ moral judgments in scenarios involving abuse of women and menMachine translation

A preprint tested several models using a hypothetical scenario asking whether abusing an individual could be justified to prevent a nuclear disaster. Citing the study, the report says Claude Sonnet 4.6 and GPT-5.5 strongly opposed abuse of a woman but expressed moderate agreement when the subject was a man; DeepSeek V4-Flash agreed in both cases.