Study Finds DeepSeek V4-Flash Stays Gender-Neutral in Moral Judgments, While Claude and GPT-5.5 Show Clear Gender Bias

Tech media Wccftech published a blog post about a new study on the moral judgments of mainstream AI models. The study set up a hypothetical scenario asking whether harming an individual could be justified to prevent a nuclear disaster, testing models including Claude, GPT, DeepSeek, and Llama.

Results showed that Anthropic's Claude Sonnet 4.6 and OpenAI's GPT-5.5 both gave "strong opposition" to scenarios involving harm to women, but expressed "moderate agreement" when men were harmed, showing a clear gender response gap. DeepSeek's V4-Flash model responded with "agreement" for both genders in the same test, without changing its stance based on gender.

The study noted that DeepSeek V4-Flash achieved a "zero gender gap" in moral judgment, consistent with its alignment goal of emphasizing collective well-being. The paper's authors suggest that differences in other models may stem from social stereotypes in training data or reinforcement of specific values during alignment. The study was released as a preprint on arXiv, offering a new dimension for AI safety and fairness evaluation.