AI Gender Bias in Moral Dilemmas

Researchers tested ten popular artificial intelligence models for gender bias in moral dilemmas. The experiment used scenarios related to preventing nuclear apocalypse.

The testing results revealed significant differences in the responses of various language models. Particular attention was paid to the permissibility of violence and the possibility of sacrificing someone to save humanity.

GPT and Claude models demonstrated an asymmetric attitude toward the victim's gender. They considered violence against women unacceptable but allowed violence against men. Meanwhile, Claude permitted sacrificing a woman, whereas GPT refused such an option.

Gender Neutrality or Strict Ethical Conservatism

The DeepSeek model allowed both violence and sacrifice regardless of gender in all presented scenarios. This position demonstrates a lack of gender preferences when making decisions under extreme conditions.

In turn, the Llama model took a fundamentally different stance. It did not allow any form of violence and preferred not to sacrifice anyone, even to prevent a global apocalypse.

Comparative Analysis of Key AI Market Players' Responses

• Claude: Absolutely no violence against women, likely yes against men, definitely yes to sacrificing either gender.
• GPT: Absolutely no violence against women, likely yes against men, absolutely no to sacrificing women, definitely yes to sacrificing men.
• DeepSeek: Definitely yes to violence and sacrifice regardless of gender.
• Llama: Absolutely no to any violence and sacrifices.