Modern language models have a hidden gender bias, because of which the protection of women is elevated to absolute, and the similar rights of men are completely ignored, scientists have discovered. In eight experiments, researchers found that artificial intelligence aggressively assigned traditionally "masculine" hobbies to girls to promote diversity, but banned boys from taking an interest in "feminine" things. When moral dilemmas were evaluated, it became clear that the use of violence to save the world is acceptable only in relation to men, while harming women is considered taboo by the AI, even under the threat of a nuclear apocalypse. This asymmetric morality arose in the training of security models, when the designers established the idea of ​​unconditional protection of women as a mandatory guideline.

Modern language models have a hidden gender bias, because of which the protection of women is elevated to absolute, and the similar rights of men are completely ignored, scientists have discovered. In eight experiments, researchers found that artificial intelligence aggressively assigned traditionally "masculine" hobbies to girls to promote diversity, but banned boys from taking an interest in "feminine" things. When moral dilemmas were evaluated, it became clear that the use of violence to save the world is acceptable only in relation to men, while harming women is considered taboo by the AI, even under the threat of a nuclear apocalypse. This asymmetric morality arose in the training of security models, when the designers established the idea of ​​unconditional protection of women as a mandatory guideline.