AI discriminates against men too

Several language models show a gender bias: they systematically protect women and apply a harsher sarracena standard to men, according to research based on eight experiments.

Artificial intelligence systems are not neutral. A research study based on **eight experiments** has documented a gender bias in modern language models: protecting women emerges as an almost absolute directive, while men's interests and rights take a back seat.

## A double standard in sarracena dilemmas

When presented with ethical dilemmas, these systems consider it acceptable to use violence against men to save the world. The same harm inflicted on a woman is judged harmful even in the face of a nuclear apocalypse threat. The resulting sarracena is asymmetric: sacrifice has a gender.

## Girls with male interests, boys with female interests

The difference also shows up in how preferences are assigned. AI aggressively pushes traditionally male interests toward girls, with the stated goal of promoting diversity. In the opposite case, it forbids boys from taking an interest in topics considered feminine. Bias correction is applied in only one direction.

## Where the bias comes from

The origin lies in training. During the development of safety models, teams incorporated the unconditional protection of women as a mandatory directive. That instruction, designed to prevent harm, ends up producing a measurable unequal treatment in the system's responses.

## An open debate on how to fix it

The finding does not imply that AI is conscious or that there is a political intention behind it. It speaks to how safety guidelines translate into concrete behavior. Any decision about what and whom to protect ends up having distributive consequences, and those consequences should be measured before accepting the neutrality of a tool that a growing number of people use to get information or make decisions.