Dutch AI models show extreme bias against transgender people but little bias against gay people.

Computer text models can carry unfair stereotypes, but scientists have rarely checked Dutch models for anti-queer bias. To fix this, researchers built a Dutch dataset called WinoQueer-NL. It pairs sentences showing queer stereotypes with matching counter-stereotypes. They checked these ideas with Dutch queer participants to make sure the examples fit Dutch culture. Then, they tested computer models by comparing how likely a model was to choose the stereotypical sentence over the counter-stereotypical one. Overall, the models seemed neutral on average. However, deep down, the models showed huge unfair differences. Some models chose harmful stereotypes up to 97% of the time for transgender identities, but only 6% for gay identities. Transgender and non-binary identities faced the highest bias scores.

Why it matters

This matters because biased AI can reinforce harmful stereotypes for LGBTQ+ people in the Netherlands. By revealing how bias targets transgender and non-binary identities more severely, it points to the need for tailored solutions to make AI systems fairer for marginalized groups.

Questions answered