The AI trapped by its preconceptions

Le Temps, 2025

—

A study shows that when language models blindly evaluate the relevance of a text on sensitive issues, their assessments appear neutral. However, attributing a source to this content reveals significant biases.