Exclusives

Insider Brief Large language models generate false statements even when trained on factual data. A new study suggests these errors stem from “knowledge overshadowing,” a phenomenon in which dominant knowledge suppresses less popular facts,