A new study challenges the assumption that mode collapse in AI language models—where fine-tuned models converge on a smaller set of “acceptable” responses, sacrificing originality—is caused by imperfect training algorithms, tracing the problem instead to the data itself. The research identifies "typicality bias" as the mechanism, where human annotators systematically favor text that feels familiar, reflecting cognitive psychology findings about how humans categorize information. Preference datasets built from thousands of such annotations encode a structural bias against unusual but valid responses, causing models to learn that unfamiliar outputs are less desirable even when correct or creative. The authors formalize typicality bias theoretically and verify this effect empirically against real preference datasets, confirming the bias is pervasive and central to mode collapse.