The Clean Data Trap: How Hidden Traits Can Survive AI Filtering

A Nature study on subliminal learning suggests that language models can pass on bias and misalignment through data that looks harmless, forcing AI teams to think beyond content filters and toward lineage security.