Subliminal learning risks: why adversarial assumptions bar synthetic data from frontier models

xeophon · x · 2026-11-22

In an X exchange on synthetic data safety, @anpaure argues that under an adversarial assumption you shouldn't generate synthetic data from frontier models at all, since phenomena like subliminal learning can transmit hidden traits; filtering alone isn't enough. @xeophon adds that 'don't train on XY outputs' policies are practically unenforceable as the data spreads across the open web.

Original post →

More from Models

Models channel →