The Normalization of Deviance in AI

I worry that over the last N months, we haven't really reduced the correctness/security risks at all, we've merely worked out more places we can plug in the risky thing.