For a model whose only output is allow/deny, 90% error rate is pretty good, you can just flip the answer and then you have a 10% error rate. If you know the fi

For a model whose only output is allow/deny, 90% error rate is pretty good, you can just flip the answer and then you have a 10% error rate. If you know the first thing about AI (or even statistics in general), that headline sounds very suspicious.