One in three AI scribe notes carries a verified clinical error

What's the error rate for humans?