Human-in-the-loop can improve AI safety, but it usually does not by default. Putting a person behind an approval button only helps when the consequence…
