Because, even today, you can’t and will never have a 100% reliable answer.

You need to have at least 2 different validators to reduce the probability of errors. And you can’t just say, let’s run this check twice by AI as they will have the same flaw. You need to check it with a different point of view (being in term of technology or ressource/people).

This is the principle we apply in aeronautics since decades, and even with these layers of precautions and security, you still have accident.

ML is like the aircraft industry a century ago, safety rules will be written with the blood of the victims of this technology.