There is no "correct" way to use AI in order to avoid bad results. The only prudent approach is to assume all results are bad until proven otherwise.

But this approach negates much of the incentive to pay for questionable results.

> The only prudent approach is to assume all results are bad until proven otherwise.

As is true with results from people.

> But this approach negates much of the incentive to pay for questionable results.

I'm not sure that follows. Even the crack-addled human informant has always been paid for questionable results.