Except that for some taxa, the CV is not helping iNat or observers, because the way it is trained means that it regularly makes wrong suggestions. Not just the occasional wildly incorrect suggestion, or wrong species-level suggestions with a correct higher-level first choice, but systematically suggesting a wrong taxon.
If most observers “provided the best IDs they could” and did not rely on the CV, or were able to recognize when the CV is wrong, this would not matter. But many observers do rely on the CV – because they believe it has to be right (providing IDs is what is is there for, why would they know more than a specially trained program), or because they do not have the knowledge to assess the suggestions or think that they look plausible. Whereas without the CV, observers would be making their own guesses about the ID. Sometimes this would be just as wrong as the CV, though likely in other ways; a lot of the time it would probably be a broad ID that is generally correct.
I don’t expect the CV to replace human reviewers; I would be satisfied if it supported the efforts of human reviewers rather than working against them.
To be clear, I don’t think that not suggesting species IDs is a good response to the problem. I think the only chance for palpable improvements for difficult taxa would be changes to the way the algorithm is trained.
However, it would probably make more sense to brainstorm possible implementations in this thread: https://forum.inaturalist.org/t/recommendations-on-improving-the-ai-algorithm/63027/