[cauch]

I doubt it is so naively simple as "querying a database". For example, before the publication of this article, the majority of the "databases" were simply saying "the inventor is Cort", which is one example of thing that the A.I. will get trivially "wrong" when reviewing the paper.

And if it is true, then this A.I. review method can also be applied to mathematics, theoretical physics, computer science, ...

And even if it is doable, you will still have things that will pass the review process when ideally it should not. It is just an hard limitation, same as the one in "justice" (impossible to not sometimes judge a guilty person "innocent" or an innocent person "guilty", anyone who thinks otherwise just don't understand how complicated it is)