I don't know how much "hidden" knowledge there is in AI research. In biotech stuff, those can make or break an experiment easily and there is just so many of them and they are all trivial when known but absolutely critical. Some of them actually has zero reason to exist but they still exist, similar to a block of legacy code that was commented as "do not remove, shit breaks when removed". Nobody knows why and everybody just follow it religiously.

So the same approach may or may not work for AI. Forcing stuff to go underground in AI research may not hinder them that much...