This seems like it was proven ages ago with the no-free-lunch theorem.

Humans could not learn to function unless their brains encode a useful prior for learning about the world. That prior means "preconfigured with instructions for understanding the world".

The short form of the no-free-lunch theorem is that if there is no prior (i.e. all possible universes are equally likely) then for any learning problem P there are an equal number of universes that learning system A will outdo learning system B on that problem.

If not all universes are equally likely, one learning system can vastly outdo another or even most other learning systems. Not equally likely is the assumption built into brains. Without that, you can't learn effectively.

So the biology is just implementation of that general principle. The details of how that implementation works are interesting, but whether we are preconfigured for learning was never in question.

And we've proven empirically that this usable "prior" can be quite small.

What exact assumptions does human brain encode and how does it use them, however, is an area of research. We are nowhere near being able to list out all of those useful inductive biases - let alone extract them and apply them to AIs.