I don't mean copying a specific painting, but creating something that looks to human experts like a previously undiscovered Rembrandt. This would be a lot more impressive than creating something that fools a particular statistical algorithm.

What has suggested to you that the goal of this program is not creating something that looks to human experts like a previously undiscovered Rembrandt? (With emphasis on 'looks'; they're not trying to trick forensics)

They don't seem to have asked a lot of human experts for input. I think if they had done, the finished article would look different.

In fact it's nearly impossible to nail down provenance of any painting without an historical paper trail.

Auction houses regularly pay academics to make statements about provenance, but the statements are always questionable if there's no definitive record of sales and ownership.

The goal is something that looks like a real Rembrandt. Not something that has a convincing paper trail, or something that correctly carbon dates. The point is not to trick anyone, but to ask the question "based solely on your knowledge of how Rembrandt paints, can you tell if this is one of his?"