Originally Posted by Wautier

It's unchartered territories, but common sense dictates that you would have to have trained it on content that you own or have permission to use.

So, should the authorities or anyone else need any proof, you would simply show the documentation of the original dataset, and if you really do own it, and if it wasn't some type of usenet or tumblr 100GB pack of amateur pictures shit, you should be just fine.

But, train it on potentially-CP type of selfies, you could get into trouble, maybe.. hypothetically -- eventually... But, remember the unchartered territories part... it's the wild wild west.