[denton-scratch]

But they train their models on everything, regardless of the licence. It follows that the resulting derivative work likely mixes stuff that is under incompatible licences, with the result that it can't be distributed at all.