The judge's decision in the Getty vs. Stability AI case today is very disappointing. Stability was judged not guilty of 'secondary infringement'. The case turned on the question of whether Stability's model weights are an 'infringing copy' of what they're trained on. Judge said no. I think this is wrong for 2 reasons: 1. Law is clear: something is an infringing copy if you import it into the UK and "its making in the UK would have constituted an infringement of the copyright in the work in question". That is clearly the case here. The judge said the model weights aren't a copy - but that doesn't seem to align with this portion of the law, which IMO defines an 'infringing copy' very clearly (read it below). 2. It is assumed throughout that Stable Diffusion doesn't contain copies of its training data. I think this is incorrect. If it can recreate its training data (which it can), it functionally contains copies. At the extreme, if I write an algorithm that perfectly recreates a single work of art, this judge would seem to be suggesting that the algorithm doesn't contain a copy. But the copy *is* stored - it's simply a question of *how* it is stored. The case is obviously limited in scope, as the primary infringement claims were dropped because Getty couldn't prove that development happened in the UK. But it's still disappointing, and I hope it will be appealed (I don't know if this is possible - perhaps others do).