> The idea that AI training is fair use isn't so obvious

> Observe how all the major copyright victories amount to judges dismissing cases on grounds of "Well you don't have an example specific to your work" rather than addressing whether such uses are acceptable as a collective whole.

Well, all a judge can/should do is to apply current law to the case before them. In the case of generative AI then it seems that it's mostly going to be copyright and "right of publicity" (reproducing someone else's likeness/voice) that apply.

Copyright infringment is all about having published something based on someone else's work - AFAIK it doesn't have anything to say about someone/something having the potential to infringe (e.g. training an AI) if they haven't actually done it. It has to be about the generated artifact.

Of course copyright law wasn't designed with generative AI in mind, and maybe now that it is here we need new laws to protect creative content. For example, should OpenAI be able to copy Studio Ghibli's "trademark" style without requiring permission?

> Well, all a judge can/should do is to apply current law to the case before them

This is true, and I do not mean to suggest it is bad. But rather, that it leaves uncertainty. These cases can all be struck down without reducing the possibility that if one does stick, the entire industry is at stake.

> Copyright infringment is all about having published something based on someone else's work - AFAIK it doesn't have anything to say about someone/something having the potential to infringe (e.g. training an AI) if they haven't actually done it. It has to be about the generated artifact.

A notable problem here is that AI models are not "standalone products" but tools provided as a service. This complicates the situation.

Take Disney/Universal's case against Midjourney, which is both about the models but also the provision of services.

Even if only the latter gets deemed illegal, that's ruinous for the big AI companies. What good is OpenAI if they can't provide ChatGPT? Who would license a LLM if the act of using it creates constant legal risks?