> If the outputs were sufficiently different from the most similar training examples would it be exempt from copyright liabilities?
This is a great kind of question to ask, and can point us to the right areas of non-deep-learning copyright to think about – but not in a way that lends itself to a quick settled answer in a comment section, I suspect. https://law.marquette.edu/facultyblog/tag/fairey-v-ap/
> On the other extreme, a single copyrighted image in the training set could spoil the whole model, it would be impossible to make sure the training set is 100% clean.
Where possessing stolen goods is illegal, it takes quite a lot of effort to determine that an e.g. antiquities dealer has met the reasonable standard of inquiry for 100% of their inventory – but that doesn't mean that isn't the legal requirement. That's an area where we know the legal standard, and this is much newer ground, so it very well may be that tech that depends on datasets too big to take responsibility for may not end up being a great foundation for anything with legal liability attached.
It will be a Napster moment all over again, people want to generate images they don't care about copyrights, and the technology is here. There's already a good enough model (Stable Diffusion) released into the open, it can be executed on any desktop computer. This tech is still very new, it will mature in a few years and we'll get used to it.
This is a great kind of question to ask, and can point us to the right areas of non-deep-learning copyright to think about – but not in a way that lends itself to a quick settled answer in a comment section, I suspect. https://law.marquette.edu/facultyblog/tag/fairey-v-ap/
> On the other extreme, a single copyrighted image in the training set could spoil the whole model, it would be impossible to make sure the training set is 100% clean.
Where possessing stolen goods is illegal, it takes quite a lot of effort to determine that an e.g. antiquities dealer has met the reasonable standard of inquiry for 100% of their inventory – but that doesn't mean that isn't the legal requirement. That's an area where we know the legal standard, and this is much newer ground, so it very well may be that tech that depends on datasets too big to take responsibility for may not end up being a great foundation for anything with legal liability attached.