> Each training would be a separate infringement.

False. Training was found to be a legitimate use. The liability was specifically, solely, for copyright infringement specifically due to getting the works in the first place, not training on those works.

> And it could be argued that each use by a user of the model could be considered a separate infringement.

No, it could not.

If this standard was applied to copyright infringement on BitTorrent, someone who helped share one file to 100 other users would get hit with 100 copyright infringement instances, not one.

> Generally copyright fines are increased if the persons doing the infringing action know what they are doing. Aka, ‘willful infringement.’ It’s hard to imagine companies like OpenAI were unaware of the possibility of their actions being considered infringement.

That's already accounted for when I said this was in the normal range for liability per copyright violation.

> Often restitution of infringement includes money made by the infringer. So not simply, “your book is worth $3000.” But rather? “Your book is worth $3000 AND this company has derived an additional $50,000 of revenue from it.”

Depends on the details; however, as previously noted, the judge *explicitly noted* that training was not itself an offence, only the piracy to get the training data was. Any revenue derived from the offence had to be shown to be in the period between the offence and when they bought the same works, because they were found to be allowed to use those works in this manner.