Substantially similar is the standard
https://crsreports.congress.gov/product/pdf/LSB/LSB10922
> Under U.S. case law, copyright owners may be able to show that such outputs infringe their copyrights if the AI program both (1) had access to their works and (2) created “substantially similar” outputs
From Shaw v Lindheim.
> Their own lawyers would be on their case if they were infringing halfway through the pipeline.
This is why many of these AI companies are their own entities with sponsorships or shares held by other, larger companies. If they're on the hook for obscene copyright infringement, they just close down.