But you aren’t using someone else’s service. You are using model weights of something that is explicitly uncopyrightable; like public domain.
It may be against XYZ’s TOS to scrape public domain content from their website, but if you obtained that content elsewhere, XYZ’s TOS doesn’t apply to you.
Laundering stolen property is not a means of escaping the law - and the terms of service were part & parcel of the related code, models, weights, etc. If they were not included in the related downloads by a third-party, upon discovery, Facebook would simply send a cease and desist notice to the party currently violating the terms with a copy of the terms, at which point, they would be in direct violation of the terms. If party failed to comply, Facebook would also send legal notices to anyone contributing to the breach of contract such as payment providers, hosts, etc.
Dollars to donuts any judge would rule that the 7 billion weights constitute a copyright violation because you didn’t spontaneously come up with those weights on your own.
Heck, this “it’s just a number” logic didn’t pan out for DeCSS [1] and that was a fixed small encryption key where it wasn’t as clear cut. So why would things suddenly shift here where there’s a very clear line that this is a derived work from FB’s model? The only thing that would help you I think is if FB decided to not enforce their copyright but I don’t think they’ve made clarifying statements either way. If I recall correctly, unlike trademarks, copyrights don’t have requirements that you must go after all instances of infringement to maintain your rights so FB can be totally selective about which instances they choose to prosecute.
Copyright isn't some magical concept that is applied on the whims of a judge. It has very specific conditions of which at least 2 aren't satisfied by model weights (creativity, human authorship).
The DeCSS legal issues have nothing to do with the copyrightability of ai models and stem solely from it circumventing copyright protection.
If you want to restrict models then propose new laws. Don't try to pretend that coycopyright somehow applies here.
It may be against XYZ’s TOS to scrape public domain content from their website, but if you obtained that content elsewhere, XYZ’s TOS doesn’t apply to you.
It’s a weird world with AI and copyright.