Comment by mmahemoff
3 days ago
In this case, Party B is sandwiched between Party A (scraper such as Meta) and Party C (the third-party content provider). The scraping agreement dictates A-B relationship and the licensing agreement dictates B-C relationship. Devil is in the fine print as far as what is allowed in both agreements.
Party B has already committed to uphold B-C. Entering A-B may put them at risk of violating that agreement by appearing to grant Party A rights that Party C never granted to anyone but Party B (and maybe even rights that Party B doesn't enjoy). Party B's options are don't consent to A-B (e.g., withdraw from Meta program), take down the Party C content (which would suck if they've already paid for it), or negotiate a carve-out with one or both of Parties A and C.
(IANAL)
I am not an expert in licensing legalese, but that strikes me as awfully wrong. If I am e.g. a musician and I license my work to be used by B for specific purposes and B unilaterally signs an agreement with A that gives them rights to my stuff, the question is gonna be what MY licensing terms allow. They have an agreement with me. Whether they enter other agreements is their decision to make, but they still have to fulfill my terms.
Sure, they should not sign things that put them into the spot of having to break my licensing agreement with me, but the fact that these kind of legal stretches are even seen as valid in the US reassures me in my stance to avoid business relations with the US. I prefer countries with a legal system that can't easily be bent out of shape using money.
You really hit the nail on the head. I can't sign over rights to stuff I don't own. Is there a good chance that this move could blow up in Meta's face, in which the Party B's organize (or are ordered) en masse not sign the license agreement and leave Zuck without the legal ability to steal while also losing marketshare?
I think more and more people like you around the world are realizing not only how corrupt the U.S. is, but that it's power and influence are in a sharp state of decline and perhaps not worth the risk at this point.
That's the (evil) genius in Meta's terms: they get the right to scrap out of thin air, and the only party liable to the scraping is B.
IANAL, but copyright binds everyone regardless of entering into any agreement. In that case Meta would still be potentially liable for its infringement, exactly because "thin-air right" is not a thing. B could not grant it, because they don't have it.
A license is a contract or contract term waiving certain rights to sue. That’s what a copyright license is. And yes, you cannot issue a valid license if you don’t have the right to do that.
What these kinds of platform agreements usually say is that you, the licensor, agree that you are fully representing that you can issue that license, and that any infringement is thus your fault and not the platform-licensee’s. Under the DMCA and other sources of law this is plausible.
4 replies →
I hereby grant all HN users and not users the right to copy, modify and redistribute any and all text and binary on the Internet and outside the Internet, already existing or yet to be created, forever and 1 day.
The problem there is that both B and C are pulley much smaller than Meta. Copyright and other IP law at the scale of meta has become something that only really works for those with significant resources.
AFAIK scraping itself doesn't go against copyright, only redistribution matters with respect to copyright (which is why B and C must have an agreement together).
Buying (or stealing) stolen goods is still a crime.
If they really want to be spiteful they can rewrite all the news articles they post to Meta using Muse. Then we can finally see if the theories of statistical model collapse hold up at the trillion param level model scale.
I'd really be curious to see how much of Muse and similar 2026 models are actually from scraping, especially from this hypothetical unsorted random sites.
My understanding is that most of the gains after the early GPT-3 "add more weights" era have come from post-training from synthetic data, especially through reinforcement learning.
I couldn't quickly find any public statements about the actual breakdown of training data, other than almost every lab saying that post-training size was "scaling" or growing "by an order of magnitude" from their previous major release.