- Disney and NBCUniversal have teamed up to sue Midjourney.
- The companies allege that the platform used its copyright protected material to train its model and that users can generate content that infringes on Disney and Universal’s copyrighted material.
- The scathing lawsuit requests that Midjourney be made to pay up for the damage it has caused the two companies.
I say this as a massive AI critic: Disney does not have a legitimate grievance here.
AI training data is scraping. Scraping is — and must continue to be — fair use. As Cory Doctorow (fellow AI critic) says: Scraping against the wishes of the scraped is good, actually.
I want generative AI firms to get taken down. But I want them to be taken down for the right reasons.
Their products are toxic to communication and collaboration.
They are the embodiment of a pathology that sees humanity — what they might call inefficiency, disagreement, incoherence, emotionality, bias, chaos, disobedience — as a problem, and technology as the answer.
Dismantle them on the basis of what their poison does to public discourse, shared knowledge, connection to each other, mental well-being, fair competition, privacy, labor dignity, and personal identity.
Not because they didn’t pay the fucking Mickey Mouse toll.
You did not read your source. Some quotes you apparently missed:
Please read your source before posting it and claiming it says something it doesn’t actually say.
Now why does Doctrow distinguish between good scraping and bad scraping, and even between good LLM training and bad LLM training in his post?
Because the good applications are actually covered by fair use while the bad parts aren’t.
Because fair use isn’t actually about what is done (scraping, LLM training, …) but about who does it (researchers, non-profit vs. companies, for-profit) and for what purpose (research, critique, teaching, news reporting vs. making a profit by putting original copyright owners out of work).
That’s the whole point of fair use. It’s even in the name. It’s about the use, and the use needs to be fair. It’s not called “Allowed techniques, don’t care if it’s fair”.
Are you saying that the mere action of scraping is fair use, or that absolutely anything you do with the data you scrape is also fair use?
The Doctorow article does not say “scraping is good actually” - it says “scraping in X circumstance is good” and “scraping in Y circumstance is bad”, and wraps up by admitting the obvious and glaring contradiction.
I’d say that scraping as a verb implies an element of intent. It’s about compiling information about a body of work, not simply making a copy, and therefore if you can accurately call it “scraping” then it’s always fair use. (Accuse me of “No True Scotsman” if you would like.)
But since it involves making a copy (even if only a temporary one) of licensed material, there’s the potential that you’re doing one thing with that copy which is fair use, and another thing with the copy that isn’t fair use.
Take archive.org for example:
It doesn’t only contain information about the work, but also a copy (or copies, plural) of the work itself. You could argue (and many have) that archive.org only claims to be about preserving an accurate history of a piece of content, but functionally mostly serves as a way to distribute unlicensed copies of that content.
I don’t personally think that’s a justified accusation, because I think they do everything in their power to be as fair as possible, and there’s a massive public benefit to having a service like this. But it does illustrate how you could easily have a scenario where the stated purpose is fair use but the actual implementation is not, and the infringing material was “scraped” in the first place.
But in the case of gen AI, I think it’s pretty clear that the residual data from the source content is much closer to a linguistic analysis than to an internet archive. So it’s firmly in the fair use category, in my opinion.
Edit: And to be clear, when I say it’s fair use, I only mean in the strict sense of following copyright law. I don’t mean that it is (or should be) clear of all other legal considerations.
I think you make some compelling points overall, but fair use has always been more complex than this. The intent is taken into account when evaluating whether something is fair use, but so is the actual impact — “fair use” is a designation applied to the overall situation, not to any singular factors (so a stated purpose can’t be fair use)
Yes, that’s a good addition.
Overall, my point was not that scraping is a universal moral good, but that legislating tighter boundaries for scraping in an effort to curb AI abuses is a bad approach.
We have better tools to combat this, and placing new limits on scraping will do collateral damage that we should not accept.
And at the very least, the portfolio value of Disney’s IP holdings should not be the motivating force behind AI regulation.
Tbh, this is not a question about scraping at all.
Scraping is just a rather neutral tool that can be used for all sorts of purposes, legal and illegal.
Neither does the technique justify the purpose nor does outlawing the technique fix the actual problem.
Fair use only applies for a certain set of use cases and has a strict set of restrictions applied to it.
The permitted use cases are: “criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research”.
And the two relevant restrictions are:
(Quoted from 17 U.S.C. § 107)
And here the differences between archive.org and AI become obvious. While archive.org can be abused as some kind of file sharing system or to circumvent paywalls or ads, its intended purpose is for research, and it’s firmly non-profit and doesn’t compete with copyright holders.
AI, on the other hand, is almost always commercial, and its main purpose is to replace human labour, specifically of the copyright owners. It might not be an actual problem for Disney’s bottom line, but it’s a massive problem for smaller artists, stock photographers, translators, and many other professions.
That way, it clearly doesn’t apply to the use cases for fair use while violating the restrictions.
And for that, it doesn’t matter if the training data is acquired using scraping (without permission) or some other way (without permission to use it for AI training).
I think the distinction between data acquisition and data application is important. Consider the parallel of photography; you are legally and ethically entitled to take a photo of anything that you can see from public (ie, you can “scrape” it). But that doesn’t mean that you can do anything you want with those photos. Distinguishing them makes the scraping part a lot less muddy.