- cross-posted to:
- hackernews
You must log in or register to comment.
“None of Ross’s possible defenses holds water. I reject them all,” wrote US District Court of Delaware judge Stephanos Bibas, in a summary judgement.
The AI company lost. Good news for a change.
Paywall
Thanks works fine, I can see you also included it as part of the post. 👍
tragic. no one should need to pay to read the law
That’s literally not what the ruling is about. It was about an AI bro company using proprietary, copyrighted materials to train its AI, which they obtained by questionable means, after being denied license to do so by the IP owners. Further, after training the AI with unlicensed materials, they launched a competing product.
Whether you support IP or not, the AI company is clearly in the wrong here.
It’s a pretty definitive example of many AI companies being little more than leeches, stealing others’ work and repackaging it as their own. All with zero long-term consideration of “what do we do when there’s noone left to leech off of because we undermined the ability of those make the source data to make a living, while unnecessarily driving increased emissions and consumption of potable water for something that provides little actual value do humanity as a whole?”
Whether you support IP or not, the AI company is clearly in the wrong here.
they’re both wrong to restrict access. if legal analysis is necessary to understand the law, then restricting access to that analysis, or it’s free dissemination, is also wrong.
Ah, fuck, is that what the case is about? That sucks; that’s the kind of case where they both need to lose:
- The law shouldn’t be copyrightable
- AI companies shouldn’t be allowed to ‘launder’ copyright (and more to the point, copyleft) by reproducing chunks of copyrighted works divorced from their license
If I were more conspiracy-minded, I would almost think that somebody intentionally decided to resolve this case first in order to guarantee that they set a disastrous precedent.
It’s not what this case about. Reuters runs a service called Westlaw that provides access to a bunch of legal materials, including summaries and explanations of cases that are written by its lawyers. Ross Intelligence wanted access to those summaries, so that it could train AI to make a competing product. As you can imagine, Reuters said no to this.
So, Ross bought summaries from someone else, another company that did have access to Westlaw, and used those to train its AI. Today, the court found (among other things), that a few thousand of the summaries that Ross’s AI produced are way too similar to Westlaw’s summaries for it to be a coincidence. Ross had argued (among other things) that its summaries were only similar because they were describing the law, and Reuters doesn’t/can’t have a copyright on the law. The court rejected this argument, saying, essentially “Yeah, it’s true that Reuters doesn’t have a copyright on the law, but it does have a copyright on the summaries that its lawyers write. It takes skill and judgment to decide which parts of a law or decision are important for people doing legal research, and to present them in a way that’s easy for people to understand. You clearly copied many of them.”
This isn’t an exhaustive discussion of all the issues covered in the opinion, because I’m a sleepy lawyer, but it’s the most important part.
Today, the court found (among other things), that a few thousand of the summaries that Ross’s AI produced are way too similar to Westlaw’s summaries for it to be a coincidence.
This is probably just inevitable when your dataset is not large enough. I would be interested in seeing the LLM’s output compared against the original texts; I do remember the early ChatGPT producing some borderline copies of sentences that you could find online (with one or two words changed).
I’m not a lawyer, but I’m also not entirely unfamiliar with this sort of thing. In particular, I remember Georgia v. Public.Resource.Org and thus do not accept at face value the notion that the data in question being “summaries and explanations of cases” necessarily means Westlaw is in the right. Even if the Westlaw materials aren’t “officially” incorporated into the law itself the way Georgia did, that doesn’t mean Westlaw should necessarily be entitled to monopolize them, especially if the judicial system is heavily leaning upon them to inform its decisions.
Though the headnotes were drawn directly from uncopyrightable judicial opinions, the court analogized them to the choices made by a sculptor in selecting what to remove from a slab of marble. Thus, even though the words or phrases used in the headnotes might be found in the underlying opinions, Thompson Reuters’ selection of which words and phrases to use was entitled to copyright protection. Interestingly, the court stated that “even a headnote taken verbatim from an opinion is a carefully chosen fraction of the whole,” which “expresses the editor’s idea about what the important point of law from the opinion is.” According to the court, that is enough of a “creative spark” to be copyrightable. In other words, even if a work is selected entirely from the public domain, the simple act of selection is enough to give rise to copyright protection.
The court distinguished cases holding that intermediate copying of computer source code was fair use, reasoning that those courts held that the intermediate copying was necessary to “reverse engineer access to the unprotected functional elements within a program.” Here, copying Thompson Reuters’ protected expression was not needed to gain access to underlying ideas.
It sounds like the case you mentioned had a government entity doing the annotation, which makes it public even though it’s not literally the law.
Reuters seems to have argued that while the law and cases are public, their tagging, summarization and keyword highlighting is editorial.
The judge agreed and highlighted that since westlaw isn’t required to view the documents that everyone is entitled to see, training using their copy, including the headers, isn’t justified.It’s much like how a set of stories being in the public domain means you can copy each of them, but my collection of those stories has curation that makes it so you can’t copy my collection as a whole, assuming my work curating the collection was in some way creative and not just “alphabetical order”.
Another major point of the ruling seems to rely on the company aiming to directly compete with Reuters, which undermines the fair use argument.
I don’t trust that judge’s ability to determine whether they were copied if it wasn’t verbatim. which is what copyright is. to control an idea, you need a patent.
I don’t think that’s the best argument in favor of AI if you cared to make that argument. The infringement wasn’t for their parsing of the law, but for their parsing of the annotations and commentary added by westlaw.
If processing copy written material is infringement then what they did is definitively infringement.
The law is freely available to read without westlaw. They weren’t making the law available to everyone, they were making a paid product to compete with the westlaw paid product. Regardless of justification they don’t deserve any sympathy for altruism.A better argument would be around if training on the words of someone you paid to analyze an analysis produces something similar to the original, is it sufficiently distinct to actually be copy written? Is training itself actually infringement?
So if stealing copyrighted information to train an AI isn’t fair use, then isn’t pretty much every commercial LLM illegal?
