Tag: Anthropic

When the End Does Not Justify the Means, Anthropic’s $1.5 Billion Lesson

“Fair Use” Does Not Justify Piracy

A hand-written note on a white paper that reads 'END ≠ MEANS'.

Image: Author

The stunning announcement on September 5 that AI company Anthropic had agreed to a USD$1.5 billion out-of-court settlement to settle a class-action lawsuit brought by a group of authors was ground breaking in terms of its size, and goes to disprove the old adage that “the end justifies the means”. It is still not clear if the “end” (i.e. using copyrighted content without authorization to train AI algorithms) is legal, although preliminary indications are that at least in the US this may be the case. However, even if what Anthropic and other AI companies have been doing is ultimately determined to be fair use under US law—which is by no means certain—downloading and storing pirated content is clearly not legal, even if it is to be used for a fair use purpose. In other words, the piracy stands alone and must be judged as such, separate from whatever ultimate use to which the pirated content may be put.

Ironically, in the end, Anthropic did not even use much of the pirated content it had collected for training its platform, Claude. It seems to have had second thoughts about using content from online pirate libraries such as LibGen (Library Genesis) and PiLiMi (Pirate Library Mirror) and instead went out and purchased single physical copies of many works, disassembling and then digitizing them page by page for its Central Library, after which it destroyed the hard copies. Why go to all this trouble? Why not just access a legal online library? That’s because when you access a digital work, you don’t actually purchase it. You purchase a licence to use it, and that licence comes with conditions, such as likely prohibiting use for AI training. Anthropic would have been exposing itself to additional legal risk by violating the terms of the licence, so instead of negotiating a training licence, they took the easy way out by downloading content from pirate sites LibGen and PiLiMi. Later, having second thoughts, they purchased physical copies of the works they wanted to ingest and then scanned them. But it was too late. The piracy had already occurred.

When the decision in the Bartz v Anthropic case was released this summer, I commented that the findings were a mixed bag for AI developers. A very expensive mixed bag, it turns out. In the Anthropic case, there were clearly some interim “wins” for the AI industry. Anthropic’s unauthorized use of the works of the plaintiffs (authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, who filed a class action suit) was ruled by the judge (William Alsup) to be “exceedingly tranformative” thus tipping the scales to qualify as a fair use. In addition, he ruled that Anthropic’s unauthorized digitization of the purchased books to also be fair and not infringing. However, it was the downloading and storing of the pirated works that got Anthropic into hot water. Even though the intended use of the pirated works was to train Claude, a so-called transformative fair use, this did not excuse the piracy. While Alsup did not specifically rule that use of pirated materials invalidates a fair use determination (i.e. he ruled that the piracy and the AI training were separate acts), his ruling exposes a weak flank for the AI companies. For example, the US Copyright Office has stated that the knowing use of pirated or illegally accessed works as training data weighs against a fair-use defence. In short, the end does not justify the means.

The piracy finding was significant because Judge Alsup decreed that this element of the case would be sent to a jury to determine the extent of damages. (In Canada and the UK, judges rather than juries normally play this role). Given that under US law statutory damages start at $750 for each work infringed but can go up to $150,000 per work for willful infringement, Anthropic could have been on the hook for tens of billions of dollars in damages for the almost 500,000 works at issue. (Over 7 million works were inventoried by the pirate websites and downloaded by Anthropic but the limitations on who qualifies for the class action reduced the number of actionable works to just 7 percent of the total). As deep as its pockets are (Anthropic is backed by Amazon), if a jury awarded damages toward the higher end of the scale, the company could have been bankrupted.

Thus, Anthropic had lots of incentive to settle (including keeping the fair use findings unchallenged). As it stands, the $1.5 billion payout, while large in total, amounts only to about $3000 per infringed work, not the minimum but not really financially significant for the plaintiffs. This amount will probably have to be split between authors and publishers, with some of the funds covering costs, so no authors are going to be buying a new house on the proceeds. The real beneficiaries will be the law firms that represented them. The messy process of deciding who gets what that has led Judge Alsup to suspend the proposed settlement in its current form and require greater clarity as to how the payouts will be managed. The number of works eligible for payment is limited by the fact that to qualify they have to meet three criteria;

1) they were downloaded by Anthropic from LibGen or PiLiMi in August 2022

2) they have an ISBN or ASIN (Amazon Standard Identification Number) and, importantly,

 3) they were registered with the US Copyright Office (USCO) within five years of publication, and prior to either June 2021 or July 2022, (depending on the library at issue).

Any other works do not qualify. Registration with the USCO is not a requirement for copyright protection but in a peculiarity of US law, without registration a copyright holder cannot bring legal action in the US.

While the settlement has been welcomed in copyright circles, and could set a standard for settlement in other pending cases where pirated material has been downloaded for AI training by companies such as META and OpenAI, it doesn’t settle the overriding question of whether the unauthorized use of non-pirated materials for AI training is legal. With the settlement, the Anthropic case is closed, including with respect to the fair use findings. There will be no appeal, another benefit for Anthropic. However, there are still a number of other cases working their way through the US courts, so the question of whether unauthorized use of copyrighted content for AI training constitutes fair use is far from settled.

The Anthropic settlement, especially its size, has caught people’s attention. It may result in AI developers deciding it is better to resort to licensing solutions to access content rather than risking the uncertain results of litigation. On the other hand, payments like this could be one-offs, a speed bump for deep pocketed AI companies who will continue to trample on the rights of creators if they can get away with it. In the Anthropic case, while the company must destroy its pirated database, it is not required to “unlearn” the pirated content that it ingested. Moreover, even if this case leads to more payments to authors, which would be welcome, there are still many copyright-related conundra to be resolved. It should not be necessary to have to constantly resort to litigation to assert creator’s rights given that, as the Anthropic case shows, only a very limited number of rightsholders benefit from specific cases. Broad licensing solutions are required. This would also help address the problem of AI platforms producing outputs that bear close resemblance to, or compete with, the content on which they have been trained.

While Bartz v Anthropic is a decision that applies only to the US, and only to this one very specific circumstance, it will be studied closely elsewhere in countries that do not follow the unpredictable US process of determining fair use, for example in fair dealing countries like the UK, Canada, Australia, New Zealand and elsewhere, and in EU countries. In Canada, the unauthorized use of copyrighted works for training commercial AI models is a live issue. With the possible exception of research, unauthorized use such as that undertaken by Anthropic is unlikely to fall into any of the fair dealing categories (in Canada, they are education, research, private study, criticism, review, news reporting, parody and satire) nor is there a Text and Data Mining (TDM) exception in Canadian law. As Canada and other countries come to grips with the copyright/AI training dilemma, the principle of how content is accessed will surely be an important principle. Just as fair use (if indeed AI training is determined to be fair use) does not justify piracy in the US, licit access is required in Canada to exercise fair dealing user rights, including where TPM’s (technological protection measures, aka digital locks) are in place to protect that content.

Judge Alsup’s decision upholds the important principle that the end (if legal) does not justify the means (if illegal). This is a key takeaway from the Anthropic case, imperfect as the outcomes of that case were. Meanwhile the legal process of determining how and on what terms AI developers should have access to copyrighted content to train their algorithms continues.

© Hugh Stephens, 2025. All Rights Reserved.

Hold the Champagne: The Two AI Training/Copyright Decisions Released in the US Last Week Were a Mixed Bag for AI Developers

Illustration of a champagne bottle being popped, enclosed in a red circle with a slash indicating 'no champagne'.

Image: Shutterstock.com

Last week I wrote about the questionable ethics of META’s use of pirated content to train its AI model, Llama, pointing out the ethical issues involved with META’s admitted use of pirated online libraries, such as LibGen (Library Genesis), to feed content to Llama for training purposes. This is quite apart from whatever legal issues that may arise from the widespread practice of ingesting copyrighted content for AI training by making an unauthorized copy from any source (such as a legitimate library, through purchase of a single copy of a work, or from publicly available internet sources, for example) not to mention the additional element of taking that content from pirate sources. The day after that blog was posted the first of what will be a series of legal decisions in the US regarding cases brought by authors and copyright holders against AI companies was issued, followed by another a day later. Both cases were heard in the Northern District of California, in the same San Franciso court house, but handled by different judges.

I updated last week’s blog to make reference to the Bartz v Anthropic case (hereafter “Anthropic”), but given the importance of that decision, combined with a decision released in another California court room a day later (Kadrey et al v META), these cases merit further exploration–especially since they were widely trumpeted by AI advocates as opening the door to unauthorized use of copyrighted content for AI training on the basis of “fair use”.

Fair use is the complex legal doctrine used in the US to determine exceptions to copyright protection. US readers are well aware of the intricacies and idiosyncrasies of fair use but for those not overly familiar with how it works, here is a short summation I drew from a blog post on fair use vs fair dealing that I wrote a few years ago.

In the US context, fair use is an affirmative defence against copyright infringement and is determined by the courts on a case by case basis, judged against several fairness factors (purpose and character of the use, the nature of the work copied, the amount and substantiality of the amount of the work used, and the effect of the use on the value of the original work)… Fair use is not defined by law. Some examples are given in US law of areas where the use is likely to be fair (criticism, comment, news reporting, teaching, scholarship, research) but these are illustrative and not exhaustive. In short, it is the courts that decide. This in turn can lead to extensive litigation as to what is and is not fair use, and it is worth noting that different judicial circuits in the US have at times come up with conflicting interpretations.

Or, for that matter, two different judges in the same circuit delivering decisions just days apart on similar issues but with some significantly different outcomes, as we saw last week (although in these cases both found fair use by AI developers with regard to the copyrighted works at issue).

On the Anthropic case, US District Judge William Alsup ruled, on summary judgement, that the use of copyrighted works for AI training, even though done without authorization, is highly transformative and does not substitute for the original work (“The technology at issue was among the most transformative many of us will see in our lifetimes”). It thus qualifies, according to Alsup, as fair use because the transformative nature of the use overrides or swallows the three other fair use factors, including the important fourth factor (effect of the use on the value of the work). He notes there was no allegation that the output of Anthropic’s model, known as “Claude”, produced content infringing the works of the plaintiffs. However, Judge Alsup then went on to consider the legality of Anthropic’s actions to download more than 7 million works from pirate libraries (such as Books3, Library Genesis and the Pirate Library Mirror) to constitute its reference library, which it initially planned to use for AI training. He concluded this was a prima facie case of copyright infringement, whether Anthropic intended to use some or all of the pirated works to train Claude or not. (“Anthropic seems to believe that because some of the works it copied were sometimes used in training LLMs (Large Language Models), Anthropic was entitled to take for free all the works in the world and keep them forever with no further accounting “.) Damages, to be decided at trial, could be substantial. Alsop did not, however, rule explicitly on whether or not the use of pirated works for AI training purposes could be a fair use.

Because of the controversial nature of Alsup’s findings on transformation and fair use, there is no question that this case will be appealed. While there have been many criticisms of the fair use elements of Alsup’s ruling, a particularly clear and trenchant analysis was put forth by Kevin Madigan of the Copyright Alliance (Fair Use Decision Fumbles Training Analysis but Sends Clear Piracy Message).

The second case last week to reach the decision stage was Kadrey et al v META. In this case District Judge Vince Chhabria found that META’s use of the works of the plaintiffs, thirteen noted fiction writers, to train its AI model (“Llama”) was also fair use. Chhabria, like Alsup, found that META’s use was transformative on the first fairness factor dealing with the purpose and character of the use (“There is no serious question that Meta’s use of the plaintiffs’ books had a “further purpose” and “different character” than the books—that it was highly transformative.”) but unlike Alsup, Chhabria put much greater emphasis on market harm, (the fourth fairness factor dealing with the effect of use on the value of the work) suggesting that it could be determinative. Unfortunately for the plaintiffs, however, Chhabria considered their arguments with respect to market harm to be unconvincing. There was no evidence that Llama’s output reproduced their works in any substantial way or substituted for the specific works at play nor was there evidence, according to the judge, that the unauthorized copying deprived the authors of licensing opportunities.

Chhabria suggested that a far more cogent argument would have been that use (unauthorized reproduction) of copyrighted books to train a Large Language Model might harm the market for those works by enabling the rapid generation of countless similar works that compete with the originals, even if the works themselves are not infringing. In other words, causing indirect substitution for the works rather than direct substitution. This is the theory of “market dilution”, which was also put forward speculatively by the US Copyright Office in its recent Pre-Publication Report on AI and copyright. Since this wasn’t presented as an argument, Chhabria could not rule on it but in effect he is inviting future litigants to pursue this line of argument, noting that his decision on fair use relates only to the works of the thirteen authors who brought the case.

The clearest way to illustrate his line of reasoning is to quote directly,

“In cases involving uses like Meta’s, it seems like the plaintiffs will often win, at least where those cases have better-developed records on the market effects of the defendant’s use. No matter how transformative LLM training may be, it’s hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works that could significantly harm the market for those books”.

This editorializing, known in legal circles as obiter dicta, is not binding nor precedential, yet will undoubtedly have some influence given Chhabria’s stature. It is likely that one of these days Judge Chhabria will have the opportunity to put these theories into practice when ruling on a similar case, but one where the plaintiffs have made a better case for market harm. He has provided them a roadmap.

While these two cases have fired the first shots in what is going to be a lengthy war, they do not seem to be dispositive. There are enough caveats and nuances to be able to conclude that the AI developers are far from being out of the woods. Both “victories” have a sting in their tail, especially Judge Alsup’s finding on piracy. Neither copyright advocates nor AI developers should be breaking out the champagne just yet. But whichever way it turns out, there will be some sure winners; the lawyers for each side.

© Hugh Stephens, 2025.