To build models, AI companies acquire terabytes of data from internet scraping, pirate libraries, user-generated materials, scanning purchased books, or third-party data licensing. Image: Thomas Gaulkin/Reihaneh Golpayegani / Tornado / Licenced by CC-BY 4.0
By Matt Blaszczyk
July 14, 2026
“We’re going to live in a world where all of the capital and power concentrates into the hands of five or six companies. That’s not a free society. That’s totalitarianism.” This is how Joseph Gordon-Levitt described the likely consequences of artificial intelligence as it’s currently being integrated into society. The actor and campaigner is not calling for the destruction of the technology. In fact, he finds AI “quite exciting.” What Gordon-Levitt opposes is “theft:” AI companies’ failure to pay rightsholders for the use of their works in their models’ training data. Whether the law agrees is another question.
British legal philosopher and positivist H. L. A. Hart explained that law is a normative system analytically separate from morality or the norms of the artistic or academic worlds. This is why something can be immoral but legal or vice versa.
For example, people may morally condemn a student who parrots information from an online source without citation; those in higher education might call this plagiarism. But, as far as the law is concerned, the student can parrot away. Copyright does not protect information or ideas; it only protects original expression. United States copyright law does not generally even mandate crediting one’s sources.
Similarly, copyright law protects works of authorship, such as emails or selfies, which the artistic world would never think to hold much value. But sometimes the law refuses to protect works that artists find valuable, including certain forms of experimental art or dance. For this reason, it’s worth keeping things in perspective: whether one is talking about the law as it is, as it ought to be, or about what’s morally good or aesthetically beautiful. If talking about the law, then one needs to look at the world through its eyes, and apply legal rules and principles, asking if the law recognizes a right in a particular work or an infringement in how a work is used. With respect to copyright and AI, this can be frustratingly difficult.
Currently, more than 100 lawsuits are making their way through United States courts, testing whether AI companies may use copyrighted works to train their models without permission or payment to copyright owners. The first two decisions, Bartz v. Anthropic and Kadrey v. Meta, both from a single federal district court in California, suggest that the training itself is likely fair use, though the two judges split over the use of books pirated from shadow or unauthorized libraries and over whether copyright should evolve to shield authors from competing with machines trained on their own works. Neither ruling binds other courts, and the harder questions—who and what exactly infringes, and what a courtroom victory would deliver for creators—remain open. Litigation will not yield bright-line answers anytime soon, and courts are unlikely to deem AI training simply “theft.” This uncertainty serves the biggest players on both the content and AI sides, while most individual creators cannot count on copyright to ensure they are compensated for AI training that aims to replace them. One way out of the legal quagmire would lie in political action: the introduction of a licensing regime that could ensure compensation for actual creators, an opt-out mechanism for those unwilling to settle for a statutory fee, and a broader democratic debate concerning the power imbalances that both technology and media impose on the people.
Copyright basics. The statute protects original works, which exhibit a minimal level of authorial creativity. Such works are different from the physical copies in which they subsist. For this reason, a novelist doesn’t lose copyright when they sell a single paperback but retains a property right that subsists in all copies of the book. In fact, copyright owners (initially authors or their employers) have a couple of distinct rights, the most important among them being the “right of reproduction,” which allows a novelist to control who copies their work and to sue if another book contains literally the same material or one that is very close to the original.
Courts test this latter type of copying through the requirement of “substantial similarity,” which determines how much copying is too much. Different U.S. courts developed inconsistent formulations of this test, each giving rather limited guidance. One approach asks if an ordinary observer would recognize that the work in question copies the original. Another asks to compare the “total concept and feel” of both works. Yet another, often favored in the software context, makes it explicit that ideas and unoriginal elements embodied in the works should be filtered out, and that the comparison should be made only between the protected material of both works. For this reason, not every debonair spy infringes on Casino Royale and not every vampire novel infringes on Twilight. Rather, the law encourages authors to create competing original works and to “build freely upon the ideas and information conveyed by a work.” This is how the Supreme Court has understood the Intellectual Property Clause of the Constitution: a limited grant of rights to authors in original works to incentivize creativity.
Fair use. That said, a lot could potentially be infringing. What is necessary to show is that someone factually copied the original work and took too much in the eyes of the jury or the judge. After all, Mission Impossible’s Ethan Hunt is pretty close to the James Bond, and similar worries apply to parodies of the Bella Swan saga of Twilight. (By way of background, in the United States, copyright can extend not just to works, but also sufficiently distinctive characters featured in works.) Even more so, much of journalism, scholarship, critique, or commentary reproduces original works and could be actionable. This is where the famous doctrine of “fair use” comes into play: Some uses of works, even if substantially similar, are taken by law to be fair and non-infringing—requiring neither compensation nor consent—based on several factors. The statute names four of them, though other considerations can be taken into account:
In practice, the first and fourth factors have come to dominate judicial analyses. When considering the purpose of the use, judges often consider how “transformative” the use is; that is, whether the new use gives the original work a new purpose, character, or meaning, rather than just replacing the original product. The most canonical examples of transformative works are parodies, which turn the message of the work on its head, as well as digital databases, such as Google Books and HathiTrust, that copy books to allow online word searches.
But not all uses are equally transformative, and the analysis is balanced out by consideration of market effects under the fourth factor, with the Supreme Court recently signaling special attention to commercial substitution. Here, the canonical examples of works not being fair use are when they simply replace the original, when a second printer sells unauthorized copies of a book at a lower price, or when movies can be downloaded from piracy sites for free.
These are relatively easy examples, and courts would not struggle with them much; most of the controversies would be factual, procedural, or evidentiary, stemming from legal grunt work, rather than the underlying vagueness of rules. But the most interesting cases are hard: the facts are difficult, there is no easy precedent to rely on, and there is no one obvious way to apply even the clearest of rules.
What counts as a copy, when is a copy substantially similar, when is a use fair, and who made a copy and with what intent, are all questions that invite legal interpretation and argument (which is also why appellate lawyers have jobs). But when approaching facts as complex and unprecedented as those in generative AI training, judges face difficult dilemmas: They must balance established legal principles with policy and equity. Judges cannot easily reach for answers in the statute, its preparatory materials, interpreting precedent, and certainly not in the dozen words found in the IP Clause of the Constitution; they need to consider arguments, technical expertise, and not be swayed by rhetoric. And finally, they need to decide. As positivists taught long ago, “in this area men cannot live by deduction alone.”
Training and copyright cases. To build a model, AI companies need to acquire terabytes of data from internet scraping, pirate libraries, user-generated materials, scanning purchased books, or third-party data licensing. The data is then compiled and filtered, first in “pre-training,” which prioritizes large quantities of data, and then more selectively in “fine-tuning,” when the model is optimized with higher-quality data. Later, these models power consumer-facing AI applications that allow users to generate outputs based on prompts. Several steps in the process involve making copies, sometimes by different actors; the overarching question is whether such copies infringe authors’ rights.
Legal debates in this area have focused on a couple of related questions: Are copies of copyrighted works transient and never perceived by a human being? Does the AI simply learn general patterns from the data—ideas or information, which are unprotectable—or does it memorize the expressive aspects of works it was trained on, with the copies somehow remaining in the AI? These questions are, in part, technical and in part normative, and are often answered with metaphors. For example, President Trump said: “When a person reads a book or an article, you’ve gained…knowledge. That does not mean that you’re violating copyright laws…Of course, you can’t copy or plagiarize an article, but if you read an article and learn from it, we have to allow AI to use that pool of knowledge without going through the complexity of contract negotiations.”
This view aligns with a theory supported by several legal scholars that training generative AI should be allowed, as it typically does not convey the author’s original expression to new audiences. By design, they submit, AI does not “memorize” the works it was trained on; it only learns from the patterns in those works.
Exceptionally, when some works appear many times in the dataset or are highly abstract—as copyrighted characters, such as Snoopy, Superman, or Mickey Mouse, tend to be—the model can reproduce them, which would invite liability, while training generally would not. Although these metaphors do not describe the technical process perfectly, their basic point is that typically, AI allows users to create new works that do not resemble anything in the dataset, and AI is not an infringement machine; instead, the underlying training should be deemed fair.
The first two district court cases on generative AI, Kadrey v. Meta and Bartz v. Anthropic, offer some support for this view. In Bartz, the court held that Anthropic’s training of the AI model behind Claude was fair use, opining that such training was “spectacularly” transformative, and that none of the outputs were substantially similar to the ingested works. The court was not willing to admit that the authors behind the class action were entitled to a licensing market for large language model training, either. However, the court found that training the AI on pirated books—those downloaded from Library Genesis, rather than scanned from physical copies—was unfair. It is the piracy, not the training, which the court condemned, leading to a $1.5 billion settlement, the biggest in copyright history.
The other case, Kadrey, is most interesting not because of its holding but its obiter dictum—the non-binding remarks given by Judge Chhabria. The court held that the training was fair use since, on the facts, no matter how hard the plaintiffs tried, they could not retrieve more than short quotations of their works from Meta’s AI. And so, they could not show substantial similarity under settled doctrine, as explained above.
Unlike the Bartz court, the Kadrey court did not think “piracy,” that is, downloading from Library Genesis, was unfair. However, Judge Chhabria regretted that the plaintiffs had not argued a different, bolder theory of harm, recently supported by the U.S. Copyright Office. Consider the hypothetical: What if the market for biographies is flooded with AI-generated books, rendering human biography writers unable to compete and reducing the value of their works? According to the dilution theory, this could infringe copyright law even if plaintiffs cannot show substantial similarity between their works and AI’s outputs. Finally, Judge Chhabria opined that, in his view, AI training generally is unfair and should not go uncompensated; Kadrey, however, was regrettably neither the time nor the place for such a holding.
Legal takeaways. The biggest takeaway from these cases is legal uncertainty. Both district court decisions have been decided on narrow facts. On the one hand, they seem to indicate that training, in and of itself, is lawful. On the other hand, the two judges differed in their approach to shadow libraries and their general views on AI training. In Bartz, Judge Alsup opined that people cannot expect copyright to save them from competing with AI-generated outputs if those are not substantially similar, while in Kadrey, Judge Chhabria wished to endorse a theory that would do just that. Therefore, it’s still unclear whether AI training is legal. This may even be a wrong, simplistic question. Instead of having one bright-line rule, courts will analyze cases in a fact-intensive way, avoiding broad pronouncements or subscribing to metaphors that AI “learns” or “steals,” depending on whom you ask.
One factor that may influence fair use is whether an AI service provider implements safety measures to minimize users’ generation of substantially similar outputs; for example, users may be prohibited from requesting a Superman short story. Another can be whether, in AI training, the company ignored technical measures that are akin to opt-outs: paywalls, robots.txt-encoded opt-outs (which tell web crawlers what they can access on a site), and other access controls.
Most importantly, experts are awaiting cases in which plaintiffs can demonstrate that the AI generates infringing output—for example, releases outputs containing Superman when prompted by users or, as The New York Times alleges, reproduces large chunks of text the AI was trained on. When those cases are decided, it will become clear whether judges accept the claim that it is the AI companies, not users themselves, who are infringing.
After all, there is something artificial about how litigators have been framing what and who is infringing: They sue AI companies for both the training of the AI model and for producing infringing outputs, likely because the training, uncoupled from outputs, would be fair use, while outputs uncoupled from training would be a direct responsibility of users, rather than companies. The latter could only be held responsible for contributing to the infringement, which is difficult to prove.
The fate of the dilution theory, which gets rid of substantial similarity altogether, is unclear. Certainly, it would be a step into uncharted territory. It goes against much of the copyright doctrine and theory and, arguably, the constitutional goals and limitations of copyright law, which limit authors’ rights. However, this may not lead to its defeat in courts or Congress. After all, the law can be changed, and the Constitution reinterpreted in accordance with the needs of the moment—realistically, that is what appellate courts, and the Supreme Court especially, quite simply do. The rest is policy.
AI training’s political economy. The dilution theory and the words of Gordon-Levitt quoted above have an intuitive appeal: People feel that there is something unfair about AI companies training on people’s works without consent or compensation. This is likely because consent matters on its own, or because artists are starving while AI companies are well-off, and seem deeply tied to the all-powerful Big Tech. Some would even say that while authors create, AI only releases slop. I examine these worries at length in a forthcoming article, AI and the Culture Industry, but it may be helpful to break them down into a couple of paragraphs.
First, the current legal uncertainty benefits the most powerful stakeholders on both the AI and content sides: not the starving artists, but big movie studios, and not small tech startups, but the giants of the AI era. Only the biggest market players on both sides can afford the costs of litigation and potential liability, or settle and strike licensing deals, such as News Corp’s reported $250 million agreement with OpenAI. This effectively pushes out both AI start-ups with little capital and individual rightsholders who lack the practical ability to play the litigation and bargaining games. Big studios can even combine the licensing of their copyrights with investment in AI firms—such as Disney’s $1 billion deal with OpenAI, which ultimately collapsed after it turned out that OpenAI’s video generator, Sora, was neither popular nor profitable. What this means is that the concentration of power that pundits warn about may be the very result of their advocacy, intentionally or not.
Second, creative employees who work for corporations or those who assign their rights cannot sue in the first place: they are invisible to copyright law and are unlikely to see a dime from deals signed by Disney or big publishers, especially if they are not celebrities or members of a strong union. Empirical research continues to show that copyright distributions are painfully unequal, and most creators receive only 12 to 22 percent of their income from copyright. Sometimes, corporations such as social media platforms even sign deals over access to their user-generated training data—like posts on Reddit—even though they do not own the copyrights. While these themes are highlighted in copyright scholarship and occasional open letters that ask publishers to “stop the misuse” of creators’ rights in AI deals, they go unnoticed in the speeches of celebrities. This does not mean that licensing will not benefit creative workers; rather, it is not guaranteed by copyright law, though it may result from union negotiations.
Third, licensing is unlikely to prevent creative work displacement or automation. In fact, many licensing agreements between publishers and AI companies have terms concerning incorporating AI by the publisher to automate creative work. For example, when the Associated Press signed its deal with OpenAI, it announced it would build on “nearly a decade to use automation to make its journalism more effective [and] help local news outlets integrate the technology into their operations.” Since then, the Associated Press has reduced staff while the company’s vice president proclaimed, “We’re not a newspaper company, and we haven’t been for quite some time.” And they are not alone, with further examples from Axel Springer, Condé Nast, Vox, and others.
There is also the end-game question of both goals and remedies. What do people want to achieve in a particular lawsuit and in the whole ecosystem? Is it a vindication of consent or achieving compensation? This matters because courts could award damages, prohibit AI firms from using datasets containing infringing works, or even order their destruction. Currently, litigators have requested all three types of remedies in different lawsuits. It remains to be seen how many plaintiffs with strong claims will litigate to the end rather than settle, and whether the courts will be willing to entertain the destruction of datasets or models.
On the one hand, the destruction of datasets that power models widely used by Americans seems very drastic. Many would argue it’s disproportionate, especially given the uncertainty over fair use in this context, and the importance of the technology to the economy. On the other hand, monetary liability can be insufficient to deter the biggest market players who can pay for their violations. It also leads to the same asymmetries of power described above; it pushes out start-ups and may not compensate those plaintiffs who don’t have the means of litigating. Perhaps large class actions, such as the one in Bartz, will vindicate ordinary creatives, but will the sum of money be enough if their work effectively trains their replacement?
Finally, this is not the first time people’s data has been used by AI companies; similar battles have been fought in privacy law. There, class actions have enriched mostly lawyers, while consent is all too easily given. To test it quickly: how many terms of service, including click-wrap and shrink-wrap agreements, have you signed lately, and how many included terms on AI training? I could easily guess that if the fate of AI companies were genuinely imperiled, Congress would pass legislation to save them from liability—it does not take a great cynic to foresee that legal actions will not shake up the foundations of the economic system, regardless of whether they should.
This is especially clear at a time when the government places significant weight on developing AI domestically rather than abroad for economic, political, and even national security reasons. The most recent proposal by Sen. Bernie Sanders I-Vt and reported talks between OpenAI and the Trump administration would involve the government taking an equity stake in the biggest AI companies. Even if these plans fail, it would be difficult to imagine the United States ceding the terrain to other jurisdictions, and, to the extent that international competition is concerned, copyright doctrine may prove the least important factor—for better or worse, the AI race seems to be ongoing.
Where does this leave us? AI’s impact on the creative economy is complicated. It cannot be wished away. This future should not be decided behind closed doors of C-suites in technological and media conglomerates but through a broad democratic debate.
What’s needed is a vision for the future where AI technology is used for human flourishing, including the creative and scientific potential it opens, while minimizing technological harms, including creative displacement and a feeling of people’s work being re-appropriated by someone else. Copyright law, as it exists in the 50-year-old statute, does not offer appropriate remedies. The courts are unlikely to be the best designers of such a reform, either. Judges are tasked with applying the statute to the dispute before them. This does not mean that all the lawsuits against AI companies will be unsuccessful, nor that they should not take place. It is important, however, to manage expectations.
Looking ahead, perhaps a mandatory licensing regime could be developed to ensure compensation for creators at large and reduce some of the asymmetries in the status quo; it could even include an opt-out mechanism. Such proposals are currently being debated both domestically and in Europe. However, copyright must form part of a larger reform. The first step remains the one this article began with: distinguishing the law as it is from morals and from policy and refusing to let the rhetoric of celebrities or conglomerates blur the three.
Matt Blaszczyk
Mateusz “Matt” Blaszczyk is an assistant professor at the University of Georgia School of Law, where he teaches property, law and technology, … Read More
Mateusz “Matt” Blaszczyk is an assistant professor at the University of Georgia School of Law, where he teaches property, law and technology, … Read More
Generously supported by
the Future of Life Institute
The Bulletin elevates expert voices above the noise. But as an independent nonprofit organization, our operations depend on the support of readers like you. Help us continue to deliver quality journalism that holds leaders accountable. Your support of our work at any level is important. In return, we promise our coverage will be understandable, influential, vigilant, solution-oriented, and fair-minded. Together we can make a difference.
Keywords: AI, Anthropic, Bartz v. Anthropic, Google books, Kadrey v. Meta, Library Genesis, OpenAI, artificial intelligence, copyright, creative content, lawsuits, settlements, training data
Topics: Disruptive Technologies, Expert Analysis, The AI Power Trip
By Tik Root | Climate Change
By Matt Blaszczyk | Disruptive Technologies, Expert Analysis, The AI Power Trip
By S.M. Amadae, Tom Stevenson | Analysis, Nuclear Weapons
By Dan Drollette Jr | Artificial Intelligence, Chemical Weapons, Climate Change, Magazine, Nuclear Weapons, Satire, Special Topics
By Jack Kennedy | Analysis, Nuclear Weapons
By Amy Hawkins | Analysis, Special Topics
By Anila Yoganathan | Analysis, Nuclear Energy, Science Denial
By Sasan Karimi | Analysis, Nuclear Energy, Nuclear Weapons
Follow our Iran coverage
Copyright © 2026 Bulletin of the Atomic Scientists. All rights reserved. Registered 501(c)(3). EIN: 36-2136497 Terms of UsePrivacy Policy
PO Box 15461, Chicago, IL 60615 | 773.834.3779