Abstract
The copyright battles regarding Generative AI are part of an historical trend whereby new technologies eventually facilitate new markets to deal with intellectual property rights. While that process occurs, given the uncertainty regarding both the value of works for Generative AI as well as their potential commercial impact for creators, it is best to opt for ex post mechanisms that take effect once uncertainty has been resolved.
Students sitting in clear lectures may decide not to worry about the textbook. Summaries of books can have the same impact. You are given so good a description of a documentary that there is no need to view it. Someone tells you the plot twist in a movie that you decided not to go to the theatre.
In all of these cases, someone has consumed copy-protected works and generated something that might deprive the creator of a sale. This is quite normal. If in each case, the consent of the creator was required, there would be real damage to education and public discourse.
AI providers argue that their models fall into the same category. Responding to the New York Times lawsuit, OpenAI likened training1 to the broad education people receive and called any one sector, news included, a ‘tiny slice’ of the training data. There is an obvious appeal to extending the rights of people to AI models. But it invites the question: why doesn’t the law treat these people as infringers? We need to answer that to understand whether it carries over to machines.
Generative AI raises the stakes on both sides. Because models extract statistical patterns from vast bodies of text, images and code, access to content is worth more to them than to any individual reader. And it is precisely because they are readily available and cheap that the scope for potential harm is greater too. On the flip-side, that scale also makes obtaining permission in advance far more challenging.
In the US, the question is being litigated as one of fair use. The EU has legislated instead. Article 4 of the Copyright in the Digital Single Market Directive2 allows works that are lawfully accessible to be mined for text and data unless rights holders have reserved their rights, and the AI Act3 obliges providers of general-purpose models to respect those reservations and to publish a ‘sufficiently detailed summary’ of their training content. Both debates are premised on the notion that, once permission is required, a workable market for licences will emerge to supply it.
The economics comes down to two questions: when can the parties bargain, and when does harm become visible? My Journal of Law and Economics article4, and my amicus brief in the OpenAI copyright litigation5, argue that for general training on large corpora the answers are ‘not before training’ and ‘only after deployment’. Policy should, therefore, let general training proceed, compensate creators whose protected expression is demonstrably harmed by what models produce, and support licensing where bargaining can work, as it can for the retrieval of current content.
I. Creation and use
Economically, the aim of copyright policy ought to be to encourage content to be created and used. Much of copyright law, quite sensibly, puts control in creators’ hands, even though this deters some use, because without that control much content would never be produced and there would be nothing to use.
Generative AI adds a second investment—the one AI providers make in building useful models—and the two are linked. Better reporting, books and software can make for better training material. But if AI outputs stand in for that content, creators’ earnings fall and so may their investment in producing it, which arguably could eventually lower the quality of AI too. The harm that matters for copyright is expressive substitution, which my paper calls ‘leakage’: a model lets someone obtain a substitute for a creator’s protected expression without buying it from the creator.
Why not let creators and AI providers bargain their way to a solution? If bargaining cost nothing, they could settle on terms reflecting what access is worth and any harm they anticipated, and, as Coase taught, the initial allocation of rights would affect who gets paid but not what is created or used. The trouble comes in applying that goal to a setting with millions of works, owners who are hard to identify, and a technology whose consequences no one has yet seen.
My paper separates two cases. In the first, a ‘small’ AI model, training uses a specific body of content whose owner can see that it has been used and strike a deal with the AI provider. Here the publishers’ position that copyright protection should govern is right: protection supports creators’ incentives and the quality of training data, and bargaining balances both against any harm to the creator. The second case, the ‘large’ model, describes most of the generative AI now in litigation, and here that position runs into trouble.
II. Bargaining over large models
Three features of large models rule out bargaining before training. The first is scale. One well-known dataset, Books3, held almost 200,000 titles6, including several of my own books. That tells me nothing about what my book contributed to the resulting model or whether the model has cost me a sale. Web-scale training involves far more documents and millions of rights holders. Finding those owners, checking their claims and negotiating with each could well cost more than their works are worth to the model, and if any one of them can block training, holdouts compound the problem.
The second is how little any one work contributes. Annie Liang and Jay Lu, in Creative Ownership in the Age of AI7, make this precise by asking whether an output could have been produced had a given work been absent from the training data. If works add to knowledge in small increments, dependence on any one work fades as the corpus grows; if contributions are heavy-tailed (i.e., comprising sizeable amounts), a distinctive work can remain essential. Supply and demand is explained in thousands of textbooks, is the first case; a breakthrough found nowhere else is the second. In other words, a collection can be valuable without any item in it being indispensable, so the value of the whole says little about what a single work is worth to an AI provider.
The third is uncertainty about harm. Until a model is trained and released, neither side can tell which works it will harm commercially. One book might teach a model a great deal without the model ever generating anything resembling its prose, while another work, perhaps one duplicated many times across the web, ends up memorised and regurgitated. Harm, if it arises, appears only after deployment, so contracting over it in advance would mean putting a price on an unknown distribution of outcomes across millions of works.
Thus, for general pre-training, policy has to begin from the fact that bargaining in advance is not feasible. A legal right to negotiate is of little use when the negotiations themselves cannot be conducted.
III. Full copyright or none
With bargaining in advance ruled out, my model sets two polar regimes against each other. Full copyright gives every creator a veto over training on their work. Because unauthorised use of a registered work exposes an AI provider to statutory damages of $750 to $30,000 per work—and up to $150,000 if the infringement is wilful8—whether or not the owner lost anything, AI providers in the model leave copy-protected content out. Creators keep their baseline incentives, but society gives up training uses that would have harmed no one, and AI becomes more expensive to build, less capable, or both.
Without liability, AI providers can draw on the whole corpus and owe creators nothing. Training benefits, but each creator whose expression an output replaces bears the loss, and the lower returns they can anticipate weaken investment in the very material that training needs.
The comparison between the two turns on how much content lowers the cost of training relative to the commercial harm creators can expect. When access is very valuable and harmful substitution is rare, welfare is higher without liability. This resolves the puzzle about human learning: an educated public is enormously valuable, while the chance that any one person’s learning damages a creator’s commercial interests is small. The point is not that suing every student who reads a textbook would cost too much; it is that such learning seldom does damage, so permitting it leaves creators’ incentives largely intact. AI pushes up both sides of the comparison, since access is worth more but models can also reproduce material on a large scale. That is why blanket immunity is harder to defend for machines than for people.
IV. Compensation after the fact
The difficulty with both extremes is that they treat every work alike before training, whereas harm, when it occurs, falls on particular works and becomes visible only afterwards. Calabresi and Melamed’s contrast between property rules, which require permission before use, and liability rules, which permit use and require compensation afterwards9, suggests a way out. Lemley and Shapiro’s analysis of probabilistic patents10, and my own analysis of remix rights11, point the same way: competing investments can be reconciled without agreement before every use.
The mechanism I propose, which I call an ‘ex post fair use-like mechanism’ (not the catchiest of names, I know, but I am an economist rather than a marketer), works as follows. AI providers may train on whatever content they wish without seeking permission work by work. Once a model is deployed, any creator whose commercial interests suffer harm above a threshold can recover lost profits: what they would have earned had the AI never existed, less what they actually earn. Note that, unlike ordinary copyright, the creator cannot stop their work being used in training, and damages restore lost profits rather than taking a statutory or punitive form for every training copy.
This mechanism can deliver higher welfare than either extreme, provided the compensation owed does not make AI provision unviable. This is because creators are, in effect, insured. Knowing that a serious loss from expressive substitution would be made good, they still have reason to invest in valuable work, and the lower the threshold and the fuller the compensation, the closer their incentives come to those under full protection. Compensation goes only to works that are actually harmed, so nobody has to bargain over millions of works on the chance that a handful will one day suffer. Put simply, creators are insured against the harm they fear, and AI providers are spared negotiations that cannot happen before training.
Because they pay for harmful outputs, AI providers also have reason to curb memorisation and regurgitation, through deduplicating training data, filtering outputs and similar safeguards. In Pigouvian terms, they internalise the harm, so any precaution that costs less than the damage it averts becomes worth taking. Neither extreme provides that incentive.
The mechanism is not without practical difficulties, of course. Can lost profits really be measured? Perhaps not with precision, but what matters is whether the most serious cases can be identified and compensated. Claims are also costly to bring, and smaller creators may struggle to obtain what they are owed; a threshold keeps administrative costs down, though some losses will then go unpaid.
This mechanism is not explicitly proposed in copyright law but is something that could be imposed legislatively. What the proposal offers is a way of assessing market harm within that inquiry.
V. The fourth factor
Fair use’s fourth factor directs courts to the effect of the use on the potential market for, or value of, the copyrighted work. Applied carefully, it asks what exactly was lost, whether the challenged use caused the loss, and whether a licence could realistically have been obtained. The four principles in my amicus brief address these questions.
First, copyright cares about substitution for protected expression, not about every lost sale. Facts and ideas lie outside copyright, so a chatbot that tells a reader a fact found in a book may cost the author a sale without causing harm copyright recognises. Authors Guild v. Google drew this line12 for a snippet giving a researcher a historical fact. In Twin Peaks13 and Nihon Keizai Shimbun14, by contrast, the defendants supplied close recountings of television episodes and abstracts that were largely translations of newspaper articles, so users received a replacement for the protected expression itself. That is the substitution the Supreme Court in Warhol called ‘copyright’s bête noire’15. A compensation rule should follow the same line, covering outputs that copy or closely track protected expression rather than every sale an AI displaces by having learned something useful.
Second, a claim of market dilution needs evidence of causation. The worry is that machine-generated works will become so plentiful that human creators earn less, even if no single output infringes. Campbell tells courts16 to consider what would happen if the defendant’s kind of use became widespread, so scale is relevant, but the counterfactual remains: how would the plaintiff have fared without the challenged use? Competition for a novelist’s readers can intensify whether or not the novel itself was ever used for training. A decline in earnings, on its own, does not reveal what share of it, if any, is due to the defendant.
The two June 2025 rulings from the Northern District of California show how much turns on this. In Kadrey v. Meta17, Judge Chhabria thought dilution could often decide the fourth factor, with news possibly among the most exposed markets, yet found the plaintiffs had produced nothing of substance to support it. In Bartz v. Anthropic18, Judge Alsup rejected the theory, likening the authors’ complaint to an objection that ‘training schoolchildren to write well would result in an explosion of competing works’, which, he wrote, is ‘not the kind of competitive or creative displacement that concerns the Copyright Act’. That is the analogy I began with, applied by a court. Whichever view prevails, the economic requirement is the same: a credible estimate of how the plaintiff’s market would have looked without the use.
Third, licensing markets must be judged as they stood at the time of the use, and for that use. Agreements with the Financial Times, News Corp, Reddit and the Associated Press19 are often cited as evidence that training could have been licensed all along. A functioning market can matter, as in American Geophysical Union v. Texaco20, where the Copyright Clearance Center already gave institutions a practical way to license photocopies of journal articles. But an agreement signed now says little about what could have been negotiated when earlier models were built. Recent work by Martin Beraja and Noam Yuchtman21 suggests why. They document how, for oil reservoirs, radio spectrum and consumer data, property rights over newly usable resources came to be settled after innovators had found valuable uses for them, often through litigation and regulation, and they treat the online content used to train AI as a current case of the same pattern. Extrapolating from history, today’s deals are partly a consequence of the training that preceded them.
Bill Graham Archives adds a warning22 against letting an owner’s wish to be paid for a use decide whether the use is fair; otherwise a licence would be needed because the owner would have sold one had a licence been needed. It also matters what today’s deals cover. Many package the right to train with archive access or live feeds, display rights, technology services or the settlement of legal exposure, and the AP’s 2023 agreement23 also gave the news agency the use of OpenAI’s technology and product expertise. A payment for such a bundle reveals little about whether, or at what cost, training rights could be cleared across millions of separate owners. These deals should be scrutinised as evidence, not treated as a reason to skip the scrutiny.
Fourth, pre-training on a general corpus should be kept separate from the retrieval of identifiable sources. A chatbot that answers a question about this morning’s news by fetching today’s articles, through retrieval-augmented generation or live grounding in web search, is not relying on what it learned in pre-training. The publisher and article are usually known, and a fresh report may be indispensable to the output in the sense Liang and Lu define, because nothing else available to the model contains it. A service that relies on the same publishers day after day can agree terms with them on delivery, attribution, display and payment. In the terms of my paper, retrieval has the economics of the small model, not the large one. Licensing makes good economic sense for retrieval services that draw on publishers repeatedly and stand in for a visit to their sites, and winning the argument over pre-training would not settle how every chatbot feature should be treated.
VI. Washington and Brussels
The AI Act requires providers to honour the reservations that the EU’s opt-out allows rights holders to make. What matters for policy is what follows when they do. Where clearing permissions is costly, a reservation is more likely to mean less training than a stream of royalties, and the largest AI providers, able to strike major deals or carry the cost of litigation, may come out ahead of new entrants. That is a risk to weigh rather than evidence that European training has stopped, but it follows from the cost of bargaining in advance.
The opt-out is better suited to a narrower role. Some creators object to any use of their work, whatever its commercial effect; that is a moral rights concern, and precisely the case ex post compensation cannot address. An opt-out protects them, and YouTube’s Content ID system24, which lets rights holders choose whether to block or monetise matching uploads, shows that such a choice can be administered at scale. The AI Act’s disclosure of training content makes it easier for creators to discover uses of their work and pursue claims. Neither, however, transforms millions of negotiations into something that can actually be carried out. Washington faces the same risk when plaintiffs invoke a potential licensing market without showing how far it extends, or seek statutory damages for training copies unrelated to any harm.
VII. Conclusion
We let people learn without anyone’s permission, not simply because pursuing them would be impractical but because their learning seldom harms the interest that copyright protects. AI raises the magnitudes on both sides of that comparison and makes clearing rights in advance far harder. The right response is to let general training proceed where no bargain could be struck, to compensate creators when harm to their expression can be shown, and to encourage licensing for identifiable uses where a bargain is within reach.
The hard part will be measurement. An ex post regime asks courts to determine which outputs substitute for which works and how much creators lost as a result. The 2025 decisions suggest this is where the evidence is currently thinnest, and so where litigants, courts and regulators should direct their effort. I am optimistic that it can be done, and that the outcome need not be the doom and gloom that content creators and AI providers each fear.