back
Involuntary Credit
EdgeFiles

Involuntary Credit

About the Author

Sources

A proposed AI copyright bargain has creators finance their own compensation

Imagine a novelist whose income has fallen and who suspects an AI service is the reason. Proving it takes expert help, and the expert can't be paid with royalties the novelist believes should have been earned. A compensation system that overlooks this problem can promise to protect creators while asking them to finance that protection themselves.

That novelist is the person I kept thinking about while reading "Permission to Learn," an essay by economist Joshua Gans published in the Network Law Review on October 28. Gans, a professor at the University of Toronto, would let AI companies train their models on whatever content they want, work by work, without asking permission.

Once a model is deployed, creators whose losses exceed a threshold could recover their lost profits: what they would have earned had the AI never existed, minus what they actually earn.

They couldn't stop their work from being used for training, and statutory damages, the fixed per-work sums courts can award without proof of loss, would no longer apply to training copies.

Gans isn't a distant observer. Several of his own books were in Books3, a dataset of nearly 200,000 pirated books that a developer named Shawn Presser built in 2020 and that later became training material for AI language models, including Meta's LLaMA and BloombergGPT. He writes that their presence tells him nothing about what they contributed to a model or whether a model cost him a sale. He's right, and that uncertainty is the core of his argument. It's also the core of my objection, because someone has to pay to resolve it.

I find his effort to separate useful learning from damaging substitution persuasive. I'm less persuaded that a promise of eventual payment can replace control before use unless creators can afford to collect. Gans describes creators under his system as, in effect, insured. Insurance is only worth what it pays out, and this policy comes with an unspecified threshold, expensive claims, and a payer whose ability to pay is a condition of the theory.

Whoever finances the period between the lost income and the payment helps decide whether the right exists in practice.

Permission could arrive long before payment

Gans acknowledges that his compensation mechanism isn't part of copyright law and would have to be created by legislation. He also offers it as a way to assess market harm under fair use, the doctrine that allows some unlicensed uses of copyrighted work, and he has taken that argument to court. In the copyright litigation brought against OpenAI by The New York Times, the Authors Guild and others, he submitted an amicus brief, a filing in which an outside party advises the court. On October 30, OpenAI included it among excerpts from 20 filings backing fair use.

The order of events matters. A court could rule that training is fair use. Compensation would need Congress. If the first happens and the second doesn't, creators lose the training claim and never receive the new remedy. They'd still have ordinary infringement claims against outputs that copy their work, at the ordinary cost of bringing them.

I'm not suggesting Gans wants that outcome. His essay argues for payment and supports licensing where bargaining can work. But a proposal that depends on two halves should be judged on what happens if only one of them arrives.

Permission is expensive to organize

Gans's underlying research, published in the Journal of Law and Economics, examines a real coordination problem: the transaction costs of negotiating over a vast training corpus can prevent bargaining. Its finding that compensation after use can beat both full copyright and no copyright is a model result about expected social welfare, an economist's measure of overall benefit, balancing content creation against AI development.

Consider what permission for every item would demand. A developer would have to find out who controls the rights to each work and reach an agreement before including it. Each negotiation consumes resources even when the work's individual contribution is too small to justify a substantial payment. Repeated across a broad training set, the process of buying permission can cost more than the permission itself.

Creators should take that seriously. A right to negotiate produces little income if the prospective buyer walks away. Developers should take the other side just as seriously, because the work has to exist before it can become training material. A policy that makes model development cheaper by making professional creation less sustainable damages the supply it depends on.

There's also a middle ground worth defending. In its May 2025 report on generative AI training, still labeled pre-publication, the U.S. Copyright Office concluded that voluntary licensing may work in some contexts, while warning that those successes don't establish that it's feasible for every training need. It favored further market development over immediate government intervention. That's a reason to investigate which negotiations fail and why. A concentrated catalog with clear rights is a different problem from scattered works whose owners are hard to find. Collective licensing could reduce the number of transactions, although it would need the authority to represent those works and a credible way to distribute the revenue. Neither a few successful deals nor a few impossible negotiations settles the entire market.

Gans presents his mechanism as a way to handle the period while new licensing markets develop, and he says those markets should be judged as they stood at the time of the use. I'd hold him to that. A model trained in 2028 should be judged against the licensing market of 2028. Any exception justified by bargaining difficulty needs that continuing test written into it. Otherwise, temporary administrative difficulty becomes a permanent transfer of value.

The model assumes away the hard part

The paper's result rests on assumptions worth stating plainly. In the public working versions, negotiation over large models is assumed to be impossible. Each creator's harm is revealed to everyone after the model launches, and above the threshold the provider must pay full compensation. Nobody hires an expert, argues over access to records, or gives up on a valid claim. The paper notes that the costs of calculating damages afterward would have to be considered in practice.

Those costs are exactly where the essay runs into trouble. Gans concedes that claims are costly to bring and that smaller creators "may struggle to obtain what they are owed." He adds that a threshold keeps administrative costs down, although some losses would then go unpaid.

That concession affects more than fairness. The compensation is supposed to keep creators investing, because they expect to be made whole if serious harm occurs. The liability is supposed to push AI companies to stop models from reproducing the material they trained on, because harmful outputs carry a price. Both incentives depend on claims actually being paid. If a valid claim is routinely too expensive to pursue, creators can't count on the money and providers have less reason to avoid the harm.

Much of my work concerns the gap between what a governance document promises and what a system actually does once it's running. This is the same kind of gap. On paper, the injured creator is compensated. In practice, that depends on whether the creator can afford to prove the injury.

Lost profits require a world that never happened

The appeal of paying for actual damage is its apparent precision. But precision in the principle doesn't supply precision in the evidence.

Return to the novelist, and assume an AI service produces passages that closely follow the book's distinctive scenes. A reader can compare the texts. Establishing the income the novelist would otherwise have earned is much harder. Some users might have bought the book, while others would never have become customers. Sales might also have fallen because a promotion ended or a distributor stopped carrying it. An honest assessment has to separate those explanations.

Gans's own wording adds a complication. His formula compares actual earnings with what the creator would have earned had the AI never existed. Elsewhere in the essay, he says compensation should cover outputs that copy or closely track protected expression, meaning the specific wording and creative choices copyright covers rather than ideas and facts, and he asks how a plaintiff would have fared without the challenged use. Those are different comparisons. Removing all AI competition could overstate the loss caused by copying, while removing a single infringing output requires a much narrower estimate. The essay doesn't reconcile them, and a claimant would have to pay experts to argue about which one applies.

Limiting compensation to copied expression protects the public as well as the developer. U.S. copyright doesn't protect ideas, so a reader who gets an idea from an AI system and decides not to buy a book hasn't established infringement. Nor does a decline in a creative business's revenue identify the defendant responsible for it.

The same limit leaves out a harm many authors care about. Consider romance novelists losing sales to a flood of AI-generated titles, none of which copies their work. Lawyers call this market dilution. Courts weigh four factors when deciding fair use, and the fourth is the effect on the market for the original work. The Copyright Office said that analysis shouldn't ignore dilution, and two federal judges in California split on the question in June 2025. As Gans recounts, Judge Vince Chhabria thought in Kadrey v. Meta that dilution could often decide the fourth factor, although the plaintiffs offered little evidence for it. Judge William Alsup rejected the theory in Bartz v. Anthropic.

Gans demands evidence that the AI actually caused the loss, and his compensation rule follows copied expression. He may be right about the law. If he is, his system would pay for the harm that's easiest to identify and leave out the flood of AI-generated books the Authors Guild is now arguing about in court.

The remedy he would replace is the one that got authors paid

Statutory damages generally run from $750 to $30,000 per work, and up to $150,000 for willful infringement. Gans has a point that they can bear little relationship to actual harm. But they also serve a purpose. Section 504 of the Copyright Act lets an owner choose them instead of proving actual losses and the infringer's profits, which spares the owner from reconstructing what would have happened. Section 412 limits them, along with attorney's fees, to works registered on time, so today's system already excludes creators who never registered.

Bartz shows what that remedy can do. Authors sued Anthropic over its use of their books. Judge Alsup held that training was fair use but rejected fair use for the company's central library of pirated books, and he allowed the authors to sue as a class on behalf of the owners of registered books in the pirate libraries. That exposed Anthropic to statutory damages across hundreds of thousands of works. The company settled for $1.5 billion, about $3,000 per work before fees and costs, shared among each work's rights holders. The court approved the deal on July 20 and called it the largest copyright class action settlement in history.

The case concerned how the books were acquired, so it doesn't test Gans's proposal directly, but it does show who gets paid and why. A class action built on statutory damages produced a recovery no individual novelist could have financed. Registration rules helped decide who was included: a work qualified only if it was registered within five years of publication and either before Anthropic downloaded it or within three months of publication. The court tied those limits to the class's ability to pursue statutory damages. It also warned that a multi-billion-dollar verdict could itself raise constitutional problems, so this leverage has limits.

The essay offers lost profits without saying what would replace that leverage. It doesn't mention the infringer's profits, which the statute allows today. It also targets statutory damages for training copies without saying whether they'd survive for an infringing output. For the novelist whose scenes were copied, that unanswered question could matter more than anything else in the proposal.

A threshold chooses which creators can collect

Costly claims and unpaid losses wouldn't fall evenly. A substantial loss for a freelance illustrator can be too small to justify an expensive proceeding. A publisher holding a large catalog can spread investigative costs across more works. Even identical legal rights would leave them with different practical choices.

A threshold adds a second filter, and the essay doesn't say how it would be measured. It refers to a creator's harm in one sentence and to harmed works in another. If the threshold applies separately to each work, repeated small losses could escape compensation even when their combined effect threatens a livelihood. If losses can be combined across a catalog, the rule favors whoever has assembled one. An earlier version of the paper suggests harm could be measured as a share of a creator's revenue, which would help the freelancer. I'd start there.

The model also contains a less comfortable condition. The threshold can be set low only as long as the AI provider stays financially viable. In other words, the level of protection depends partly on the defendant's finances.

And if valid claims exceed what a provider can pay, creators have supplied more than training material. They've extended involuntary credit.

Requiring financial backing for claims would make the promise more credible while raising the cost of entry for developers. That trade deserves an explicit decision.

My test for a workable replacement begins with a credible claim that can be brought without betting a livelihood on it. Congress has already created a low-cost forum for small copyright disputes. The Copyright Claims Board charges $100 in filing fees and caps damages at $30,000, but a respondent can opt out of the proceeding, and an AI company facing thousands of claims would have every reason to do so.

A serious version of this proposal would close that exit, pay for shared investigation and give claimants proportionate access to relevant records. Developers also need protection against speculative demands for an entire business's confidential records, so the procedure has to distinguish a focused request from a fishing expedition. None of this follows automatically from labeling the remedy compensation.

Some creators will simply say no

Gans also recognizes that some creators object to any use of their work, whatever its commercial effect, and that compensation can't answer that objection. He would keep an opt-out for them and points to Content ID, the YouTube system that lets rights holders block or monetize matching uploads, as evidence that such a choice can be managed at scale.

I welcome the concession, but I don't see how it holds together. If any creator can opt out, the veto he argued against returns. If only creators with objections of conscience qualify, someone has to decide whose conscience counts.

Europe already operates the first version. The EU's copyright directive lets rights holders reserve their works from text and data mining, the automated analysis AI training relies on, and the AI Act requires providers of general-purpose models to respect those reservations and publish a summary of their training content.

Gans warns that this may lead to less training rather than more royalties and may favor the largest providers. It's a fair warning. I split my time between New York City and Berlin, and only one of those places has written an opt-out into law.

Retrieval brings the bargaining table back

Here I agree with Gans. Retrieval means a chatbot fetching a specific source to answer a question, which is different from the general training that happens before a model is released. A service fetching today's article knows the publisher and can agree on terms, so Gans treats this as a case where bargaining works, and licensing makes economic sense. Winning the argument over general training wouldn't settle it.

An agency building a research assistant for a client offers a useful example. The model's general training is one operation. Fetching a licensed market report to answer a question is another. The report can be identified, its supplier can be contacted, and the agency can decide whether access justifies the terms. Calling the whole product AI doesn't merge those transactions.

Agencies can be claimants and customers

For a creative business, the bargain cuts both ways. An agency might want inexpensive access to a capable model while objecting to another service reproducing work from its own portfolio. Those positions can coexist, but they require the agency to keep four questions apart: what permission covers training, what terms cover retrieved material, whether an output infringes someone else's work, and what copyright protects in the finished campaign.

On that last question, the Copyright Office's January 2025 analysis ties protection for AI outputs to a human determining enough of the expressive elements. Prompts alone aren't sufficient, while creative arrangement or modification can qualify. A client buying an exclusive campaign should ask what, exactly, is exclusive, and a contract between agency and client can't settle whether copyright protects the machine-generated parts.

Either system can favor the firms with money

Licensing before training can advantage developers who can negotiate large deals. Liability after deployment can advantage firms that can finance uncertain claims. Moving the expense doesn't guarantee a more competitive market.

Creators unable to fund an investigation would also have reason to join organizations that can do it for them, and those intermediaries' fees and terms would determine how much compensation reaches the person who made the work.

I take this proposal seriously because it forces a difficult question into view: who pays. Useful models and sustainable creative work are both worth investing in, and reducing the cost of obtaining permission can serve both. The bargain weakens if the savings depend on injured creators being unable to establish what they're owed.

Before asking creators to exchange control for compensation, the law needs a convincing answer to how they collect, and how they pay their bills until then.