A paper in Science says there is still no proof that AI dilutes the market for human work. The problem is that, in the real world, the one who has to prove the damage is the author — and the big tech companies just won another round.
Source: Market dilution, fair use, and generative AI — Science, 15 October 2026
There is a quiet inversion happening in the debate about artificial intelligence and copyright. For decades the copyright system started from a simple premise: whoever creates owns what they create, and whoever uses it needs a licence. Now the creator is being asked to prove that they are losing. And whoever uses it is only being asked to exist.
The paper Market dilution, fair use, and generative AI, published in Science on 15 October by the legal scholars Abhishek Nagaraj, Pamela Samuelson and Christopher Jon Sprigman, formalises that inversion with academic elegance. The three argue that the market dilution theory has no clear definition and that there is no evidence yet that AI output is actually harming the market for human work. While there is no proof, they say, fair use cannot be denied to training.
The reasoning is technically correct and politically revealing.
What market dilution proposes
The dilution theory was born from a court decision, not from a law. In June 2025, in the case Kadrey v. Meta, the judge ruled for Meta because the authors had not presented convincing proof of market harm. But he left a door open: he suggested that the ability of AI to generate content at volume could dilute the market for human work even when the output does not copy the protected expression of any specific work. A well-argued case, he said, could defeat the fair use defence.
It is an important distinction. Copyright law has always dealt with market substitution: someone copies the expression of a work and shifts the demand for it. Dilution proposes something different and broader. Without copying anything at all, the mass production of synthetic works would compete with human work as a class.
The three authors answer that creating competing works from the same ideas, facts or methods was never an infringement. And they are right on that specific point. What they do not address is the cumulative effect of a system that produces competition at industrial scale, with no marginal cost and no author.
The proof demanded from those already losing
Here is the knot. The paper asks for case-by-case proof of damage that is structural. It would be like demanding that each farmer prove, individually, that the change in rainfall harmed them, while refusing to measure the rain.
And there is a piece of data that points the other way. The global IFRRO survey, released at the end of October with nearly two thousand respondents from the text and image sector, shows that more than 90% report a direct negative impact on income, 96% consider the current copyright framework insufficient, and more than 75% believe AI will displace creative workers. That is not statistical proof of dilution in the academic sense. It is a picture of the market that the paper does not mention.
Berkeley Haas captured the weight of the question when quoting Nagaraj: much is at stake, because the theory could redraw the boundaries of fair use for those who build AI models. The summary adds that creators and users would feel the consequences too. It is the only passage in the material where creators appear as a party with an interest, rather than as an object of analysis.
Why now, and why like this
It is no coincidence that this argument is gaining academic weight at exactly this moment. Artificial intelligence has stopped being a laboratory promise and become economic infrastructure. It sustains entire sectors in the United States and sits at the centre of the technology dispute with China and the BRICS bloc. In that arrangement, the question of who pays for the training input stops being legal and becomes geopolitical.
When an asset becomes strategic, the burden of proof tends to migrate to whoever is in the way. That is what we are watching. The big tech companies got, once again, the support they needed — now not only in Congress and in the agencies, but in how justice is perceived and in academic output. And it is worth asking how those studies are designed, which cut-offs they adopt and who funds the building of the point of view that reaches the conclusions. This is not about doubting the authors, but about applying to them the same rigour demanded of anyone who measures damage done to others.
The argument nobody is making
This article is not a condemnation of artificial intelligence. That would be a misreading and a strategic mistake. Generative AI is a legitimate tool, and the author of this text uses it every day in his own production.
The point is different, and more serious than the fight over fair use: without payment and without the continuity of human production, AI loses the raw material it depends on. A model trained mostly on the output of other models does not learn the world. It recombines what has already been recombined, with loss at every generation. In a short time that turns into hallucination, repetition and an inability to produce new method.
Discouraging human production is not the artists’ problem. It is AI’s own problem. The diversity, the contradiction, the creative error and the lived experience that feed a model come from people who need to make a living from it. If that chain breaks, artificial intelligence gets stuck inside its own mirror.
What is up to us, now
The challenge of our time is not to choose between AI and copyright. It is to build a coexistence in which both keep existing, and not to let the libraries put their own interests above everything else.
That means three practical things. First, real licensing, with a price and a trail, instead of fair use as the default. Second, transparency about what went into the training, because without it there is no way to prove any damage at all. Third, a debate about the burden of proof that acknowledges the asymmetry: on one side, a company able to measure everything; on the other, an author who has to argue in court to prove they lost income.
As long as the discussion stays in the abstract, it will be won by whoever has more lawyers. The way forward is to bring data, method and a concrete case to the table, and to turn what is today an academic thesis into public policy.
