On 20 July 2026, a federal judge in California granted final approval to the largest settlement to date between authors and an artificial intelligence company: one and a half billion dollars, approximately three thousand per book utilised. What was resolved there is not what is typically reported. It was not resolved whether a machine can write literature. It was not resolved whether training a model on protected books is lawful, a matter which remains open and which the agreement itself expressly leaves undecided. What was resolved is more limited and far more effective: that pirated copies had been used.
A year earlier, another group of authors had lost a similar case against a different company. The judge conceded that the training constituted fair use in that specific file, and explained why: the plaintiffs had failed to prove market harm. The judge added a caveat that is rarely cited, yet remains the most compelling aspect of the entire matter: that in most cases, copying protected work to train systems of this nature will likely be unlawful if it prejudices the market for those works. They did not lose because the training is legitimate. They lost because they failed to prove the prejudice.
Furthermore, a third case completes the landscape. A major image agency sued an image generation company in the United Kingdom, and in the midst of the trial, it was compelled to abandon its central claim regarding training. This was not because the underlying argument lacked merit, but because it could not prove that the training had occurred within British territory. It secured only a minor victory, relating to watermarks reproduced in the generated images.
Let us consider these three elements together, and a pattern useful for this discussion emerges. Where a claim is factual and proven—these specific files, obtained in this specific manner—it advances. Where it is diffuse, it falters. In no instance, neither for nor against, has the question of whether the output is art or not determined anything.
It is worth recalling the events of February 2025, when a major auction house announced the first sale dedicated entirely to work created with artificial intelligence. More than three thousand people signed a letter requesting its cancellation; by the time the sale closed, the signatures numbered nearly six thousand five hundred. The letter did not halt anything: the auction proceeded and reached a total of 728,784 dollars, commissions included. Their argument was not that machines cannot create art. It was that these models, and the companies behind them, exploit human artists by using their work without permission or payment to build products that subsequently compete with them.
In this conversation, there are two distinct grievances that are often presented together and which do not carry the same weight. One asserts that the work of thousands of people was taken without request or payment. The other asserts that what is produced by these systems cannot be art, because art requires a human being behind it. The former is an assertion about what occurred; the latter, about what art is. In the eyes of the courts, the difference is not academic: one is proven with files, and the other has remained unresolved for a century.
I maintain, in my research, that what makes something function as art does not depend on the agent that produces it nor the medium with which it is produced. I cannot, therefore, conclude that generative production is excluded by definition, and I am aware that this sounds like a concession to those who are witnessing the decline of their commissions. Yet the same criterion allows me to state what the enthusiastic defence of these tools never does: that the vast majority of what they produce does not reorganise anything whatsoever. That something is possible does not mean it occurs, and it almost never occurs.
These two positions are perfectly compatible with agreeing entirely with the letter. One may believe that the medium determines nothing and simultaneously that there has been large-scale expropriation. There is, however, a conflict when the protest relies on metaphysics, because then a single counterexample suffices to make it appear refuted, while the labour issue remains obscured by a discussion of aesthetics.
There are also legal tools that almost no one uses. The 2019 European directive permits text and data mining, yes, but with one condition: that the rights holder has not expressly reserved their rights, and for content published on the internet, it mentions machine-readable means. That is to say, an artist can say no. The European Artificial Intelligence Act has required providers of general-purpose models since August 2025 to have a copyright compliance policy and to publish a sufficiently detailed summary of the content used for training, according to an official template. On paper, the framework is in place.
On the ground, it functions poorly. Opt-out systems exist—there is an initiative that registered over one billion works marked by their authors—but they never succeeded in having the industry adopt a common standard, and they have received serious criticism for the opacity of their own procedures. The technical problem is that it is one thing for a reservation of rights to be machine-readable and quite another for it to be machine-executable. Without that, saying no is a gesture without a recipient.
There is also a fact that dismantles the most frequently repeated excuse, which is that licensing at that scale is impossible. It is not: it is being done. A major image agency signed a six-year agreement in 2023 to supply training data. A news agency licensed its archive the same month. A large publishing group signed a multi-year global agreement the following year. The amounts are not always published, but the agreements exist and are announced by the parties themselves. Thus, the material is indeed paid for. It is paid to those with the scale to negotiate. An individual illustrator, however, is not.
That is the strong version of the grievance, and it is not the one occupying the debate. A survey by a British professional association, conducted at the beginning of 2024 with approximately eight hundred responses, found that twenty-six per cent of illustrators and thirty-six per cent of translators had already lost work due to generative AI, and that more than a third had experienced a decrease in income. It is a partisan survey, but an independent study on freelance labour markets points in the same direction with more modest figures: declines of two per cent in the number of contracts and five per cent in income in the most exposed occupations.
And yet another figure. In a survey of more than three hundred galleries conducted in February 2026, sixty-one per cent responded that none of their artists use artificial intelligence. Fifty-seven per cent responded that they themselves do use it, to draft their emails and communications. The technology has not entered the art world where it was feared. It has not entered through the work; it has entered through the labour that surrounds the work. Through gallery texts, press releases, descriptions, translations, correspondence. That is to say, through the jobs that were already the lowest paid in the sector.
In Spain, the matter is still in the positioning phase. The entity that manages the rights of text authors publicly stated in September 2024 that training with protected work without authorisation could infringe European regulations, and in December 2025 it published recommendations for its members on how to respond when a company proposes that they assign rights for training. The visual arts entity has published materials regarding the issue. Neither has initiated legal action, as far as is known.
I conclude with what I would grant without reservation to the signatories of that letter, and it is not insignificant. They are correct regarding the facts, they are correct regarding the scale, and they are correct that the lack of consent is not rectified by labelling it innovation. I would only ask that they do not shift the ground. A robust complaint does not need to maintain that art requires a human hand. It suffices to maintain that the work of thousands of human hands was utilised without permission to build what now competes with them. That is verifiable, it is recent, it has identifiable parties responsible, and, as we have just seen, it can be won. The other question can wait. It has been waiting for a century and has suffered no harm.
On the open conversation
This text intersects with two lines of my work: that what makes a work function as art does not depend on the agent or the medium, and that ethical responsibility is played out in infrastructures—training sets, labour conditions, regimes of visibility—rather than in the content of the works. Hence, I support the artists in their fundamental complaint while, conversely, disputing the terrain upon which they sometimes formulate it. If anyone wishes to intervene from the perspectives of professional illustration, translation, copyright, collective management, or commissioned cultural work, this notebook remains open, and in this instance, I am primarily interested in those who are witnessing the decline of their own.
Sources
Bartz v. Anthropic (District Court, Northern District of California). Final approval of the $1.5 billion settlement, 20 July 2026; Associated Press coverage, 21 July 2026.
Kadrey v. Meta Platforms (3:23-cv-03417, Northern District of California). Partial ruling on fair use, 25 June 2025.
Getty Images (US) Inc. et al. v. Stability AI Limited [2025] EWHC 2863 (Ch), High Court of Justice, 4 November 2025. Parallel action in the Northern District of California (3:25-cv-06891), filed on 14 August 2025.
Andersen v. Stability AI et al. (3:23-cv-00201, Northern District of California), filed on 13 January 2023; ongoing.
Thomson Reuters v. Ross Intelligence (District of Delaware), 11 February 2025.
Christie’s, Augmented Intelligence (online sale no. 23942, 20 February – 5 March 2025): total sale $728,784, including commissions; Christie’s results statement, 5 March 2025. Regarding the protest letter: The Art Newspaper, 10 February 2025 (3,576 signatures at time of publication) and 5 March 2025 (nearly 6,500 at the close of the sale).
Directive (EU) 2019/790, Article 4 (text and data mining and reservation of rights).
Regulation (EU) 2024/1689, Article 53 (copyright policy and summary of training content); European Commission template, 24 July 2025.
TechCrunch, 3 May 2023, regarding Spawning and HaveIBeenTrained.
Society of Authors, survey on artificial intelligence, 11 April 2024 (approximately 800 responses).
Brookings Institution, 8 July 2025, regarding effects on freelance labour markets.
Press releases from Shutterstock (11 July 2023), Associated Press (13 July 2023), and News Corp (22 May 2024) regarding licensing agreements with OpenAI.
CEDRO, public position of 11 September 2024 and recommendations of 16 December 2025. VEGAP, Ojo al dato. Inteligencia artificial y artes visuales, 21 June 2024.
Artsy, The Artsy AI Survey 2026, 18 March 2026 (more than 300 gallery professionals).
Esteban Ruiz, Juan A. Art as Structural Surplus. Zenodo, 2026. https://doi.org/10.5281/zenodo.20020706