No work co-signed by ChatGPT appears in the official list of winners and finalists for the 2025 Pulitzer Prize. The European Union responded with Article 50 of its AI regulation, which requires generative systems to label their output. But neither Brussels nor London has defined how to compensate authors whose texts, images, and musical scores were used to train these systems. Regulating the label on the finished product without touching the raw material amounts to treating the symptom while letting the cause flourish.

The essentials

  • The AI Act mandates labeling of AI-generated content but includes no mechanism to compensate creators whose works fed the training corpora.
  • The United Kingdom did not extend its text and data mining exception to commercial AI use. The current British exception remains reserved for non-commercial research; commercial training generally requires a license or other legal basis.
  • Midjourney is being sued in the United States for unauthorized use of artist corpora; the American Writers Guild of America struck over this same issue in 2023.
  • European and British decisions from 2026 to 2028 should lock in the legal framework for a generation of creators.
  • The tension between AI innovation and copyright remains unresolved: no jurisdiction has yet found a balance that preserves both.

The AI Act regulates the exit, not the entry

Article 50 of the European AI regulation, which took effect in 2024, requires suppliers of generative systems to indicate that their output is automated. Deepfakes must be labeled. Chatbots must identify themselves. The consumer must know.

This is real progress on transparency of use. The regulation requires documentation on the provenance, acquisition, and selection of data for the AI models in scope, without necessarily demanding detailed public disclosure work by work. It does not create automatic compensation, but it includes obligations to respect copyright and transparency aimed at rights holders.

This silence is a choice. During European Parliament deliberations between 2022 and 2024, several amendments proposed requiring developers to list the corpora used and create a right to compensation. They did not survive. What remains, according to the Culture Action Europe report published in May 2026, is a text that regulates the distribution of generated content without addressing the conditions of its production.

The gap is all the more visible because the text and data mining exemption, provided for in the 2019 European Copyright Directive, already authorizes research to mine protected corpora. The AI Act did not close this door for commercial training.

The United Kingdom chooses industry

London made a clearer choice. By late 2025, the United Kingdom had not adopted this extension; the option of a commercial exception with opt-out was subsequently rejected as the preferred option in March 2026. A developer can legally train a model on protected works without seeking permission or paying royalties, unless the rights holder has explicitly put in place a restriction system.

The reasoning is economic: the United Kingdom wants to remain competitive against the United States and China in the race for language models. Imposing authors’ oversight of each training corpus would slow industry. British consultation favored a commercial exception with opt-out, where the burden falls on the author to protect themselves rather than on the company to ask permission.

This logic reverses the burden of proof. In the United Kingdom, the current exception does not create a general commercial regime: commercial training generally requires a license or other legal basis. An opt-out could impose obligations on developers close to an opt-in in certain scenarios, according to the UK report, without this being established as a general effect.

The contrast with France’s position is striking. France defended, in European negotiations, a more protective approach to authors’ rights. But European harmonization has not been achieved, and the post-Brexit British framework applies to an AI market in full expansion.

Midjourney, WGA: what the lawsuits reveal

Midjourney, one of the main AI image generators, faces lawsuits from several American artists who argue that their styles and works were incorporated into training corpora without consent or compensation. The complaints target the initial collection, not the output produced by the tool.

The American writers’ union struck for nearly five months in 2023, notably to obtain guarantees against the use of their scripts to train AIs. The final agreement provided minimal contractual protections, without resolving the fundamental question: who owns the right to the data used to build these models.

These conflicts concern the ingestion of an entire corpus to extract statistical patterns, not the copying of an isolated work. American courts have already rendered decisions on the merits concerning certain uses of protected works to train AI systems, even if questions about generative models remain contentious. European systems have established a framework for training through text and data mining, but its practical application and limits remain contested.

The AI industry was built on a bet that training on public or semi-public corpora would be tolerated. That bet remains open.

The datasphere as a question of power

Stéphane Grumbach, researcher at INRIA and author of L’empire des algorithmes, analyzes how the digital reconfigures power relations and governance in a datasphere whose flows can escape control. This reading illuminates training corpora: many large models are trained on very large volumes of data, whose composition and time period vary by model. The corpora of some models may combine multiple types of texts and data. Creators and rights holders participated in consultations and certain content is subject to licenses.

This imbalance results from design choices by the industry and a definition of rights that remains contested. Control of corpora influences model orientation, and this influences global symbolic production. The wage bifurcation that AI produces in the world of work has its equivalent in creation: a share of productivity gains concentrates among platforms, while certain creators face reduced outlets.

A competing reading deserves to be stated honestly. Philippe Aghion, whose work on Schumpeterian growth emphasizes creative destruction as the engine of progress, would argue that restricting access to corpora slows a technology that produces net gains for the economy and, ultimately, for creators themselves. Less capable models would be less useful, less adopted, and would generate less demand for quality human content.

The argument is coherent, but it assumes gains are redistributed. There is no general dividend for artists, but certain uses of protected content can be compensated by license. Market segments like commercial illustration and stock music have experienced contractions without clear emergence of compensatory outlets.

The coming years will set the rules

Decisions made between 2026 and 2028 should shape the framework for a generation. Several texts are underway or in preparation.

The Commission has already clarified part of the practical articulation through guidelines and implementation tools, even if legal questions remain. The Council of Europe is working on a framework convention on AI that includes cultural provisions. The May 2026 Culture Action Europe report recommends an obligation to be transparent about training corpora and a collective remuneration mechanism, modeled on collective management of music rights.

This model exists and works. Rights collection societies like SACEM in France or PRS in the United Kingdom collect royalties on the use of protected music. SACEM redistributes them according to statutory formulas or, for certain mechanical rights, contractual ones. PRS redistributes according to its distribution policy, usage data, and work registration information.

Such a mechanism may require transparency about corpora. The regulation calls for a public summary of training content for general-purpose AI models, without going as far as individual identification of each work.

Corpus transparency is thus the prerequisite lock for any compensation system. Without a list of sources, no calculation basis. Without a calculation basis, no redistribution. Several EU member states are pushing for a disclosure obligation in the implementing acts of the AI Act. Industry resistance is strong: OpenAI, Google DeepMind, and Anthropic have all argued that revealing the composition of their corpora would expose them to unfair competition and an unmanageable volume of litigation.

AI and labor markets in emerging countries illustrate a similar dynamic: the gains of technology concentrate where capital is, while those who provide the material—here, the works—absorb the losses without compensation. This dynamic runs through the entire AI economy, far beyond cultural creation.

Cultural losses absent regulation

The prospective question is not abstract. If the current framework freezes—legal and uncompensated training, a labeling obligation without right to compensation—the effects on cultural creation would be structural.

Artists who live from standardized commissions, illustration, background music, advertising copy, may face substitution. Some segments of these markets have undergone transformations without the emergence of compensatory outlets being established. A sustained contraction of standardized creation revenues could affect the renewal of future training corpora if no compensation mechanism accompanies it. The model feeds on human creation; if human creation is no longer economically viable, the model loses its source.

A regulation that internalizes this cost, rather than bearing it solely by creators, would allow the industry to continue developing. A levy on the revenues of generative platforms, redistributed via collective management societies, would make it possible to fund human creation. The analogy with the blank media tax of the 1980s, which had financed rights holders during the transition to cassette, is imperfect, but it shows that such a mechanism has already worked.

The real difficulty is as much technical as political. A remuneration mechanism can rely on work-by-work attribution, but can also use collective licenses or other distribution formulas. Precise measurement of each work’s contribution remains technically difficult. Some researchers are working on data attribution methods, but they remain experimental.

The signal to watch in the coming months: the compliance of existing models with the AI Act’s transparency obligations, and final judgments in Midjourney and similar lawsuits in the United States. These two tipping points should indicate whether the window remains open for rebalancing, or whether the market has already organized itself around a de facto standard that the texts will merely ratify.

The capacity of states to build real capabilities largely determines their ability to regulate technology sectors that evolve faster than law. Europe has chosen to be the jurisdiction that regulates: it remains to be seen whether it will regulate the entire problem, or only the visible part.

Sources

  1. Culture Action Europe, AI and Cultural and Creative Sectors: Recommendations from the Digital AI Action Group, May 2026, cultureactioneurope.org
  2. Stéphane Grumbach, L’empire des algorithmes : Une géopolitique du contrôle à l’ère de l’anthropocène, Dunod, dunod.com
  3. Regulation (EU) on Artificial Intelligence (AI Act), Article 50, European Parliament and Council of the EU, 2024, Futurium EU, futurium.ec.europa.eu
  4. Directive (EU) 2019/790 on copyright and related rights in the digital single market, Articles 3 and 4 on the text-and-data mining exemption
  5. Writers Guild of America, strike settlement agreement, October 2023, WGA
  6. Policy Review, LLM blind spots, March 2025
  7. Council of Europe, work on the framework convention on artificial intelligence, 2024-2026