The machine and the muse: what French law actually does to artificial intelligence

Alison Leroy                                                                                                                 

UGGC Avocats, Paris                                      

a.leroy@uggc.com

Sylvie Savoie

UGGC Avocats, Paris and Beijing 

s.savoie@uggc.com

There is a foundational irony in the quarrel raging between human creators and learning machines: both feed on the same raw material. The novelist devours libraries, steeps themself in other people’s languages and distils accumulated influences into something that bears their mark: a voice, a gaze, what the law calls personality. The language model ingests billions of pages, encodes statistics and returns probabilities. The resemblance is close enough to unsettle. It is superficial enough that the law has something genuine to say.

As of July 2026, no generative artificial intelligence (AI) model provider has been condemned in France for copyright infringement. A Paris court in May 2025 ordered internet service providers (ISPs) to block an anonymous AI-powered site systematically pillaging press publishers[1]. But model providers themselves have not yet faced a French judgment on the merits. Proceedings against Meta are pending before the Paris court[2]. That fact is not a mark of judicial delay but is representative of a legal system that thinks before it strikes. Three developments loom: the full application of the European Union’s AI Act to high-risk systems in August 2026, with the obligations placed on general-purpose AI providers already biting since August 2025[3]; the Court of Justice of the European Union’s (CJEU) ruling on the Budapest referral; and the Darcos bill in France, adopted by the Senate in April 2026 and now pending before the National Assembly[4], which would create a presumption of use in favour of rightsholders. What follows traces not the absence of law, but its paradox: the rights are coherent, the creditors are real and the machine was built, in structure if not in intent, to elude them both.

The elusive safe harbour: how a mining exception became a shield for giants

Everything begins with an exception. When OpenAI, Meta, Google or Mistral absorb billions of tokens to train their models, they invoke a clause tucked into France’s Intellectual Property (IP) Code: the text-and-data mining exemption, which permits the reproduction of lawfully accessible works unless the rightsholder objects via machine-readable means[5]. Silence equals consent. The clause is the fulcrum on which the entire AI and copyright debate rests, a lever made of balsa: elegant in theory, fragile under the weight now placed on it. The exemption was drafted with academic research in mind, not the industrial ingestion of centuries of cultural inheritance.

Practitioners identify three structural fault lines. The first is philosophical: does statistical training constitute extracting information within the statute’s meaning or does it merely encode parameters that carry no trace? No French court has answered that question. The second is material: the statute requires the destruction of copies after mining, yet training data dissolves into numerical weights impregnated into the model irreversibly, the way ink saturates paper. The third is practical: the Code of Good Practice adopted in July 2025 specifies opt-out modalities[6] but is silent on pre-existing training runs or authors with no digital footprint.

The first front began in March 2025, when France’s principal publishing and authors’ societies filed suit against Meta before the Paris court[7]: copyright infringement and unfair competition, for training Llama on some 200,000 books partly drawn from a pirate library, among them works by Nothomb, Carrère and Ernaux. Meta invoked fair use, a doctrine with no French counterpart. The case is pending. Whatever the court decides, every model provider will feel the ruling.

A second front has opened in Europe. A Hungarian referral to the CJEU raises four questions that strike at the heart of generative AI[8]: whether chatbot responses constitute communication to the public; whether training constitutes reproduction; whether the text-and-data mining exception applies at all; and whether liability attaches to the model provider. The Advocate General’s conclusions are expected in autumn 2026. The judgment will shape French law. While that answer is pending, a concrete ruling has arrived from Munich. In November 2025, a court ordered OpenAI to stop reproducing nine protected song lyrics, rejecting the mining exception outright[9]. Full memorisation, the judges held, exceeds information extraction.

The creditors: the rights that predate AI

The mining exception is the most contested ground, but it is not the only one. France’s sui generis database right, a distinctly European invention without an equivalent in leading technology jurisdictions, protects any producer who demonstrates substantial investment in gathering, verifying or presenting data[10]. Extraction of a qualitatively or quantitatively substantial portion is prohibited. Repeated and systematic extraction that manifestly exceeds normal use is forbidden. Any operator scraping an editorial platform to build a training dataset faces direct exposure, regardless of what happens in regard to the mining exception[11].

The neighbouring rights of press publishers have produced the most concrete results. Google was fined €250m by France’s competition authority in March 2024[12], closing a saga that opened in 2020 over the platform’s failure to negotiate in good faith after their new rights came into force. A bill adopted by the National Assembly in March 2026 and by the Senate in June 2026 goes further: it would authorise the audiovisual regulator (Autorité de régulation de la communication audiovisuelle et numérique or ARCOM) to impose sanctions of up to one per cent of global turnover on operators who circumvent press neighbouring rights.

Music has attracted no comparable fine, but the exposure is structurally identical. Training music-generation models likely constitutes covered reproduction. The text-and-data mining exception applies by cross-reference, with the same gaps identified. A Grand Chamber ruling in Luxembourg in April 2026 narrowed the pastiche exception further[13]: it covers only creations presenting perceptible differences and engaging a recognisable artistic dialogue with the original. For AI output that recombines rather than transforms, the pastiche shield is fragile.

That fragility reaches beyond melodies and pastiche into trademark law. No French court has yet ruled on whether reproducing a registered mark in a model’s outputs constitutes infringement. A London court has come close to an answer: in November 2025, the High Court of Justice held in Getty Images v Stability AI that training on watermarked photographs and producing outputs that reproduce those watermarks infringes the mark-holder’s registered rights[14]. The principle is structural: the platform operator, not the user, bears responsibility for what the model produces in trade. The decision remains subject to an appeal. But both systems draw from the same European directive and the mining exception offers no shelter here. The problem waits patiently in every jurisdiction where such models operate.

The creative void: neither author nor inventor

The rights assembled above press against the machine from the outside. Behind them lies a different question: not what can be claimed against it, but what cannot be given to it. Can the machine be a creator? French IP law answers with a philosophical intuition inherited from Romanticism: the work is the expression of the author’s personality, a projection of selfhood onto the page, the canvas, the score. Authorship is reserved for the human creator alone. Case law bars attribution to a legal entity or a machine[15]. The criterion is unexpectedly radical: it asks not whether the output is statistically original but whether it bears the trace of a human who made choices for specific reasons. A diffusion model navigates a probability space. It makes no choices.

Recent case law traces those contours with precision. A French court in June 2024 held that only the author can identify what reflects their personality, which is obvious until applied to a diffusion model[16]. A 2024 ruling found that originality does not flow from an arbitrary aesthetic choice. A decision issued in February 2026 confirmed that commonplace elements may manifest creative effort if they confer a distinctive physiognomy[17], a phrase drawn from settled French case law, now tested in the AI era. The result is unambiguous: output produced by a minimal prompt attracts no copyright protection. That incapacity does not exhaust the problem. Where AI reproduces or distorts elements of pre-existing works, the original author’s moral right, which is perpetual, inalienable and imprescriptible under French law, is potentially infringed. The text-and-data mining exception offers no cover: moral rights are incapable of waiver by any contract or statute.

In patent law, the same problem takes a different form. French law vests patent rights in the inventor or their successor in title. The European Patent Office has confirmed a machine cannot hold that title[18]. The distinction that matters is between the AI assistant, where an engineer bears responsibility for the result, and the autonomous AI, which identifies the problem and the solution without human input. The first preserves the human inventor. The second exposes a gap that positive law has not filled, and the gap is growing.

Copyright asks whether there is a personality in the work; moral rights ask whether that personality can be erased or distorted; patent law asks whether there is a mind behind the invention. In the age of generative AI, all three questions land in the same empty room.

The senator’s wager: a presumption of use and the politics of proof

That empty room has a practical corollary: rights have been violated and someone must prove it. In June 2025, France’s highest advisory body on IP found that rightsholders cannot exercise their rights because they do not know whether their content was used[19]. The proposed solution, a statutory presumption of use where circumstantial evidence exists, was taken up by a bill filed by a cross-party group of senators led by Laure Darcos.

Adopted by the Senate at its first reading on 8 April 2026[20], the bill would create a presumption: in any civil dispute, a work protected by copyright or a neighbouring right shall be presumed to have been used by the AI model provider, where any element relating to its development, deployment or output renders such use plausible. The Council of State issued a favourable opinion on 19 March 2026, confirming that the presumption targets use, not infringement, and raises no constitutional or conventional obstacle. The bill is now before the National Assembly, whose decision will determine whether rightsholders gain the evidentiary lever the Senate voted to give them.

In practice, if adopted, circumstantial evidence such as output-work similarities, documented presence in a known dataset or query-test results would shift the burden of proof to the provider, creating strong incentives to maintain auditable training logs. Failing adoption, the EU’s AI Act remains the sole framework: general-purpose AI providers must publish a summary of the training data used, a requirement that has applied since August 2025[21].

See everything, prove nothing: the paradox that trade secrets built

That transparency, however, opens another front: trade secrets. French commercial law protects information not generally known, carrying commercial value by reason of its secrecy and subject to reasonable protective measures[22]. Model weights, architecture, hyperparameters and training dataset composition satisfy all three conditions. The paradox is exquisite: verifying copyright compliance requires the very data that evidences the violation. Access the provider may legitimately refuse. The law does, however, contain a seam: the trade secret shield is inoperable when disclosure is required by the investigatory or sanctioning powers of authorities, a carve-out the AI Act fits with precision. A civil judge can already examine a disputed document alone or restrict access to one named individual per party. That instrument has not yet been turned against an AI provider. The standoff illuminates a structural feature of the AI era: the systems most in need of scrutiny are, by design, the least legible. Opacity is not a bug. For many providers, it is the product.

The foundational irony has not been resolved. Both the novelist and the language model feed on the same archive. The law, so far, has found a way to say what the novelist is owed without finding a way to make the machine pay for it. Three developments may shift the evidentiary terrain: the EU AI Act’s full obligations on model providers, the CJEU’s answer to the question from Budapest and the Darcos presumption, if it clears the legislature. None of them reaches the structural problem: the machine’s opacity[23]. The systems most in need of legal scrutiny were built, by design, to resist it. The rights exist: copyright, neighbouring rights, database rights, trademark and patent protection. The remedies exist. What is missing is the grip. Opacity, for now, remains part of the product.

[1] Paris Judicial Court, 3rd Chamber, 3rd Section, May 7, 2025, Case No. RG 25/02294 (New DayFr — AI/press blocking order).