Legal AI News

The Darcos Law on AI and Copyright in France: 20,000 Signatures to Force the Vote — AI News

AI news · AI copyright · AI France: full legal analysis of an evidentiary revolution stalled at the National Assembly.

April 8, 202616 min read

Article updated April 8, 2026 — AI news, AI copyright, AI France

20,000 cultural professionals have signed. The Alliance de la presse d'information générale, every French daily newspaper, hundreds of journalists, photographers, artists, directors, composers — all calling for the same thing: that the French National Assembly urgently take up the bill known as the Darcos law on AI and copyright.

The text, adopted unanimously by the Senate on April 8, 2026, is now waiting only to be placed on the deputies' agenda. So far, the National Assembly remains silent.

Yet this law could be the legal turning point the creative world has been waiting for since the rise of ChatGPT. Because it does not merely ask AI giants politely to respect copyright. It flips everything: from now on, it will be up to platforms to prove that they did NOT scrape protected works. No longer up to the artists.

As a legal professional trained in AI, I went through the full legislative dossier — the 496 pages of the Senate report, the Conseil d'État opinion, the floor debates, the hearings. What I found goes well beyond a simple quarrel between creators and tech. It is a fundamental legal battle on intellectual property in the age of artificial intelligence.

And the timing is brutal: while France hesitates, generative AIs swallow billions of French content items every day — without paying a cent.

Darcos law: presumption of exploitation at the heart of AI news

The bill consists of a single article. Just one provision — but explosive enough to require the Conseil d'État opinion.

Here is the heart of the mechanism, as adopted by the Senate on April 8, 2026:

"Unless proven otherwise, an item protected by copyright or by a related right is presumed to have been exploited by the artificial intelligence system, where any indication relating to the development or deployment of that system or to the result it generated makes such exploitation likely."

In plain legal terms: if an AI produces content that resembles your work, it is up to the AI provider to prove that they did NOT use it for training. No longer up to the artist to prove the looting.

It is a Copernican reversal of the law of evidence.

AI and copyright in France: today's evidentiary nightmare

Today, if you are a photographer and discover that a generative AI produces images in your exact style — your distinctive visual signature, your characteristic framing, your color treatment — you are stuck.

To sue, you must prove that your photos were in the training set. But these datasets are opaque. Midjourney, Stable Diffusion, DALL-E — none of them publishes the exhaustive list of ingested works. You may suspect, observe an unsettling resemblance, but without access to the datasets, proof is impossible.

Senator Pierre Ouzoulias put it bluntly during Senate debates, according to Public Sénat: "It is absolutely impossible. AI harvests, or rather trawls, with vast fine-mesh nets that catch every species, including protected ones."

Senator Laure Darcos, interviewed by France Inter: "Until now, it was for the artist to prove that an AI provider had used their original work to generate other images or sounds. With this bill, it will be for AI providers to justify that they did not use that original work in training their generative AI model."

This is what makes the text legally fascinating. It does not create a new right. It does not redefine copyright. It only changes the burden of proof in civil litigation.

But that "only" changes everything.

Legal architecture: rebuttable presumption, not absolute

The Conseil d'État validated the constitutionality and conventionality of the text in its opinion of March 19, 2026.

Crucial point: this is a rebuttable presumption, not an absolute one.

The AI provider can rebut the presumption by showing:

That the content was not used in training
That the opt-out (robots.txt, refusal metadata) was respected
That access was unlawful
That the use falls under the Text and Data Mining (TDM) exception of EU directive 2019/790

But it is now up to the provider to bring forth evidence — its training logs, exclusion lists, filtering mechanisms.

Opacity becomes a handicap for platforms, no longer a shield.

20,000 signatures: a petition to force the Assembly's hand

The petition reported by Journal du Net brings together an impressive spectrum of creators: the Alliance de la presse d'information générale (Le Monde, Le Figaro, Libération, Les Échos, regional dailies), the SNEP (recorded music), the SAIF (visual arts), collective management organizations, plus thousands of journalists, photographers, directors, composers, screenwriters, graphic designers and translators.

Their demand is simple: that the National Assembly place the text on the agenda before the parliamentary recess.

Because every passing day means billions more tokens scraped from French content — without authorization, without compensation.

AI France vs. global AI: why France is moving alone

Why a national law when AI is global and the EU has adopted the AI Act?

Because Europe missed the boat on copyright.

The EU AI Regulation 2024/1689 imposes transparency obligations on AI model providers (Article 53 requires a "sufficiently detailed summary" of training content). But the regulation says nothing about evidentiary mechanisms in litigation — and does not reverse the burden of proof.

Worse: EU directive 2019/790 created a copyright exception for Text and Data Mining (TDM) — precisely the activity of generative AIs. The TDM exception allows reproduction and extraction of online content unless the rightsholder has expressed reservations "in an appropriate manner, in particular through machine-readable means" (i.e., a robots.txt that says "don't scrape me"). Largely theoretical: few creators know how to implement it, nothing forces scrapers to respect it, and even with robots.txt in place, proving it was ignored is… an evidentiary nightmare.

Hence the Darcos law. France says: Europe did not go far enough, we are taking the lead.

Three judicial scenarios

Case 1: photographer vs. Midjourney

Before: Sophie, a fashion photographer, sees Midjourney generate images in her exact style. She sues. She must prove her photos are in the training set. Impossible — the dataset is not published. She loses.

After: Sophie sues with an "indication": 20 Midjourney generations showing her signature style. The presumption kicks in. Midjourney must prove they did not use her portfolio. They either expose their logs or pay damages.

Case 2: newspaper vs. ChatGPT

Before: Le Monde finds ChatGPT quoting near-identical passages from paywalled articles. They sue OpenAI. They must prove use. OpenAI invokes "TDM exception, lawful access, no robots.txt." Le Monde must prove the contrary.

After: Le Monde shows the citations. Presumption activated. OpenAI must prove (1) lawful access, (2) robots.txt was respected, (3) the TDM exception applies, (4) no "unjustified prejudice" to the newspaper's interests.

Case 3: composer vs. an AI music model

Before: Jean, a film score composer, hears a Suno output with harmonic progressions and orchestrations strangely similar to his soundtracks. He suspects, but proves nothing.

After: He files 10 extracts showing the similarities. Presumption. Suno must open its musical training datasets and prove Jean's works are not in them. Or negotiate a retroactive license.

In all three cases, the balance of power tips.

Limits of the text: what it does NOT do

This law is not a magic wand.

It does not create an automatic right to remuneration. It only eases evidence in litigation.

It applies only in France. A US-based AI provider can ignore it… until a French artist sues in a French court.

It does not solve remuneration. The text helps prove unlawful use; the amount of damages is left to the judge — no scale.

It assumes creators able to litigate. Hence the importance of collective management organizations.

It does not prevent training, it makes it risky. Strong incentive to (1) negotiate licenses upstream, (2) filter protected content, (3) implement respected opt-outs.

This is regulation by legal risk.

My analysis: between hope and lucidity

What this law does well. It identifies the real problem — evidentiary asymmetry — and answers it surgically. It sends a strong political signal: France will not let tech giants strip-mine its cultural heritage for free. And it creates a European precedent: if it works in France, Germany, Italy and Spain will follow. That is how the GDPR was born.

What it does not do. It does not directly compensate creators. It does not address the economics of generative AI. It may arrive too late: billions of French works are already inside GPT-4, Claude 3, Gemini, Llama 3.

What is still needed. A mandatory collective licensing system, on the model of private copy levies. Total dataset transparency, beyond a "summary." Deterrent sanctions proportional to global revenue, on the GDPR model (up to 4% of worldwide turnover).

Above all: the Assembly must vote on this text.

Conclusion: 20,000 signatures to force the hand

The 20,000 petitioners are not asking for the moon. They are asking elected officials to do their job: examine a text adopted unanimously by the Senate, validated by the Conseil d'État, and demanded by the entire cultural sector.

If the Assembly blocks the bill under tech lobbying pressure, it sends a catastrophic signal. If it adopts it, France becomes a pioneer once again — as it did with the right neighboring publishers' rights law that later inspired Europe.

Until then, AI keeps scraping. Every day. Billions of tokens. For free.

Official sources

Press and analyses

AI usage statement

The author used Claude (Anthropic) as a research assistant for the systematic collection of official parliamentary sources, verification of compliance with legislative texts, and documentary structuring of the article. All legal analysis, opinions expressed, practical examples, criticism of the mechanism, and final drafting are entirely human. Legislative quotations were verified against the official sources of the Senate and the Conseil d'État before publication.

Cookie settings