
Marketing teams generate visuals in seconds, developers ship code suggested by assistants, startups train models on large datasets. Each of these activities raises a copyright question, and the answers differ between the United States and Europe. Two issues dominate: are AI outputs protected, and can protected works be used to train AI models?
This guide summarises where the law stands on both sides of the Atlantic as of late 2026, and flags the contractual steps Romain Waiss-Moreau recommends to companies that create or commission content with AI.
In the United States, the answer is now settled at the highest level available. On 2 March 2026, the Supreme Court declined to review Thaler v. Perlmutter, leaving in place the D.C. Circuit's 2025 ruling that copyright requires a human author. The Copyright Office registers works that combine human and AI contributions, but only protects the human elements, and expects applicants to disclose and exclude AI-generated material.
In France and the EU, protection requires originality, understood as the reflection of the author's free and creative choices. A purely machine-generated output does not meet that standard. Human selection, arrangement, editing or substantial modification can be protected.
| Scenario | Likely protection |
|---|---|
| Image generated from a short prompt, used as is | Little or none |
| Text written by a person, edited with AI assistance | Yes |
| AI outputs substantially selected, edited and combined by a person | Possibly, for the human contribution |
| Code generated by an assistant and largely rewritten | Possibly, for the rewritten parts |
In the United States, training is assessed under fair use, case by case. Several district courts ruled in 2025 on training by AI developers, reaching nuanced outcomes that depend on how the data was acquired and whether outputs compete with the originals. The question remains unsettled at the appellate level.
In the EU, training relies mainly on the text and data mining exception of Directive (EU) 2019/790. Commercial mining of lawfully accessible works is permitted unless the rightholder has opted out in an appropriate manner, such as machine-readable signals for online content. Since 2 August 2025, the AI Act also requires providers of general-purpose AI models to adopt a copyright policy that respects opt-outs and to publish a sufficiently detailed summary of their training content.
France is going further. A Senate bill creates a rebuttable presumption that a protected work has been used by an AI provider when evidence makes that use plausible, for instance when a model reproduces extracts or imitates an author's style closely. The Conseil d'État gave a favourable opinion on 19 March 2026 and the Senate adopted the bill unanimously on 8 April 2026. If enacted, it would shift the burden of proof onto AI providers in French proceedings.
The EU text and data mining exception allows training unless the rightholder has reserved its rights appropriately. For online content, rightholders typically combine:
None of these is foolproof. Together, they document a clear reservation that matters in litigation and in licensing negotiations with AI providers.
| Question | United States | European Union / France |
|---|---|---|
| Is purely AI-generated output protected? | No, human authorship required | No, originality requires human creative choices |
| Is AI-assisted work protected? | Yes, for the human contribution, disclosed on registration | Yes, for the human creative contribution |
| Legal basis for training on protected works | Fair use, assessed case by case | Text and data mining exception, subject to opt-out |
| Transparency on training data | No general statutory duty | AI Act summary of training content for general-purpose models |
| Burden of proof on use of a work | On the rightholder | Proposed presumption of use in France |
A U.S. agency produces a campaign for a French client, partly with an image generator and then retouched by designers. Its standard contract assigned all copyright and warranted originality. Both clauses were risky: the agency can only assign the protectable human contribution, and cannot warrant that the tool's output does not resemble existing works. The revised contract discloses AI use, limits the assignment to protectable elements, passes through the tool provider's indemnity and adds a clearance check before publication. The firm recommends this structure to any business delivering AI-assisted work to clients.
Yes. Trademark law requires distinctiveness, not originality. Copyright protection for the logo may be weak, so the trademark becomes the main asset.
Yes. Applicants must identify and exclude more than de minimis AI-generated material and describe the human contribution.
The user who exploits the infringing output is exposed. Recourse against the tool provider depends on contractual indemnities, which vary widely between providers and plans.
A sufficiently creative prompt can itself be an original text. That protection covers the prompt, not the image or text the system generates from it.
You set out your project. By the end, you will know whether one of the four services fits, or whether we should direct you elsewhere. EN or FR, Paris or New York, confidential.