Artificial intelligence raises two very practical copyright questions today: on what conditions may works be used to train a model, and to what extent can the resulting output be protected?
To determine which rights can be relied on, you have to examine how the works were used, the share of human input in the creation and the intended use of the output.
In its report of 16 July 2026, the French Higher Council for Literary and Artistic Property (CSPLA) rejects the idea of a copyright regime designed specifically for artificial intelligence. The question remains the familiar one: can an original human creation be identified in the output?
Do you use AI in a creative process, or may your works have been used to train a model? Cabinet Bouchara & Avocats helps you analyse your literary and artistic property rights.
Key points
- An AI-assisted creation can be protected where decisive human choices are found in its final form.
- Training a model means checking the conditions of access to and use of the works, in particular in the light of the text and data mining exceptions and any opt-out by rights holders. Content accessible online is not necessarily free to use.
- Terms of use may authorise the exploitation of an output without conferring copyright protection on it or guaranteeing the absence of third-party rights.
- You must be able to document what was asked of the AI, what was changed by a human and what was finally released.
AI and copyright: which questions should be kept apart?
The first thing to establish is at what point the AI comes in.
If the question concerns training the model, you have to look at which works went into the corpus and on what terms.
Once the content has been generated, two checks remain: can it be protected in itself, and does it reproduce material belonging to a third party?
Can a creation made with AI be protected by copyright?
Yes, on one condition: the human input must be identifiable and demonstrable in the output. Using AI does not, on its own, rule out copyright.
Under French law, protection requires a work of the mind created by an author (article L. 111-1 of the French Intellectual Property Code).
The CSPLA distinguishes two situations:
- hybrid creation, in which a person retains genuine creative direction and whose output may be original;
- synthetic production, generated without significant human input, which falls outside copyright.
The AI is not the author. What must be demonstrable is that the final form comes, at least in part, from human creative choices.
Which human choices support a copyright claim?
Human input can appear at different points in the creative process: before generation, over successive iterations, or at the stage of selection and retouching.
Human input at any of these points can help establish an original creation, provided the creative choices relied on are decisive enough and perceptible in the final work.
| Favourable indicators | Counter-indicators |
|---|---|
| A creative project defined upstream | A general request put to the tool |
| Instructions that determined elements visible in the output | A detailed prompt with no identifiable effect on the final form |
| Iterations guided by genuine creative choices | Successive retries until a satisfactory result appears |
| A reasoned selection among several proposals | Immediate acceptance of the first result |
| Substantial and creative retouching | Purely technical corrections |
| Decisive personal input | An interchangeable user |
The number of prompts is therefore not, in itself, a criterion of originality.
In practice, it is better to keep this material as the work proceeds than to try to reconstruct the process once a dispute has begun.
Cabinet Bouchara & Avocats’ note
“You have to be very precise in order to prove each instruction given and to trace how each request, and your own contribution, developed.”
Vanessa Bouchara, copyright lawyer
Who owns the rights in content created with AI?
If the output can be protected, the question remains who can claim authorship and the rights attached to it. Authorship belongs, unless proved otherwise, to the person or persons under whose name the work is disclosed (article L. 113-1 of the French Intellectual Property Code).
In a company, as soon as an employee, an agency or a freelance has been involved, the contracts must be examined carefully. Having commissioned or paid for a creation is not enough to transfer all copyright. The assignment must in particular specify the rights transferred and delimit their field of exploitation (article L. 131-3 of the French Intellectual Property Code).
A word of caution: the fact that terms of use state that the output “belongs” to the user does not automatically give that user a copyright. Nor do those terms guarantee that the output does not infringe a third party’s rights.
Is a detailed prompt enough to be recognised as the author?
No. A very precise prompt is not enough, on its own, to make its author the author of the output.
What counts is being able to show that those choices actually determined elements perceptible in the final form. The number of iterations, or the time spent querying the tool, is not in itself proof of originality.
A decision of the Munich court of 13 February 2026 illustrates the difficulty well: three logos generated from prompts were not recognised as works, for want of a sufficiently predominant human creative influence on their final form. The decision is German and of course does not bind the French courts, but it shows very concretely why the human contribution to the output must be traceable (analysis by the legal affairs department of the French finance ministry).
On what conditions may a work be used to train an AI?
The rules differ according to the purpose of the text and data mining. Two situations must be distinguished:
- Mining for scientific research purposes may be carried out by certain bodies on works to which they have had lawful access, subject to the conditions laid down by the French Intellectual Property Code.
- Mining for other purposes may also cover lawfully obtained works, but the rights holder may opt out in an appropriate manner (article L. 122-5-3 of the French Intellectual Property Code).
For content available online, that opt-out does not have to be reasoned. It may in particular be expressed through metadata, a machine-readable process or the site’s terms of use (article R. 122-28 of the French Intellectual Property Code).
It is also necessary to look at what the training data contains. A photograph, a text, a piece of music or the content of a database may be protected by different rights. And where the corpus itself constitutes a protected database, the rights of its producer must also be taken into account (article L. 342-3 of the French Intellectual Property Code).
The mere fact that content is accessible on the Internet therefore does not mean that it may be freely used to train an AI.
Providers of general-purpose AI models must comply with two obligations:
- put in place a policy to comply with European Union copyright law, in particular in order to take account of the rights reservations expressed by rights holders;
- draw up and make publicly available a sufficiently detailed summary of the content used to train the model (article 53 of the AI Act).
That summary must make it possible to better understand the nature of the content used for training. But it constitutes neither a copyright licence, nor an agreement on remuneration, nor an exhaustive list of all the works used.
Can AI-generated content constitute infringement?
Yes. The fact that AI-generated content is not itself protected by copyright does not mean that it may be freely exploited. It may reproduce protected elements belonging to a third party.
The output must in particular be examined in order to determine whether it reproduces original and identifiable elements of an earlier work: a character, a composition, certain graphic choices or elements specific to a visual universe. The reproduction, performance or distribution of a work in breach of its author’s rights may constitute infringement (article L. 335-3 of the French Intellectual Property Code).
An AI-generated advertising image may therefore raise difficulties if it reproduces the original features of an existing character or universe. For a brand or an agency, it is therefore the output actually published that must be examined, and not merely the prompt that was used.
Does a creation “in the style of” constitute infringement?
Mere stylistic proximity is not enough to establish infringement. The risk arises where the generated content reproduces identifiable original elements of an earlier work: a character, a composition, a setting or a recognisable graphic combination.
It is this concrete reproduction that must be compared with the work relied upon.
For a company, a brand or an agency, this check is best carried out before publication, rather than after receiving a claim.
Contact an infringement lawyer to have content analysed before it is published, or to take action where your creations have been reproduced.
What evidence should be kept where an AI is involved in the creation process?
Where a company creates using generative AI, it is well advised to keep whatever makes it possible to retrace the process: the relevant prompts, the successive versions, the edits made and, of course, the content finally published. Contracts and terms of use must also be kept.
The CSPLA also proposes a method for analysing evidence, based on examining the various elements that make it possible to establish human involvement and the creative choices found in the result.
Where a rights holder suspects an infringement of its rights, it must first gather the earlier work, the evidence of creation and ownership, the disputed content and the material making it possible to compare them.
Court action is not necessarily the first step. Depending on the case, it may be useful to have the publication formally recorded, to request explanations, and then to consider a formal notice, a negotiation or an action.
In this type of case, the Firm first examines the rights that can actually be relied upon, then the contracts and the evidence. These checks then make it possible to decide whether to opt for a formal notice, a negotiation or an infringement action.
Do you use AI to create content, or do you believe your works have been reproduced?
FAQ
Can content that is entirely AI-generated be freely reused?
No, not automatically. Even if the output is not itself protected, it may reproduce elements in which a third party holds rights. In particular, it is necessary to check for the presence of protected elements belonging to third parties, the applicable contractual terms and, depending on the circumstances, other legal grounds.
Do the terms of use of an AI tool guarantee the absence of infringement?
No. The terms of use govern the relationship between the user and the provider of the tool. They do not guarantee that the output is original or that it does not reproduce any protected element belonging to a third party.
Can a prompt be protected by copyright?
A prompt may be protected by copyright if it is itself original, that is to say if it reflects free and creative choices in its wording, its structure or its sequencing. Conversely, a purely functional or commonplace instruction, or one dictated by a technical result, is unlikely to be protectable. And even where a prompt is protected, that protection covers the prompt itself: it does not, on its own, make it possible to claim copyright in the generated content.
Can an author be remunerated where their work is used to train an AI?
The use of a work to train an AI does not, on its own, give rise to an automatic right to remuneration. Where the use does not benefit from the text and data mining exception, in particular because the rights have been validly reserved, the exploitation may require authorisation and give rise to the negotiation of a licence.
Sources
- French Intellectual Property Code: articles L. 111-1, L. 113-1, L. 122-5-3, R. 122-28, L. 131-3, L. 335-3 and L. 342-3.
- CSPLA, Mission report on the status of artificial intelligence productions, 16 July 2026.
- Regulation (EU) 2024/1689 on artificial intelligence, article 53.
- Munich Regional Court, decision of 13 February 2026.
- Legal Affairs Directorate of the French economic and financial ministries, Lettre de la DAJ, no. 402, June 2026.
- Directive (EU) 2019/790, articles 3 and 4.
