The rise of generative AI models raises many legal questions in France. At the dawn of 2025, legislation seeks a balance between protecting authors and technological innovation. This article deciphers the latest developments in copyright law, the specific protection criteria for works derived from artificial intelligence, and the responsibilities incumbent on the various stakeholders.
🔑 Shared responsibility: both the model publisher and the end user are now held accountable in case of copyright-related disputes.
📊 Required originality: for a generated work to be protected, human intervention must be significant, whether through settings, data choices, or creative editing.
⚖️ Collective management: collective management organizations are adapting their scales to include rights arising from training AIs on protected catalogs.
📝 Best practices: keeping prompt logs, specifying the creative contribution, and verifying the license of the used model reduce legal risks.
Somaire
Context and Legal Developments in 2025
Recent Revisions to the Intellectual Property Code
France has adjusted its Intellectual Property Code to incorporate the requirements of the 2021 European directive on the protection of AI-generated works. In 2024, a decree clarified the notion of fair use of data during the training phase, imposing compensation mechanisms on original rights holders. Now, the law requires model providers to document data sourcing, under penalty of financial sanctions.
Concept of AI-Assisted Creation
In practice, a work born from a generative model benefits from no protection in the absence of an identifiable human imprint. Case law has confirmed that AI, lacking legal personality, cannot claim any rights. Only works presenting a human trace — for example, corrections, complex prompt composition, or post-production work — can be considered original.
Protection Criteria for Generated Works
Originality and Human Imprint
The judge now evaluates originality based on two axes: the selection of training data and the degree of human intervention. In a landmark 2025 ruling, the Court of Cassation refused protection for a 3D illustration generated without editing, but granted it as soon as the author had added additional graphic elements.
The Role of the User: Settings and Intervention
A simple prompt (“create an urban landscape”) is not enough: the law requires precise settings (weight of styles, choice of filters, assembly of multiple requests). It is this level of customization, often documented in the software execution logs, that establishes the direct link between the human and the final work.
Responsibilities and Royalties
Civil and Criminal Liability
In the face of a disputed work, two actors can be prosecuted: the model publisher, as the instigator of the use, and the user, who triggered the generation. The legislator specified that a negligence fault (lack of license verification) can lead to civil liability. In case of serious infringement, criminal penalties are possible, aligned with those applicable to classic copyright offenses.
Collective Management and Licenses
Collective management organizations (CMOs) have deployed new framework agreements to allow remuneration of rights holders whose works are used to train AI. A summary table below compares the main adopted schemes:
| CMO | Scope | Remuneration Mechanism | Indicative Rate |
|---|---|---|---|
| SGDL | Literary texts | Annual lump sum payment | 0.5% of AI turnover |
| ADAGP | Visual works | License based on volume of images | 1.2% of AI turnover |
| SAIF | Sound works | Royalty for reproduction | 0.8% of AI turnover |
Best Practices for Creators and Companies
Documentation and Traceability of Prompts
By recording each parameter at the time of generation, the prompt becomes probative evidence. This approach, backed by a timestamping system, allows demonstrating the original nature and authorship of the work in case of dispute.
Choice of Open-Source vs Proprietary Models
Open-source models offer full transparency on data and license but may lack performance or support. Proprietary solutions, on the other hand, often provide a turnkey service accompanied by a clear license agreement. The choice will therefore depend on the balance between legal security and rendering quality.
FAQ
Can a work entirely generated by AI be protected?
No: French law requires a human creative contribution. Without intervention, the work remains royalty-free and does not open any copyright rights.
Who pays royalties for model training?
Model providers and end users may be required to pay compensation to collective management organizations, depending on the share of protected data used.
How to challenge suspected generated content?
It is necessary to gather all evidence of originality (project files, generation logs, correspondences on the prompt) and to bring the case before the competent judicial court to have the infringement recognized.
