Burden of proof is the boring part of copyright law until it isn’t. Right now, a rights-holder suing an AI company over training data use faces a significant obstacle: proving the model actually used their specific work. That’s technically and legally difficult. France’s reported legislation, identified as Article L. 331-4-1 in Jones Day’s legal analysis, flips that default.
Under the reported mechanism, use of a copyrighted work is presumed if use is “plausible.” The rights-holder doesn’t need to prove the work was in the training data, they need to show it’s plausible it was. The AI provider must then disprove use. According to Jones Day’s analysis, this provision would trigger on plausibility, not proof, restructuring the litigation risk profile for any AI company deploying in France.
That’s a structural change in copyright enforcement. Not a fine increase. Not a new registration requirement. A shift in who has to prove what.
The practical consequence is significant even before litigation begins. A “plausibility” standard for triggering presumption is low by design. A popular novel, a widely-distributed dataset, a commonly-scraped news archive, if it’s plausible that any of these appeared in a training corpus, the burden is on the AI provider to show they didn’t use it. Companies that haven’t maintained detailed data provenance records face obvious exposure. Companies that have may face the cost of litigating their records in French courts.
Don’t expect this to stay a French problem. The mechanism is watching as a potential template. The EU is wrestling with the same training data question, the copyright fight in Brussels, detailed in prior hub coverage, involves the same core tension between rights-holder attribution and AI developer immunity. If France’s rebuttable presumption model survives constitutional review and produces the first enforceable judgments, it gives other jurisdictions a functioning model to adopt.
Unanswered Questions
- What evidentiary standard satisfies 'plausibility', does a model trained on Common Crawl automatically trigger presumption for any work in that corpus?
- Does the mechanism apply to models trained outside France but used by French-domiciled users or entities?
- What documentation would constitute sufficient disproof, a negative training data audit, a data exclusion list, or cryptographic provenance records?
What to watch
Legislative status matters here. Senate passage is not final enactment, the reporting from Jones Day’s client alert covers the French Senate’s action, but full enactment requires additional steps including Presidential signature under the French legislative process. Article L. 331-4-1 as a designation comes from Jones Day’s analysis and hasn’t been independently verified against primary legislative text . Compliance teams should treat the French mechanism as a live risk to monitor, not a confirmed obligation to implement, yet.
The real question is whether the “plausibility” standard holds as written through French judicial review. Copyright presumptions have been challenged in other jurisdictions on proportionality grounds. If Article L. 331-4-1 survives its first constitutional challenge intact, it becomes the strictest training data liability standard in any major jurisdiction.