Key facts about AI development
- Without process traceability, copyright protection of AI-generated code becomes difficult to defend
- Keeping and time-stamping the dialogue with the model (prompts, responses, iterations) enables you to pre-constitute proof of originality, justify a claim to copyright protection and document the work of developers in the context of a Research Tax Credit application.
- eIDAS-qualified time stamping enables you to build up an enforceable file of evidence as you go along
- API integration allows developers to track their work without interrupting the workflow
Coding with AI: how the developer’s job is changing
The advent of code generation tools has profoundly altered the developer’s daily routine. Where once they wrote every line, now they orchestrate: they formulate instructions, evaluate model suggestions, refine them, reject or validate them. The benefits are real: repetitive tasks are speeded up, and the quality of the code produced improves for certain types of problem.
Adoption of these new practices is massive. GitHub Copilot had 20 million users in July 2025[1]. According to the Stack Overflow Developer Survey 2025, 84% of developers use or plan to use AI tools in their workflow.
This role shift is real. But it does not mean that the developer disappears from the creative process. On the contrary: it is the developer who sets the constraints, who selects, who arbitrates. His intervention remains central, provided it is documented.
What are the challenges of development with AI?
AI-assisted development raises legal issues that technical teams are rarely prepared to anticipate. The first of these concerns a fundamental prerequisite: originality.
Code originality: a difficult prerequisite to prove
The legal question raised by AI-assisted development is, in fact, a fairly old one. For software to be protected by copyright, it must be original, as defined in article L. 111-1 of the French Intellectual Property Code.
In its Pachot ruling of March 7 1986[2], the French Supreme Court clarified this criterion for software: originality presupposes a personalized effort reflecting know-how, over and above the simple implementation of an automatic and constraining logic.
In practice, this criterion is rarely documented as the case progresses. In the absence of evidence of the creative process, the judge, who is not a code technician, tends to conclude that there is no originality, thus putting an end to the infringement debate before it has even really begun.
AI-assisted development doesn’t create this problem: it reveals and amplifies it. From now on, it’s not just a question of proving that the developer made creative choices. It’s also a question of proving how much of the work is attributable to the developer, and how much to the model.
As the law currently stands, no legal or praetorian rule grants copyright protection to a work whose human part cannot be identified and documented. This places the developer who codes with AI in a particularly exposed evidential situation.
This is precisely where the traceability of interactions with the model comes in. Documenting prompts, responses and iterations throughout the creative process simultaneously solves both problems?
- it is precisely in these arbitrations (the choice of one prompt over another, the rejection of a suggestion, the reformulation of an instruction) that originality in the legal sense resides?
- this is what the judge needs to see documented in order to appreciate the human contribution to the design.
Liability and regulatory framework: who bears the burden of proof?
Liability in the event of a dispute follows the same logic: whether it’s a dispute over authorship, an infringement action or a contractual dispute over code ownership, it’s always the developer or company that must provide proof, never the model.
As for the regulatory framework, it’s still a work in progress. The senatorial report Création et IA: de la prédation au partage de la valeur [3] (July 2025) and the mission entrusted by the CSPLA to Alexandra Bensamoun [4], whose conclusions are expected in summer 2026, outline the contours of a law under construction. In these uncertain times, anticipation is better than waiting.
Traceability of AI prompts: a blind spot to be filled
A Git repository records what has been produced. It says nothing about how or why. The succession of commits traces the evolution of the code. But it doesn’t document the decisions that shaped it: why this architecture rather than another, why this prompt was reformulated three times before it came to fruition, why this suggestion from the model was discarded. Yet it is in these decisions that the developer’s originality is expressed.
At theAgence pour la Protection des Programmes conference Clean Code, Clean IP: coding in the age of AI[5], Jérémy Pappalardo reminded us that the practical challenge lies precisely in demonstrating that behind the generated code, a human being has made choices. To do this, he emphasized the need to keep model prompts and responses as preparatory design material (in the same way as mock-ups or working notes in other creative fields) and to timestamp them to establish the anteriority and chronology of the creative process, thus building up an evidence file that can be presented to a judge.
Prompt management tools exist. However, they have no probative value: they organize but do not certify. Saving prompts without timestamping them means building up a file without guaranteeing its date certainty or integrity, both of which are crucial in court.
Qualified timestamping via API: native integration into the IA development workflow
Qualified timestamping as defined by the eIDAS regulation cryptographically associates an exact date and time with a digital document, with a legal presumption of accuracy recognized throughout the European Union.
Applied to prompts and code versions, it enables you to build up a structured file of evidence as you go along: each significant stage in the process (selected prompt, validated version, intermediate deliverable) is dated and sealed in an enforceable way.
Prepare a file of opposable evidence
Documenting interactions with the model is not enough if this documentation is not certified. A log file, a screenshot, a conversation export: none of these elements guarantees the date or integrity of the data in the legal sense. eIDAS-certified timestamping gives each step in the process a definite date and verifiable integrity, thus constituting an evidence file that can be presented to a judge in the event of a dispute over authorship or an infringement action.
Track your work without interrupting your workflow
Adopting this functionality via API, as offered by Evidency, enables direct integration into existing development workflows, without modifying teams’ work habits. In a single call, a qualified timestamp token is generated, verifiable at any time. Its implementation requires no legal expertise: it’s a technical tool, designed for development teams.
An asset for Research Tax Credit applications
Traceability of interactions with AI is an additional advantage for companies undertaking research and development work. The Research Tax Credit requires documentation of the nature and scope of the work carried out. A time-stamped history of prompts, iterations and code versions provides a solid documentary basis for justifying the eligibility of expenditure and responding calmly to any tax audit.
For a CTO, the three benefits are therefore complementary: reducing the risk of disputes over the ownership of code produced with AI, integrating frictionless traceability into existing workflows, and optimising the company’s CIR file.
AI-assisted development: traceability as the new standard
Generative AI has transformed the speed of code production. It has not changed the rules governing its legal protection. Software produced with the help of a model remains subject to the same originality requirements as code written line by line. But a new difficulty has been added:human intervention must now be proven, not just asserted.
In this context, documenting the development process is no longer an incidental precaution. It’s a professional practice in its own right, just like versioning or unit testing. Developers and companies who integrate it into their workflows today gain a concrete advantage: the ability to defend and value their work, regardless of the role played by the model in its realization.
Sources
1 Microsoft Investor Relations, FY25 Q4 Earnings Call Transcript (July 30, 2025): https://www.microsoft.com/en-us/Investor/earnings/FY-2025-Q4/press-release-webcast
2 Cass. ass. plén., March 7, 1986, Pachot decision, pourvoi n°83-10477: https://www.legifrance.gouv.fr/juri/id/JURITEXT000007016934
3 Senate, information report no. 842 (2024-2025), Création et IA: de la prédation au partage de la valeur, rapp. A. EVREN, L. DARCOS and P. OUZOULIAS, Commission de la culture, de l’éducation, de la communication et du sport, submitted July 3, 2025: https://www.senat.fr/rap/r24-842/r24-842.html
4 https://www.culture.gouv.fr/nous-connaitre/organisation-du-ministere/conseil-superieur-de-la-propriete-litteraire-et-artistique-cspla/travaux-et-publications-du-cspla/missions-du-cspla/le-cspla-lance-une-mission-relative-a-la-protection-des-contenus-generes-avec-le-recours-a-l-ia-generative
5 AI-generated code: tracing prompts, Expertises des systèmes d’information, May 2026, N°523 p.6
Disclaimer of liability
The opinions, presentations, figures and estimates presented on the website, including the blog, are intended for information purposes only and should not be construed as legal advice. For legal advice you should contact a legal professional in your jurisdiction.
The use of the content of this website, including the blog, for commercial purposes, including resale, is prohibited without prior authorization from Evidency. The request for permission must specify the purpose and extent of reproduction. For non-commercial purposes, all material in this publication may be freely quoted or reprinted, but acknowledgement is required, together with a link to this website.
