French state employees must first turn to the artificial intelligence tools provided by the Administration. If they want to use the free version of ChatGPT, Claude or Gemini, they must first ask themselves a question: could I publish this information on a website accessible to anyone? Summarizing an already published law or generating an illustrative image are possible uses. Analyzing a citizen's file, summarizing the minutes of an internal meeting or processing medical or personnel data falls outside that use.
The first interministerial guide on the use of AI for French state employees was published in June 2026. It was drafted by the Interministerial Directorate for Digital Affairs (DINUM), the Interministerial Directorate for Public Transformation (DITP) and the General Directorate for Administration and the Civil Service (DGAFP). The document has no regulatory status and serves as a common reference when a ministry lacks its own rules. Where such rules exist, the ministry's rules prevail.
Tools public employees can use
The guide distinguishes three types of tools according to the guarantees they offer and the nature of the information to be processed.
- Administration tools: These are the first option for daily work. They include the interministerial AI assistant, Transcripts, Visio and the tools each ministry makes available to its employees.
- Commercial paid tools contracted by the Administration. These may include paid versions of ChatGPT, Claude or other services. Before using them, the employee must consult the IT service about what data can be entered into each tool.
- Free commercial tools. Their use is very limited and may be prohibited by the organization's rules. Where they are permitted, they may only be used with information that could be made public. If the employee could not publish that information on an open website, it must be kept out of these services.
The guide warns that hosting servers in the European Union is only one of the factors determining control over data. A service hosted in European territory may still be subject to foreign legislation, such as the US CLOUD Act, if the company operating or controlling it is subject to that law. For sensitive uses, the document recommends solutions with enhanced control guarantees, preferably with the SecNumCloud qualification granted by the French National Agency for the Security of Information Systems (ANSSI)
Freeing up time for employees
The goal is to free up time for public employees and preserve the ability to adapt the service to each citizen. The guide covers generative AI, which can produce text, images, sound and code, and specialized tools in common use, such as audio transcription, character recognition and machine translation.
California sues OpenAI over leaks from its AI agents
The document cites specific tasks in which these tools can help, such as transcribing meetings, summarizing lengthy reports, rephrasing technical texts, translating documents or organizing ideas before drafting a note. For the assistance to be useful, it recommends that the employee master the task being delegated, that the tool be incorporated into the work with the participation of those who know the craft, that it be used when it adds value and that sources be checked when available.
Generative AI can invent data, cite nonexistent laws or fabricate bibliographic references. Its responses are generated from statistical probabilities and can reproduce stereotypes present in the training data. Base models can answer without citing sources, and their knowledge is limited by the cutoff date of their training data.
The document also warns that a tool can be manipulated through instructions hidden in a file and that its use can cause involuntary information leaks. It adds the environmental cost. Every query requires energy and water. As an indicative reference, the guide estimates that a text query may be equivalent to between one and ten Google searches and that generating an image may consume about 60 times more than a text query.
How to assign tasks to AI and review the results
The guide compares writing instructions for AI to assigning a task to an intern. It recommends explaining the context, providing the reference document when needed, specifying the assignment briefly, indicating the response format and requesting revisions if the first result needs adjustments. When the task is complex, it suggests asking the tool to pose questions to clarify what is needed.
Before disseminating the result, the employee must review it with judgment, adapt it to the context and reread it. An annex of more than thirty pages gathers examples from human resources, project management, communication, development, legal affairs, finance and management. Each entry describes the situation, includes a reusable instruction, explains what result to expect and points out which aspects require special attention. The final check is up to the employee, who must cross-check figures, references and rules against the original document.
The guide places data protection, transparency and human review at the center of professional AI use. When the tool plays a substantial role in producing content, an analysis or the preparation of a decision, its use must be made explicit. Responsibility remains with whoever produces, decides or signs. The framework also offers other administrations a practical reference for incorporating these tools while respecting public service principles.