Assess the product before people use it
A French expert panel has recommended that consumer AI chatbots face a safety assessment before entering the European market. The report, delivered to President Emmanuel Macron on November 10, was drawn up by a commission convened by France’s Council for AI and Digital Affairs. The proposal would cover consumer chatbots embedded in other services as well as standalone assistants.
The panel wants an assessment of how a service might exploit or worsen vulnerabilities in its users, including minors. One example is a design that repeatedly flatters someone or makes an assistant seem more human to keep a conversation going. The recommendation concerns the way the product behaves, not just which model it uses.
Revisit the result when the product changes
The panel proposes regular checks after release, with another assessment after a substantial change or when a report suggests a possible breach. A new feature can change what a chatbot does even if it keeps the same name. The report calls for a way for users, associations and others to flag potential problems.
The authors also want regulators and approved researchers to have controlled access to data about how the products are used, trained and monetised. They suggest anonymised summaries, such as trends in the kinds of requests people make or the time they spend interacting. That would let researchers ask questions about use outside a test setting, without treating an increase in use by itself as evidence of harm.
What would need to happen next
The report is advice, not a law or a decision that any service has broken the law. It suggests that France could use the recommendations in discussions about changes to European AI rules. An assessment before launch and a review after a change serve different purposes: one looks for foreseeable problems, the other can use evidence from people actually using the service.
If the proposals advance, questions remain about what counts as a substantial change, who would carry out the assessments and how researchers could access useful evidence while protecting private conversations.
Sources
- Élysée: Delivery of the report on generative AI and vulnerabilitiesNovember 10 presentation, authorship and summary of the 18 recommendations.
- CIANum expert commission: Generative AI and vulnerabilities (PDF)Executive summary and recommendations 1, 3–6; tests before launch, ongoing reassessments, controlled evidence access and safeguards.
