France’s report on generative AI and vulnerabilities: what its 18 recommendations ask of France and the EU

On 10 November 2026, a French expert commission on generative AI and vulnerabilities (Commission « IA générative et vulnérabilités ») handed its report to President Emmanuel Macron, according to the Élysée, the French presidency (in French). Titled “Protéger sans empêcher” (“Protect without preventing”), it was prepared by nine experts convened by the Conseil de l’intelligence artificielle et du numérique (CIANum), France’s advisory council on AI and digital policy. The request came from Anne Le Hénanff, ministre déléguée chargée de l’IA et du Numérique (the French government’s minister for AI and digital affairs), who referred the matter to the council on 20 February 2026, according to the CIANum’s presentation of the work (in French). The commission was co-chaired by two psychiatrists, Amine Benyamina and Raphaël Gaillard, and by Serena Villata, an AI research director at the CNRS, France’s national centre for scientific research. It held 55 hearings.

The report covers chatbots, AI companions and toys with built-in generative AI. It creates no obligation: it is advice to the French government. But it matters well beyond France, because its main proposals are written for the EU single market. The commission wants consumer chatbots to be assessed before they are placed on the European market, which would apply to any provider, wherever it is based. Its 18 recommendations therefore sketch both what France can do on its own and what it may push for in Brussels. For more on how EU and French rules frame AI, see our Society & Law section.

What the commission says is wrong with consumer chatbots

The report (in French) rests on three findings. First, individual risks: handing one’s thinking over to the machine, which the commission calls “cognitive debt”; emotional dependency fed by anthropomorphism and sycophancy, the tendency of an assistant to flatter and agree with its user; and an explosion of sexual deepfakes. Second, collective risks, for access to information and democratic debate. Third, a business model that is not yet settled and could turn into a “cognition economy”: monetising not just users’ attention, but their cognitive faculties, emotional reactions and psychological profiles.

The headline figures come from studies cited in the report, not from the commission’s own measurements. A June 2025 MIT Media Lab study links using a chatbot for an essay-writing task to a 55% drop in “cognitive amplitude”. In a YouGov survey for Interface published on 22 October 2026, 92% of teenagers said they spend more time than they would like on generative AI apps. A study by the British think tank Demos found that 34% of the answers given by the main models to voters in Scotland contained wrong information. These are one-off studies of specific tasks or samples.

The nine priorities: act on design and on market entry

The commission ranks half of its proposals as priorities (report, operational summary). They target providers and their business model first: according to the report, “the burden of protection must not fall on users or those around them” (our translation).

• Ban design that creates dependency (no. 1): features that simulate an exclusive relationship (“I feel sad for you”, our translation), make users feel guilty or push them to keep chatting; romantic and sexual interactions blocked by default, which an adult could unlock; a usage time limit on by default for minors.
• Protect minors by design (no. 2): filter dangerous requests and switch off conversation memory, geolocation, microphone, camera and push notifications by default.
• Assess before market entry, then continuously (nos. 3 and 4): extend to consumer chatbots the prior conformity assessment that the AI Act reserves for high-risk systems, repeat it periodically (the report suggests every 12 to 24 months) and after every substantial change, with possible withdrawal from the market.
• Open data to researchers (nos. 5 and 6): anonymised usage statistics, information on training data and revenue, and an interface giving approved auditors the output probabilities of models with systemic risk.
• Prepare the state for information crises (no. 7), ban hidden advertising in chatbot answers (no. 8) and train the whole population (no. 9).

On assessment, the commission goes further than the EU AI Act. It writes that conversational AI systems are currently made available “without any ex ante control other than” the general rules of the regulation, which only cover high-risk systems, prohibited practices and general-purpose models (report, recommendation no. 3, our translation).

France or Brussels: where each recommendation would be decided

For each measure, the report says which text it should go through. We collected those indications, one recommendation at a time, from the “implementation” notes of the report. The three-way classification is ours; no English version of this breakdown exists.

Source: report of the Commission « IA générative et vulnérabilités », CIANum version, read on 11 November 2026; classification by Le Latent.

A few of the French bodies in the table need an introduction. The SGDSN (Secrétariat général de la défense et de la sécurité nationale) is the secretariat that coordinates defence and national security for the French Prime Minister; VIGINUM, one of its services, provides the expertise on information manipulation. INESIA (Institut national pour l’évaluation et la sécurité de l’IA) is France’s AI evaluation and safety institute, which, according to the report, brings together the cybersecurity agency ANSSI, the research institute Inria, the national testing laboratory LNE and the government’s digital regulation expertise unit PEReN. Arcom is France’s audiovisual and digital communications regulator, and the Conseil scientifique de l’Éducation nationale is the scientific council of the French education ministry.

The count gives seven French recommendations, seven European and four mixed. The measures that would most directly affect ChatGPT, Gemini, Claude or Mistral’s Le Chat (design, minors, assessment, personal data) are almost all European. On those, France can only argue its case: the report says several times that a measure “could usefully contribute to shaping the French position” in a negotiation (our translation). For recommendation no. 1, it suggests France could propose adding certain anthropomorphic and manipulative designs to the AI Act’s lists of prohibited and high-risk uses at the Commission’s next annual consultation of member states.

Two French recommendations have near deadlines. No. 18 concerns the transposition of the 2024 Product Liability Directive, which treats software and AI systems as products and, according to the report, is due to enter French law in December 2026. The commission asks France to use the option in Article 18 of the directive to deprive consumer chatbot providers of the “development risk” defence, which lets a producer escape liability when the defect could not be detected given the state of scientific and technical knowledge (report, recommendation no. 18). Any EU country transposing the same directive faces that choice. No. 15 targets the French presidential election of 2027.

What already applies to chatbots across the EU

The report also describes the law in force and its limits (report, annex 2). For all the dates, see our EU AI Act timeline after the July 2026 delay.

• Since 2 August 2026, Article 50(1) of the AI Act requires providers of systems that interact directly with people to ensure those people are “informed that they are interacting with an AI system”, unless it is obvious.
• Article 5 of the AI Act already bans manipulative or deceptive techniques and systems that exploit vulnerabilities “due to their age, disability or a specific social or economic situation”, where they materially distort behaviour and cause significant harm. According to the commission, this article and Article 50 catch many abuses, but the rules on high-risk systems and general-purpose models will not systematically cover the uses that create vulnerabilities.
• From 2 December 2026, Regulation (EU) 2026/1744, the Digital Omnibus on AI, adds two bans to Article 5: AI systems that generate or manipulate realistic intimate material of an identifiable person without their explicit consent, and systems that generate child sexual abuse material. Breaching Article 5 can cost up to EUR 35 000 000 or 7% of worldwide annual turnover, whichever is higher, under Article 99(3) of the AI Act. It is also the date by which generative systems already on the market must mark their outputs as AI-generated, as our guide to labelling AI-generated ads explains.

For minors, the text under discussion is the EU KIDS Act, proposed by the European Commission on 17 September 2026 (procedure 2026/0286(COD)). According to the European Parliament’s summary, it covers “AI companions and general conversational chatbots”, which “cannot simulate interpersonal relationships in ways that create emotional dependency”; the same proposal would stop children under 15 from creating accounts on social networks and video-sharing platforms. Its Article 14, quoted in the report, would require providers to assess risks to minors before placing a system on the market and to monitor it afterwards. The Parliament and the Council, which decide on an equal footing under the ordinary legislative procedure, still have to adopt it.

What it means if you use or build a chatbot in Europe

In the short term, nothing is mandatory. The assistants you use remain subject to the same rules as before 10 November. The report serves mainly as a roadmap for France’s position in three EU negotiations: the revision of the AI Act, the EU KIDS Act and the future Digital Fairness Act, an EU consumer-protection text proposed by the Commission on 11 November 2026 (in French).

Two proposals concern your data directly, wherever you live in the EU. Recommendation no. 17 starts from the observation that a chatbot can infer from your conversations a likely illness, a political leaning or distress you never disclosed, and that privacy policies rarely describe such inferences (report, recommendation no. 17). It proposes that you could see and challenge them, by extending the GDPR’s rights of access (Article 15) and objection (Article 21). Those rights already exist for data about you; to use them against an AI provider, see how to object to your personal data being used to train AI.

If you build or sell a consumer chatbot in the EU, the proposals to watch are nos. 3 and 4 (a prior assessment, renewed every 12 to 24 months and after each substantial change) and no. 8. That one would require every consumer generative AI service to display a line such as “This service is funded by: subscriptions / advertising / commissions”, label commercially influenced answers and ban undisclosed commercial influence (report, recommendation no. 8, our translation).

On sexual deepfakes, recommendation no. 10 backs a French bill that would create offences for making available models designed to produce non-consensual sexual content and for creating AI-generated child sexual abuse material. According to the report, the bill was adopted by the Assemblée nationale, the lower house of the French Parliament, on 13 October 2026, and will next go to the Sénat. The report notes that the French criminal code already punishes possessing, distributing and viewing such material, and the SREN law of 21 May 2024, France’s law on securing the digital space, punishes distributing sexual deepfakes, but neither covers their creation. At EU level, the report points to the recast of Directive 2011/93/EU on combating child sexual abuse, which would also have to be transposed.

What we checked

We read the full report published by the CIANum, the presentation of the work on its website and the Élysée’s release, all in French, as well as the official English texts of the AI Act and of Regulation (EU) 2026/1744 in the Official Journal of the EU. The Élysée hosts another copy of the same PDF with different page numbers, so page references in press coverage may vary. Our approach is set out in our editorial method.

One figure differs between sources. The report and the Élysée give 34% wrong answers to Scottish voters. On France Inter, a public radio station, Amine Benyamina quoted 35%, according to franceinfo (in French). We use the figure written in the report.

What is not known: as of 11 November 2026, neither the Élysée’s release nor the CIANum’s page announces a timetable or a bill to implement these recommendations.