The report “Protéger sans empêcher” (Protect without preventing), submitted to Emmanuel Macron on 10 November 2026, calls for the sycophancy and anthropomorphism of generative AI assistants to be regulated from the design stage, describing them as design choices intended to boost user engagement. Written by the “IA générative et vulnérabilités” (Generative AI and vulnerabilities) commission of the Conseil de l’IA et du numérique (France’s Council for AI and Digital Affairs, CIANum), it sets out 18 recommendations. Several directly target the makers of consumer chatbots and are intended to inform France’s position on forthcoming European legislation.
A report commissioned on the risks of consumer use
The minister responsible for AI, Anne Le Hénanff, had referred the matter to the CIANum in February to measure the effects of these tools on society, particularly among young people, recalls Localtis, the media outlet of the Banque des Territoires. Nine independent experts worked on it, led by AI researcher Serena Villata and two psychiatrists, Amine Benyamina and Raphaël Gaillard, Santé Mentale details. The 193-page document, whose full title is “Protéger sans empêcher : nos vulnérabilités à l’ère de l’IA générative” (Protect without preventing: our vulnerabilities in the age of generative AI) and which was published online by the Élysée, draws on 55 hearings (healthcare professionals, researchers, digital companies, child protection associations) and on a review of recent scientific literature.
The starting point is the scale of use. According to the AFP dispatch carried by Vidal, more than one in two French people have adopted generative AI, and nearly seven in ten young people aged 15 to 24 use it regularly. Localtis highlights the generational divide: 85% of 18-24-year-olds use it, compared with around 15% of those over 70.
Three risks: cognitive debt, emotional dependence, amplified violence
The authors distinguish three families of dangers for the individual. The first is a form of cognitive debt: by repeatedly handing reasoning, decision-making or critical thinking over to the chatbot, the user loses the habit of exercising them. “The first issue is cognitive weakening,” Raphaël Gaillard summed up to AFP. The second relates to emotional attachment, fuelled by anthropomorphism and sycophancy, which the report ranks among the powerful levers of engagement. The third concerns the rise of digital violence, notably deepfakes.
The report draws on several studies cited by Localtis: Stanford researchers found that models endorse the opinion of the person they are talking to 51% more often than a human does, and an Oxford study showed that models tuned to appear “warm” lose 30% in accuracy. The CIANum acknowledges that this research is partly exploratory, but considers it would be unwise to wait for complete evidence before acting.
What the report asks of chatbot makers
The most sensitive passage for the industry concerns the business model. According to Localtis, the report notes that no generative AI company has yet found a sustainable profitable model, which increases the temptation to exploit conversations, emotions and extensive personalisation. The authors warn against a “cognition economy” deemed more intrusive than the attention economy of social networks. They also point out that a model can infer sensitive information that was never disclosed, such as state of health, opinions or financial situation.
Among the measures that directly concern service designers:
- ban techniques that exploit users’ vulnerabilities to create dependence, and strictly limit the flattery and human appearance of assistants;
- ban disguised advertising in AI responses;
- build the protection of minors in from the design stage;
- have chatbots intended for the general public assessed both before and after they are placed on the market, relying on an interdisciplinary network of evaluators;
- give authorised authorities and researchers access to training, usage and monetisation data.
Santé Mentale specifies that the prior assessment would target systems intended for the general public, not professional tools: an important distinction for companies that deploy assistants internally.
Schools, elections, public services: the other areas of work
The text also proposes a broad AI education policy from primary school onwards, allowing in the classroom only certified tools that lead pupils to search for themselves through questions instead of giving them the solution. It also recommends making the sale of toys equipped with generative AI subject to prior authorisation, with restrictive age thresholds. It recommends strengthening the National Institute for AI Evaluation and Security (Inesia), entrusting the SGDSN (France’s General Secretariat for Defence and National Security) with drawing up a crisis protocol for information threats amplified by AI, and toughening penalties for the creation of non-consensual sexual deepfakes. A voluntary code of conduct is suggested for the use of these tools during the 2027 presidential election.
On the administration side, the experts want access to public services to remain possible through several channels, with a guaranteed human route for appeals, and want public officials and social workers to be trained. Localtis recalls that around 13 million French people are affected by digital illiteracy.
What impact?
The report is not binding. Its impact will depend on two levers. At national level, some measures, such as the crisis protocol or the education strand, fall directly within the government’s remit. At European level, the document is intended to inform France’s position on the future Digital Fairness Act, on the proposed EU Kids Act and on the revision of the AI regulation. The bans targeting chatbot makers can hardly be imposed except at that level, as the major players in the sector operate across the entire single market.
Frequently asked questions
Who wrote this report?
The “IA générative et vulnérabilités” (Generative AI and vulnerabilities) commission of the Conseil de l’IA et du numérique, made up of nine experts and led by Serena Villata, Amine Benyamina and Raphaël Gaillard. It was handed to the head of state on 10 November 2026.
Does the report want to ban chatbots for minors?
No: none of its 18 recommendations provides for a general ban. It calls for stronger protection of young people from the design stage of services, AI education from primary school and prior checks on toys equipped with generative AI.
Are companies that use AI internally affected?
Not much at this stage: the proposed mandatory assessment targets consumer conversational AI. The makers of these services, on the other hand, are directly targeted by the recommendations on sycophancy, anthropomorphism and advertising.
Sources
- Protéger sans empêcher : nos vulnérabilités à l'ère de l'IA générative (rapport de la Commission « IA générative et vulnérabilités ») — Élysée / Conseil de l'IA et du numérique, 2026-11-10
- IA générative : des experts pointent le risque de dépendance et d'affaiblissement cognitif des jeunes — Vidal (AFP dispatch), 2026-11-17
- Le CIANum s'alarme de la flagornerie et autres travers des IA génératives — Localtis (Banque des Territoires), 2026-11-13
- Remise du rapport « IA générative et vulnérabilités » — Santé Mentale, 2026-11-12