First birthday celebrated for the National AI and Digital Council (CIANum).
On July 23, 2025, the Government established this new incarnation of the National Digital Council (CNNum). Since then, it has produced, among other things, two reports on digital sovereignty and AI in the education system. It has also published 5 notes.
Don’t Say “AI Agents,” but “Agentic AI”
The first of these notes, dated February 2026, deals with agentic AI.
The CIANum suggests favoring this term over “AI agents.” This avoids, it argues, reducing the discussion to a binary opposition between automation and human control.
The note distinguishes 5 degrees of autonomy for agentic AIs:
- Simple automation based on the reproduction of rules
- Agents capable of extracting data or analyzing documents
- Agentic workflow chains (a sequence of steps in processes alternating task automation and human validation)
- Flows of semi-autonomous agents (responsive to context and able to prepare and then make decisions)
- Fully autonomous agents
The CIANum synthesizes technological advances, from longer context windows to post-training methods. It revisits technical limits (cascading effects, orchestration drift, semantic misalignment, data security, etc.) and ethical, societal and legal issues (responsibilities, algorithmic biases, transparency and intelligibility of AI…).
A portion of the note is devoted to the replacement of human labor. The CIANum notes, in particular, the slow penetration of AI into companies when the share of automated tasks represents less than 20% of a job’s tasks.
A reaction to the launch of Claude Mythos…
In April, a note on AI & cybersecurity followed. Anthropic had just launched Claude Mythos, then opened to around forty organizations that were almost exclusively American.
Projecting an acceleration in software vulnerability remediation, the CIANum warned about the potential inability of information systems to keep pace with patching. It also urged caution about the tendency of vendors to tout the “dangerousness” of their models if they were to fall into the wrong hands. A “clever” way to generate public interest—and investors, in the view of the notes. And it referenced, in particular, the “morality of AI” summit Anthropic had organized with about fifteen members of the Catholic clergy.
The note also highlighted, among other things, the need to strengthen the capabilities to evaluate models. It underscored France’s significant information asymmetry compared with the British, American and Chinese ecosystems.
… then to the suspension of access
In June 2026, another note followed after the unilateral, prior-notice suspension of access to Fable 5 and Mythos 5 for non-Americans.
The CIANum saw this as an “additional, significant, and unprecedented step” toward the generalized practical risk of a kill switch, extending the containment strategy hitherto embodied mainly by NVIDIA chip export restrictions.
The threat to European digital autonomy became a tangible reality, the CIANum summarized, urging everyone to “conduct and continuously update their risk analysis.” Echoing its previous note, it recalled the grammar of catastrophic or existential risk often deployed by the GAFAM to justify the idea that only a handful of players should possess the most advanced technologies.
Against this backdrop, the CIANum called for accelerating efforts on the European preference in public procurement. It also called for market-shaping measures (portability, interoperability, tax or regulatory incentives, etc.).
A note on frictions between GDPR and agentic AI
The latest note (July) addresses the application of the GDPR to agentic AI.
Because of its operation, this technology pushes against several principles set out by the regulation. Among them:
- Purpose (the general-purpose capabilities of agentic AI systems can complicate identifying the processing)
- Lawfulness (the autonomy of these systems can complicate linking the initial legal basis to subsequent operations)
- Data minimization (with large data flows and sharing between agents, the necessity of processing is not always easy to demonstrate)
- Accuracy (the probabilistic nature of models implies a risk of error that can propagate)
- Transparency (outputs can be hard to explain)
- Storage limitation (tensions with retaining data for training and model improvement)
Consequently, difficulties arise in effectively exercising data subjects’ rights (access, rectification, erasure, objection, restriction of processing). Agentic AI amplifies the challenge due to the multiplication of data streams and memory spaces, whose synchronization can undermine data accuracy.
A call to better compartmentalize agentic systems
The CIANum also highlights the difficulty of gauging the true degree of human supervision in the decision-making process of these systems. Yet, under the GDPR, any individual has the right not to be subjected to automated individual decisions.
Opacity also complicates assigning responsibilities. In this area, the AI Act does not impose a specific regime. But, through its approach that distinguishes actors along the value chain, it introduces a form of accountability.
The ongoing circulation of data more broadly raises fault lines from the risk of leakage and manipulation to uncontrolled reuse.
The European Commission and the European Data Protection Board (EDPB) are preparing guidelines for this year on how to align GDPR and the AI Act.
From the CIANum’s perspective, there is a push to specify the governance of automated decisions and to consider measures to detect or filter uses not foreseen or misused. It also advocates partitioning the memory of systems (one memory per agent, with no automatic access to other agents’ data). It recommends limiting their size, for example by letting data expire after a set period or by routinely replacing older data with newer items. Memory could even be segmented by processing to curb excessive data storage.
Protection of minors: targeting specific platforms displeases the CIANum
In the midst of these AI-centered notes, the CIANum published one dedicated to protecting minors online. This was in March 2026, against the backdrop of a proposed law whose main effect would be to ban social networks for anyone under 15. The National Assembly had just adopted it in its first reading.
Australia had been the first country to adopt such a ban, with a lower limit of 16 years. Effective in December 2025, it targets Facebook, Instagram, Kick, Reddit, Snapchat, Threads, TikTok, Twitch, X and YouTube.
The CIANum is wary of such an approach that targets specific platforms. It fears it would push activity to more obscure platforms. It cites Roblox as an example: not covered by the law, yet there are dozens of groups with more than 100,000 members exchanging pedophilic content and soliciting sexual acts from minors.
In March, Australia refined its approach through online safety codes. They target content (highly graphic violence, pornography, suicide promotion, etc.) and apply to a broad range of services (chatbots, app stores, online games, search engines…). The aim is to ensure children have age-appropriate online experiences.
France adopts a more restrictive stance. Its law targets only “online social network services provided by an online platform.” The CIANum regrets that this approach regrets the opportunity to regulate other platforms whose content can be harmful. Especially as uses evolve and move elsewhere.
Age verification: perspectives around the European digital wallet
For age verification, the European Commission distinguishes three methods: self-declaration, estimation (through facial recognition or cross-referencing usage data, among others), and the provision of proofs (physical identifiers or verified identity sources).
Australia did not impose specific methods on platforms. It did, however, express concerns, notably regarding Snapchat, which may not verify that facial recognition occurs on real images.
In France, the SREN law (May 2024) also leaves the choice to services that distribute pornographic content. They must nonetheless comply with the technical requirements of the associated framework (published October 2024).
Germany, Denmark, Spain and Italy have adopted similar obligations. The European Commission’s guidelines supplementing the DSA secure this approach. Discussions are under way about the European digital wallet, with each member state required to propose at least one by 2026. A pilot has been launched in five Member States – including France – to verify age for access to social networks.
The CIANum, not enthusiastic about “disconnection windows”
Implementing a single, uniform rule can raise questions, the CIANum explains: protection needs vary by age and maturity. Some platforms advocate device-level controls at account creation in an app store or via the OS. For the CIANum, this option is too broad given the relatively small number of services ultimately affected. It would also push service developers to sidestep their obligations and would be ill-suited to shared digital environments, such as within a household.
As for imposing disconnection time windows, the CIANum replies that time spent online does not reveal practices or contents viewed. It notes that South Korea implemented this approach from 2011 to 2021 and rolled it back due to its narrow scope (online gaming).
The idea of a global right to configure interfaces
The CIANum raises another concern: targeted measures for minors could unshackle digital spaces for other users. It thus calls for a global right to configure interfaces. This would include strengthened filtering options, including via third-party algorithms, alongside externally moderated services that are labeled and controlled.
The DSA requires the provision of information on moderation tools and the main parameters of recommendation algorithms. But, according to the CIANum, these are technical reports that are difficult for the general public to interpret, and do not allow a full grasp of the economic and attention-driven logics that structure the systems.
The CIANum also revisits the idea of rethinking the host vs. publisher dichotomy. Under the E-Commerce Directive, social networks enjoy a lighter regime: they are liable for user-published content only if they have actual knowledge or fail to act promptly to remove it. As early as 2014, the Conseil d’État urged moving beyond this dichotomy to establish a dedicated status. The DSA does not go that far but does distinguish among hosts, imposing a stricter regime on platforms. The 2027 reform could be a timely moment to shift toward a responsibility-based approach, the CIANum argues.
Sovereign digital: beware the competition between territories
The sovereignty report appeared in May. The CIANum notes a “persistent gap” between state rhetoric and practice, with public services still largely relying on foreign proprietary solutions.
At the interdepartmental level, the DINUM lacks resources and coercive power to enforce a unified approach. Work remains siloed among public enterprises and local authorities. Furthermore, the CIANum contends that the State has fostered a competition between territories (to access public funding), pushing each to specialise in a particular digital niche.
Nevertheless, the notes acknowledge a growing willingness to coordinate with the private sector. It cites FranceConnect and ProConnect as “models of governance and public-private cooperation.” It also notes DINUM’s involvement in the Open Interop initiative, driven by the strategic committee of the trusted software and solutions sector.
Anchoring public funding for the long term
Public funding for digital commons and open-source software remains insufficient, including at the European level, due to a lack of criteria aligned with their collaborative model.
Another observation: funding does not always have a durable horizon. The CIANum points to the Programme for Investments in the Future (PIA), which funded many projects over 18 to 36 months. Other initiatives have led to a scattered approach, such as ALT-EDIC and the LLM4EU consortium (€40 million spread across 70 participants).
Underfunding may partly stem from confusion between “free” and “gratis.” More broadly, it fuels expectations about service levels that are not realistically achievable.
A “Factory of digital commons” to govern demand
To translate governance of demand into reality, the CIANum calls for creating a “Factory of digital commons”. It could bring together the State, the Caisse des dépôts, associations of local authorities, the DSI of large enterprises, and representatives of solution providers and maintainer communities.
The Ariane (Agency for Digital and AI) would oversee the implementation of this structure, which would:
- Consolidate dependency diagnoses
- Develop a national map of the existing landscape, following the approach of the Interministerial Open-Source Software Foundation
- Grant a label to off-the-shelf solutions that meet governance, interoperability and durability criteria
- Trigger targeted funding (through its partners, through calls for commons or through a dedicated digital commons fund)
- Contribute to the active governance of open-source communities
- Make visible the legal structures suited to digital commons
SCOP, SIEG, SCIC… Legal vehicles mobilizable for digital commons
Among the mobilizable vehicles:
- SCOPs (cooperative and participatory companies) — They foster lasting attachment to the project’s founding values, independent of any majority shareholder.
- SCICs (collective-interest cooperatives) — Multipartite governance, allowing public and private actors while preserving the aim of limited profit.
- SIEGs (services of general economic interest) — Enable durable financing of the maintenance of recognized digital commons deemed to be of public interest, without the heavy procedures of traditional public procurement.
- AAC (calls for commons) — Identify general-interest challenges and fund a consortium committed to contributing to a shared public good.
- Factories of commons — Incubators without a formal legal structure that issue “calls for ideas” and select projects for an intrapreneurship phase and partnerships.
The CIANum also suggests studying the transposition into French law of social-use trusts (fiducies d’utilité sociale). Developed in Quebec, they allow organizations to share data in a trusted environment under collective governance.
At the European level, a possible lever would be to broaden the mandate of the DC-EDIC consortium, with missions resembling those that the Factory of digital commons would undertake in France.
Within the National Education system, innovative uses of AI remain in the shadows
The report on AI in education was published in June.
The CIANum highlights France’s efforts, including a framework for AI usage in education “one of the most complete internationally in terms of understanding what language models are, their effects and their legal framework.” The “last mile” remains to be accomplished. The evidence : the number of teachers trained is “particularly low” compared with other European countries. In the absence of a clear framework, those who innovate do so in the shadows: their uses are not well documented.
AI-powered tools can enable personalization of learning paths and better inclusion of students. They can also provide a means to address societal issues. The CIANum cites the Internet Sans Crainte program as proof. It runs from fourth grade through high school, with AI-ackathons where students imagine practical AI-integrated solutions to contemporary challenges.
On personalization, the CIANum mentions the Lailo platform, aimed at CP, CE1 and CE2 teachers, which helps differentiate the reading instruction based on each pupil’s progress.
Facing the risk of “cognitive resignation,” support “Socratic” designs
A downside: AI-based tools pose the risk of cognitive resignation. The CIANum cites a MIT Media Lab EEG study showing that participants who used an LLM to revise performed noticeably worse than those who used other methods. It thus calls for favoring a Socratic design that exercises critical thinking. For example, Khanmigo (not distributed in France) never gives the answer directly but guides the student with successive questions. The CIANum proposes issuing a call for proposals to select such an app for middle school and one for high school by the end of 2026.
Another risk is student isolation. This touches in particular on the so-called sycophantic bias: how the AI tells users what they want to hear. Oxford and Stanford findings are mobilized: exposed to a complacent AI, humans gradually lose the preference for human contact and emotional support and self-esteem.
Similarly, extreme individualization could erode consideration for others. A student might lose the “social patience” needed to account for others’ emotional and cognitive states.
From oral exams to collaborative projects, rethinking assessment methods
The CIANum warns that the accessibility of tools like ChatGPT creates an illusion of simplicity, while these technologies were not designed for educational use. There is also an illusion of zero cost, since the price of high-quality language models is likely to rise over the years.
The spread of AI also requires rethinking assessment models, including placing greater emphasis on processes rather than outcomes, and giving more weight to oral verification and collaborative projects.
The CIANum further encourages allowing “times without AI” to ensure that essential knowledge is consolidated. “A student who can do well without AI and also knows when and how to use it will be best prepared for the society to come,” it states.
From the “Digital Referent” to the “AI Referent”—even if the role may change
Operationalizing this educational principle will depend on teachers’ ability to adapt AI usage to pedagogical objectives. Hence the recommendation to launch a broad continuing education program. An immediate step would be to insert a module into the LPE (Bachelor in Education) and the M2E (Master of Education). On the online side, the CIANum suggests building on resources like the MOOC “AI for and by Teachers”, designed by Inria. It also proposes naming, in each secondary school, 1 or 2 AI referents. This may entail rethinking the role of the digital referent, potentially changing the person in charge.
Regarding the students, the Prime Minister has promised that starting in the 2027 school year, in the second year of high school they would have one hour per week of AI instruction. The CIANum urges going further to include a structured program of digital literacy, AI, media and information education, akin to sex education. It also recommends training school doctors on mental health and cognitive risks, and organizing, from the 2027 school year, workshops under their responsibility.
