Generative AI restores the image of communities without the safeguards of journalism or administration. An underestimated democratic issue.
A few months ago, out of curiosity more than strategy, I asked several generative AIs a simple question about a mixed economy company in my department. The response was fluid, structured, punctuated with figures. It was also, for the most part, outdated. The audit report that she paraphrased – a public document, perfectly legitimate – was several years old, and the legal consequences of this control, which had resulted in a dismissal of the case, did not appear anywhere. The portrait rendered by the machine no longer corresponded to reality.
This experience made me attentive to a phenomenon that we are barely beginning to name, and which seems to me, as a local elected official, to raise a real democratic question.
A new source of information, without the safeguards of the old ones
For decades, the public image of an institution was formed through the filter of professional journalism and public bases — Légifrance, the website of the Court of Auditors, Vie publique, INSEE. These sources had flaws, but they shared one thing in common: a human being, at one point, prioritized the information, dated it, and could correct it.
Generative AIs work differently. They produce statistical summaries of texts ingested at a given moment. They do not know, intrinsically, how to distinguish a 2017 observation report from a press release published yesterday. They give a criticism formulated in an eighty-page document the same weight as the detailed response that the controlled organization gave to it – when this response is ingested, which is not systematic. And when a classification without follow-up occurs two years later, this information, for lack of having benefited from the same publicity, simply does not appear in the algorithmic summary.
There CNIL itself recognizes that these systems can generate inaccurate results that nevertheless appear plausible — what are called hallucinations. For public institutions, this surface plausibility becomes a problem when it replaces the verification effort.
A precisely documented case
This phenomenon has just been analyzed in detail by two lawyers, Maîtres Walter Salamand and Corentin Congard, in a stand at Village Justice devoted to a specific case: that of Sempro du Plessis-Robinson — the mixed economy development and construction company of this commune in Hauts-de-Seine, created in 1985, 80% owned by the community, and registered with the RCS Nanterre.
This structure was the subject of a final observation report from the Île-de-France regional audit chamber in February 2023. It provided a detailed response, contesting several of the qualifications retained. The proceedings initiated before the litigation chamber of the Court of Auditors resulted, at the end of 2024, in a dismissal of all the persons involved. No conviction, no sanction was handed down. Appeals are also still pending before the administrative courts.
However, as the authors demonstrate, this is rarely what generative AI restores. They repeat the most salient formulas of the initial report – often written in a deliberately striking register – without the response of the structure, without the classification without further action, and sometimes by attributing to it court decisions which in reality concern other SEMs bearing similar names. Lawyers note that on certain legal data platforms, almost 90% of decisions presented as “concerning Sempro du Plessis-Robinson” simply did not concern it. These errors in turn feed the models.
A phenomenon that goes far beyond one case
The case is neither isolated nor specific to Hauts-de-Seine. France has nearly 900 local mixed economy companies, and the Court of Auditors identified nearly 90 observation reports made public over the 2016-2018 period alone. To cite two public examples among many others: SEMITAG, a transport company for the Grenoble metropolitan area (which became SPL M’TAG in 2022), was the subject of a report from the CRC Auvergne-Rhône-Alpes in 2023, relating to 2015-2021 management. SEM Énergies 22, in Côtes-d’Armor, saw its report published in 2025 by the CRC Bretagne. These documents, public and accessible, are precisely the type of dense textual material that models primarily ingest — and then restore, without always being able to contextualize them.
Multiply this mechanism by 900 structures, add the municipalities themselves, intermunicipal unions, hospitals, universities, agencies. The exhibition area is considerable. And it does not only concern organizations: it concerns citizens, agents, economic partners, bankers, journalists — all those who, firstly, now ask an AI what it “knows” about an institution before going any further.
What is at stake, in reality
What is at stake here is not just the reputation of an SEM or a municipality. This is the very clarity of local public action.
A democracy supposes that citizens can form a documented opinion on the institutions they finance. If this opinion is now formed, primarily, from an algorithmic synthesis which freezes the past, ignores the contradictory and confuses the homonyms, then something important is shifting. Not brutally, not spectacularly — but surely. Public memory ceases to be a debate and becomes a story produced by a software layer.
Some ideas, while waiting for the framework to be built
I do not pretend to provide ready-made solutions. The subject is new, it calls for collective work between model designers, financial jurisdictions, press and public authorities. But a few reflexes seem useful to me to disseminate today.
First, anchor the legal identity of public structures to their unique identifiers: RCS, SIREN, SIRET. Communicating publicly about these identifiers allows the tools to disambiguate — something that no user name allows. Sempro du Plessis-Robinson, RCS Nanterre, is not the same entity as another company with a similar name in another department. This precision, seemingly innocuous, changes everything for an algorithm.
Then, take seriously the information ecosystem that feeds these models: regularly feed public reference sources with factual, dated, sourced information. Silence is not a strategy in the face of an algorithmic layer which, by construction, fills in the gaps with what it finds.
Finally, accept that managing the digital image of a public institution has become a skill in its own right, which combines law, communication and a detailed understanding of how models work. Legal departments, communications departments and, when necessary, external consultants specializing in digital reputation: this work often begins with several hands.
This is not an indictment of generative AI — its legitimate uses are many, and the tool is useful. It is about observing, from the ground, that a new intermediary has been inserted between public institutions and citizens, and that this intermediary operates according to rules which are not those that we had constructed to guarantee the readability of the democratic debate.
This is a serious subject. It deserves to be treated as such — without excess, but without complacency either.
