ChatGPT and Local Authorities: AI Reshapes Local Perception

The Impact of Generative AIs on the Image of Local Authorities

Generative AIs are changing the perception of local authorities by providing information that bypasses the traditional filters of journalism or administration. A few months ago, a local elected official, out of curiosity rather than strategy, asked several generative AIs about a mixed-economy company in their department. The response was fluid, structured, and filled with figures, but largely outdated. The audit report it paraphrased was a legitimate public document, but it was several years old and omitted the legal follow-ups that had resulted in a dismissal.

This experience highlighted a major democratic issue: the lack of updates and contextualization of the information provided by AIs. Indeed, generative AIs produce statistical summaries of texts ingested at a specific moment, without distinguishing between recent information and that which is obsolete. This poses a significant problem when these summaries are used to form opinions about public institutions.

Information Without Human Hierarchy

Historically, the public image of institutions was shaped by professional journalism and public databases such as Légifrance or INSEE. These sources, although imperfect, benefited from a human hierarchy of information. Generative AIs, on the other hand, synthesize texts without temporal distinction, often confusing old reports with recent information. The CNIL itself acknowledges that these systems can generate inaccurate results that nonetheless appear plausible, a phenomenon known as hallucinations.

Generative AIs do not intrinsically distinguish a 2017 observation report from a press release published yesterday. They give the same weight to a critique made in an eighty-page document as to the detailed response provided by the audited organization — when that response is ingested, which is not always systematic. And when a dismissal occurs two years later, that information, lacking the same publicity, simply does not appear in the algorithmic summary. The errors of AIs, in turn, feed the models, thus reinforcing biases.

Concrete Example: Sempro of Plessis-Robinson

Two lawyers, Maîtres Walter Salamand and Corentin Congard, analyzed the case of Sempro of Plessis-Robinson. This mixed-economy company for development and construction in the Hauts-de-Seine municipality, created in 1985, is 80% owned by the community and registered with the RCS Nanterre. It was the subject of a report by the regional audit chamber of Île-de-France in February 2023. Sempro provided a detailed response, contesting several of the qualifications retained. The proceedings initiated before the litigation chamber of the Court of Accounts resulted, at the end of 2024, in a dismissal regarding all the individuals involved. However, generative AIs continue to disseminate erroneous information, sometimes incorrectly attributing judicial decisions. The lawyers found that nearly 90% of the decisions presented on certain platforms did not actually pertain to this company.

Legal actions are still pending before administrative courts concerning Sempro of Plessis-Robinson. This case illustrates how generative AIs can distort reality by omitting crucial information or misattributing decisions to incorrect entities.

A National Issue

This phenomenon is not isolated. France has approximately 900 local mixed-economy companies, and the Court of Accounts has recorded nearly 90 observation reports made public during the 2016-2018 period. AIs ingest these documents without always contextualizing them, which can affect the public perception of local institutions.

To cite two public examples among many others:

• SEMITAG, the transport company of the Grenoble agglomeration (which became SPL M'TAG in 2022), was the subject of a report by the CRC Auvergne-Rhône-Alpes in 2023, covering the management from 2015 to 2021.
• SEM Énergies 22, in the Côtes-d'Armor, had its report published in 2025 by the CRC Bretagne.

These public and accessible documents are precisely the type of dense textual material that models prioritize ingesting — and subsequently reproduce, often without being able to contextualize them.

Democratic Challenges and Reflection Paths

The issue goes beyond the mere reputation of local authorities. It touches on citizens' ability to form an informed opinion about the institutions they finance. To address these biases, it is suggested to strengthen the legal identification of public structures and regularly update public information sources with factual and dated data. Managing the digital image of institutions thus becomes an essential skill, requiring collaboration between legal experts, communicators, and digital reputation specialists.

This is not about blaming generative AI — its legitimate uses are numerous, and the tool is useful. It is about recognizing, from the ground, that a new intermediary has inserted itself between public institutions and citizens, and that this intermediary operates under rules that are not those we established to ensure the clarity of democratic debate.

What is at stake here is not just the reputation of a SEM or a municipality. It is the very readability of local public action. A democracy assumes that citizens can form a well-informed opinion about the institutions they finance. If this opinion is now formed, in the first instance, from an algorithmic summary that freezes the past, ignores counterarguments, and confuses homonyms, then something important shifts. Not abruptly, not spectacularly — but surely. Public memory ceases to be a debate and becomes a narrative produced by a software layer.

This is a serious issue. It deserves to be treated as such — without excess, but without complacency either.