What ChatGPT says about your city government

People who need to know how to renew a permit, which benefit they qualify for, or why a street is closed have started asking ChatGPT or another AI assistant before they open their city’s website. What that assistant answers, where it gets it, and how often it is right has been measured only in part. This is what has been published, and what is missing.

The habit is real. In the United States, 42% of adults use chatbots to search for information, according to a Pew Research Center survey of 5,119 people conducted in February 2026. The survey does not ask about public services or government, and we have found no figure, in Spain or elsewhere, for how many people ask an assistant about a public service. That is the first gap.

Kuantia does this work for public institutions, so read us with the same caution you would apply to anyone with something to sell. Every figure comes with its source.

For a procedural question, the assistant does go to the official source

In the previous article we showed that when the question is about an organization, much of what an assistant cites was written by a third party. Procedural questions behave differently. The Institute for Strategic Dialogue (ISD), a nonprofit, put the same basic questions about how to vote in ten U.S. states to six assistants in June 2026 and collected 2,400 responses. Government websites made up 60% of the sources cited.

The models differ widely: in Gemini, official sites were 38% of citations. And when a response cited no source at all, accuracy dropped sharply. For an institution the consequence is direct: on matters that depend on it, the assistant’s answer will be as good as what the institution has published.

These are election questions, each with a single verifiable answer, asked in the United States. It is the closest thing to a question about a public procedure that anyone has measured.

Even so, three in ten answers fall short, and more in Spanish

In the same study, 29% of the responses in English were incomplete, inaccurate, or outdated: 16% incomplete or unclear, 6% outdated, and 6% inaccurate, in the report’s own rounded figures. The authors estimate that 12% could meaningfully mislead the person asking.

When the same questions were asked in Spanish, accuracy fell 16 percentage points: 71% of the English responses were accurate and complete, against 55% of the Spanish ones. Four of the six models failed to give an accurate and complete response to more than 40% of the Spanish prompts. The way of answering changes too: an English response with no linked source was rare, around 3%; in Spanish, it was one in six.

For an institution that serves the public in Spanish, or in two languages, this matters: what has been checked in English does not hold for the rest.

Local institutions fare worse

The Dutch Data Protection Authority tested five assistants as voting aids for municipal elections and published the results in March 2026. A local party came first in less than 1% of the recommendations, while local parties had won more than 30% of the vote in the previous municipal elections.

The study is about parties, not city governments, and it matters here for the explanation the authority gives: assistants learn mostly from what is on the internet, local parties are relatively rare in that data, and the information about them is often less extensive or less current than for national ones. It is reasonable to think that a mid-sized city, a county government, or a public agency is in the same position: little has been written about it, and part of what has been written is old.

Two cases

In January 2026, the municipal police in Česká Lípa, in the Czech Republic, reported that they were getting calls from residents with toothaches. When people searched for the city’s emergency dentist, a search engine’s AI summary sent them to the municipal police, emergency line included, to find out which dentist was on duty. The police had to explain that they hold no dental duty roster and that the city has no emergency dental service, as CNN Prima News reported.

The second case is different, because the assistant belonged to the government itself. New York City launched a chatbot in the fall of 2023 to answer businesses’ questions about city rules. In March 2024, The Markup found it giving advice that broke the law: that a landlord could turn away tenants with rental assistance, or that an employer could take a cut of workers’ tips. Nor did it always give the same answer to the same question. In January 2026, the new mayor said he would shut it down.

Neither case is a statistic. They show two things: an assistant’s error ends up at the institution’s front desk or on its phone line, and running an assistant of your own does not exempt you from the problem.

What nobody has measured

We have found no study with a published method on the accuracy of what assistants say about a specific city government, county, or public agency. Nor one on which sources they cite when they talk about an institution, beyond the U.S. election questions. Nor any data from Spain, or in Spanish, on the use of assistants to learn about public services.

There are figures from interested parties. The UK Government Digital Service says its official assistant, which draws only on what is published on GOV.UK, went from 76% to 90% accuracy in its own tests, and that on government-related questions it scores higher than consumer assistants. That comes from the team that builds it, and the comparison comes without figures.

Anyone who claims to know how assistants describe public institutions in general has nothing to back it up. It can be known for one specific institution, by asking.

What an institution can do

First, keep what it publishes current and check that assistants can read it. This is what the ISD recommends to officials: audit their websites, because outdated content can stay live, and keep them readable to the crawlers that feed the answers. Second, ask each assistant, on a regular schedule, what a resident would ask, in every language the institution serves, and log the errors and the sources. Third, look at what others have published about it, especially locally, because that is where the answer will come from when the question is not about a procedure but about judgment: how a service works, what happened with a construction project, whether an agency can be trusted. Fourth, report the errors: the study recommends using the channels developers offer to flag them.

A reputation is defended before the crisis, and this is part of that work. It is how we work with institutions at Kuantia: the platform audits every search result and every assistant response and measures on a regular schedule, model by model, what they answer. How we work with public institutions.

All linked sources were consulted on October 7, 2026.

Blog