A city council announces that it will use an AI system to order repair requests in public housing. The promise seems difficult to reject: attend to urgent faults sooner, reduce unnecessary journeys and make a queue that now frustrates many people more transparent. The presentation shows heat maps and a simulation of response times. When a resident asks how anyone will explain that persistent damp was placed below another incident, she receives a reassuring answer: “the model will consider every factor.” The phrase tries to end the concern, but it opens the central question. Who decided what every factor is, and who can change the answer when the experience of one building does not fit them?
Democracy does not begin when someone protests an unfair outcome. It begins earlier, while people can still discuss whether a problem should be automated, which objective deserves priority and which limits will not be traded away to save time. Artificial intelligence organizing public services is not an administrative accessory. It distributes attention, waiting and the ability to complain. If it arrives as a finished product, residents can only accept its logic or live with its consequences. That is why it needs more democracy: not as a ceremony of approval, but as a way of sharing power over design, testing and correction.
Consultation is not asking at the end#
It is easy to mistake consultation for a questions slide at the end of a presentation. By then, the central decision is usually already made: the supplier was chosen, a budget set aside, indicators defined and a date announced. Invited people can comment on details, but they have little room to rethink the project. That participation informs; it does not govern.
It is easy to mistake consultation for a questions slide at the end of a presentation.
Early consultation asks different questions. Before buying the tool, the council should explain which concrete problem it is trying to solve. Are there too many requests? Is technical staff missing? Do urgent reports fail to reach people able to act? Each diagnosis suggests different solutions. If tenants do not know where to report a fault, improving channels and in-person help may be needed. If staffing is insufficient, a classifier does not create repair workers. Automating without distinguishing the problem can turn a shortage of resources into a system that distributes the same scarcity more elegantly.
People who know the consequences from different positions should also participate: residents, caretakers, associations of older people, maintenance staff, social workers and people who do not regularly use digital channels. They should not be asked to design an algorithm, but to describe what a spreadsheet does not easily record. A damp mark can be aesthetic in one flat and a respiratory risk in another. A broken lift has a different effect on someone who can climb six floors than on a person with limited mobility. Situated knowledge is not an anecdote added after data; it is information needed to define what counts as urgent.
Disagreement improves the project#
When an institution presents AI as inevitable, disagreement can look like irrational resistance to progress. That attitude weakens the project. Differences reveal assumptions nobody stated: that everyone can use an application, that a photograph is enough to describe a fault, that a building’s history represents its current state, or that a calculated priority can be applied without seeing the case. Hearing such objections early prevents the first learning from arriving after harm.
Disagreement needs a concrete route. In the repair example, a monitoring group could publish priority criteria in plain language and open a period to discuss them. Collecting opinions would not be enough. The institution would need to answer what it incorporates, what it rejects and why. If a group points out that reports from homes without smartphones are underrepresented, the response cannot be a promise that this “will be taken into account.” It must describe an alternative channel, who records that information and how it will be checked that it enters the process with equivalent weight.
The ability to dissent protects workers as well. A technician may warn that a classification underestimates a fault because they know a particular building has old pipes. If the system is designed so that any change lowers the shift’s productivity, that warning will become invisible. The procedure must allow someone who knows the ground to correct a priority, explain the reason and see that correction reviewed as evidence for improving the rule. Giving voice is not adding an incident button; it is giving time, authority and protection to use it.
Participation requires understandable information#
Nobody can participate meaningfully in a decision described only through marketing language or opaque documentation. An explanation need not reveal secrets that compromise security, but it should answer essentials: what data is used, where it comes from, what result the system produces, which decisions remain human, what effects an error may have and how the function can be suspended. A two-page summary, an open session and examples of real cases are often more useful than a hundred-page technical report inaccessible to the people living in the buildings.
Information has to circulate before and during the trial. Publishing that “AI improves efficiency” after six months does not make assessment possible. A simple public dashboard, by contrast, can show response times by area, the number of priorities changed by staff, complaints received, errors found and cases sent to the alternative channel. Numbers do not replace stories, but they help people ask precise questions. If one area systematically waits longer, the community can request an investigation before a municipal average hides the pattern.
Transparency also needs limits on interpretation. An isolated metric can be used to justify a false conclusion. Fewer requests do not always mean fewer problems; they may mean the channel is harder to use or that people stopped expecting a reply. Data must therefore be accompanied by context and places where people living the situation can explain it. Accountability is a dialogue between records and experience, not the display of a dashboard claiming to speak for itself.
A pilot that can stop#
Democracy does not require waiting for impossible certainty before trying an idea. It requires that a trial have visible limits and that it can stop. The council could begin with a few buildings for a defined period, without removing current channels and with one clear rule: no request connected to safety, health or accessibility will be closed through automatic classification alone. Someone should be responsible for reviewing the pilot’s effects, with a public date for deciding whether it continues, changes or ends.
A serious pilot includes uncomfortable situations deliberately. What happens if someone cannot send an image? What if two reports look alike but affect residents differently? What happens when a recommendation conflicts with the judgment of the person visiting the property? Rehearsing these questions is not sabotaging the tool. It acknowledges that public life is full of exceptions and that a trustworthy system must work when reality escapes its form.
Evaluation should not belong exclusively to the supplier or to the team that argued for the project. An independent review can include neighborhood representatives, staff using the process and someone with knowledge of housing rights. It must be able to access necessary records, hear complaints and publish understandable conclusions. If it finds disproportionate harm, it needs authority to recommend a pause or require changes before use expands. Without that possibility, evaluation risks becoming a celebration of deployment rather than protection for the community.
Democracy means retaining the ability to correct#
Some institutions fear that opening these decisions will make them slow. Yet deciding without listening often produces another kind of slowness: resources spent repairing foreseeable conflicts, staff required to answer repeated complaints and residents who cease to trust the service. Deliberation can take time at the beginning, but it prevents system speed from becoming a burden for people with less power to respond.
The key is to distribute the capacity to correct. A resident must know how to challenge a priority. A worker must be able to point out an absurd rule without exposure. An association must be able to see whether a pattern affects particular buildings unequally. Those governing must be able to suspend a contract if safeguards are not met. Each of these routes turns AI from a silent authority into a tool governed by common rules.
The repair project can end up using technology and still be more democratic than before. It can organize information, help find delays and free time for complex visits. But only if it accepts that efficiency does not by itself define the public interest. A tool’s value is also measured by how it admits the experience of people who did not design it and by how easily it can be questioned when it fails. Democracy is not artificial intelligence’s brake; it is the steering system that stops it from moving quickly toward a place nobody chose.
