On the Monday after the neighbourhood forum, the association received a polished document. An artificial-intelligence tool had read the transcripts from four tables, arranged the proposals, and produced a clear report with persuasive headings and a list of priorities. Its first page carried a reassuring sentence: “There is consensus on redesigning the square.” Participants recognised some of their own words, but they did not recognise the meeting. The proposal for the square had been the most contested item: shopkeepers worried about deliveries, families asked for shade and play space, older residents feared losing nearby benches, and another group opposed using the area as temporary parking.
The problem was not poor writing. The smoothness of the report hid a decision. To summarise, an AI must group, rank, and leave things out. Those operations can help people navigate a long discussion, but they are not neutral when applied to a public conversation. A proposal mentioned repeatedly because people debated its consequences may be read as support. A difficult condition expressed in one long intervention may vanish from a synthesis that rewards short, tidy statements. When a tool chooses “participants agreed” where the record should say “participants debated,” it changes the civic meaning of what happened.
Summarising is a form of mediation#
It helps to stop treating a summary as a shorter photocopy. A neighbourhood record does more than preserve information: it allocates recognition. It shows which matters reached the table, which reasons shaped the decision, and how people can later check whether their participation was respected. When a document says there was consensus, it can also close a future discussion; anyone who objects may appear to be challenging an agreement that has already been made. A trustworthy synthesis does not manufacture one voice from many voices.
It helps to stop treating a summary as a shorter photocopy. A neighbourhood record does more than preserve information: it allocates recognition.
AI is useful for speeding up work that otherwise gets postponed: transcribing audio, finding recurring themes, flagging related proposals, or preparing a readable first structure. It can notice that lighting came up at three tables and gather those references without making staff search line by line. Yet finding recurrence is not the same as interpreting a position. At one table, “we need more light” may mean safer routes for children walking to school. At another, it may be a criticism of intrusive floodlights and cameras. A responsible report does not merge those reasons merely because they share a word.
The practical distinction is straightforward. A tool may propose a draft; the community keeps the authority to decide what the draft represents. For that authority to be real, people need to be able to trace every important claim. A note saying that AI assisted the writing is not enough. The report should show where a conclusion came from, which contributions support it, and which relevant contributions do not fit it. Provenance is not a technical ornament. It is what makes it possible to challenge a summary without having to distrust the entire process.
Keep proposals, reasons, and debate status apart#
One concrete improvement is to avoid asking the system for “a consensus report.” Ask instead for a conversation map with separate fields. Each proposal should include a short description, reasons offered in favour, reservations raised, unanswered questions, and the current state of discussion. That state might be “broad support,” “competing alternatives,” “not discussed in sufficient time,” or “needs further evidence.” These labels are less attractive than a headline, but they prevent editorial convenience from replacing deliberation.
Consider the square again. A draft could record broad support for improving daily use of the space without claiming agreement on the solution. It could separate the request for shade from the dispute over removing parking spaces and note that people making morning deliveries have not yet been heard. It should also retain a minority proposal even if it does not prevail, such as reserving some hours with no programmed activities. A minority does not need to win in order to be represented accurately. Sometimes it names the risk that a majority will want to revisit when circumstances change.
This separation also guards against a common mistake: assuming that the number of mentions measures the strength of a reason. In an open forum, people with more time, meeting experience, or confidence may speak frequently. Someone caring for a child, speaking another language, or arriving late after work may make one decisive observation. A system that counts sentences can amplify the inequality already present in the room. Human reviewers must ask whether the report confused volume with representation and whether it left out written submissions, paper notes, or accessible channels.
Sources must be usable, not merely stored#
Linking a summary to sources does not mean putting every audio file and every name online without care. Some people need to contribute without having a sentence permanently connected to them. The answer is a proportionate evidence layer: anonymised records where needed, extracts identified by table and moment, original documents with restricted access, and a plain explanation of what is withheld for privacy. What matters is that a participant can find the basis for a statement attributed to the group and ask for a correction when it has been taken out of context.
A sound review package can include links or reference codes for the source intervention, a mark wherever the AI has paraphrased, and a list of editorial decisions. If two similar proposals were combined, say so. If a detail was removed because it could expose someone, say that too. This openness helps people distinguish legitimate protection from convenient omission. It also makes the next meeting more useful: instead of spending it arguing about whether the report reflects reality, participants can discuss what to do with the disagreements the report has made visible.
Corrections need to be possible before the text acquires authority. Many organisations publish a final PDF and then open an email address for comments. By then the document may already be shared online, quoted at meetings, and treated as the official memory. It is better to set a review window with a deadline, named contacts, and several routes: online annotations, an in-person session, a phone call, and paper amendments. Not everyone can use the same platform or respond at the same pace. Accessibility is part of the report’s accuracy.
Approval is not a button press#
Collective approval should not become a yes-or-no poll on a twenty-page document either. A more honest method first shows a change summary and asks for confirmation in parts: facts recorded, wording of proposals, disagreements, decisions made, and matters still open. A participant can then accept that their contribution was transcribed faithfully while challenging the label “consensus” for a vote that never took place. The disagreement is preserved without forcing that person to reject the whole piece of work.
For choices that require formal agreement, the community should define in advance who approves and under which rule. It might be an assembly, an elected committee, or the people who attended a particular table. The essential point is not to invent the validating body after seeing the outcome. AI cannot settle that question because it does not independently know the mandate, justified absences, or neighbourhood procedure. It can prepare compared versions and highlight changes, but it cannot confer legitimacy.
Before publication, a named reviewer should answer four questions. Does every claim of consensus have a verifiable source? Are disagreements that affect the decision described without caricature? Do received corrections have a visible response? Is it clear which version was approved and when? If any answer is no, the report is still a draft, however elegant it looks. This checklist does not slow participation down; it prevents speed from turning a misunderstanding into a decision that is hard to repair.
The best final version of that forum may not promise a solved square. It can say precisely that there is a shared intention to improve it, that three alternatives remain on the table, and that a temporary trial will be called with criteria for reviewing it. That sentence retains the energy of the gathering without claiming a unanimity that did not exist. AI can help organise complexity. The collective commitment is not to ask it to erase complexity. When sources, corrections, and approval are visible, the summary stops speaking for the neighbourhood and becomes a tool through which the neighbourhood can keep speaking.
