More than a thousand citizens' assemblies and comparable deliberative processes have now been run worldwide. The evidence on what they produce is unusually good for this kind of intervention: they depolarise opinion, they generate recommendations that hold up to scrutiny, they rebuild a measure of trust in institutions among the people who take part.
The people who run them know something else, which appears in the evaluations less often. Almost none of what an assembly builds survives it. The recommendations survive: they get written up, handed over, sometimes acted on. The understanding that produced them does not. It exists for the duration of the process, in the room and in the heads of the participants, and then it disperses.
That is not a criticism of the format. It is a description of what happens when a genuinely sophisticated practice runs without any infrastructure underneath it.
What an assembly actually builds
Watch the middle phase of a well-run assembly and what is happening is not opinion formation. Participants arrive with positions, encounter evidence and each other, and construct something considerably more interesting than a revised opinion: a shared map of why the question is hard. Which trade-offs are real and which are presentational. Where the expert testimony conflicts and what turns on the conflict. Which objections were answered and which were absorbed without being answered.
That map is the actual product. The recommendations are a projection of it: a compressed output, shaped for handover to an institution that asked a specific question. Compression is the right thing to do at that point. The problem is that the compressed version is the only version that leaves the room.
The recommendations survive. The understanding that produced them does not, and the understanding was the harder thing to build.
What is lost at handover
An institution receiving a set of recommendations receives conclusions without the reasoning that produced them. Officials implementing them cannot tell which are load-bearing and which were compromises that everyone could live with. Or where corners may have been cut by relying too much on AI. When circumstances change, and the gap between assembly and implementation is frequently years, nobody can determine whether a recommendation still applies, because the conditions it depended on were never recorded in a form anyone can interrogate.
The minority positions fare worst. In a deliberative process the considered dissent is among the most valuable material generated: these are people who engaged the full evidence and reached a different conclusion for stated reasons. In the final report they appear, if at all, as a noted reservation. When the chosen path begins producing surprises, the fastest route to understanding why is precisely that dissent, in the form it was made, before anyone knew who was right. It is almost never retrievable.
Every assembly starts at zero
The structural consequence is the one that should concern anyone who cares about this field. A thousand deliberative processes have produced a thousand separate starting points.
An assembly on housing in one city has no access to what an assembly on housing in another understood two years earlier: not the recommendations, which are published, but the reasoning: which framings turned out to be dead ends, which piece of evidence changed the room, which apparently technical question turned out to be a values question in disguise. Each new process rediscovers this at full cost. The practice has accumulated a great deal of methodological expertise among facilitators, and almost no accumulated understanding of the substance.
Compare that with any other field that runs repeated structured inquiry. Research accumulates through publication and citation. Engineering accumulates through standards and post-incident review. Deliberative democracy accumulates through the personal memory of practitioners, which is why the field's knowledge is so heavily concentrated in a small number of experienced individuals and so difficult to transfer.
The readiness gap nobody budgeted for
Two assessments this year, from Carnegie and from DemocracyNext, reached the same conclusion from different directions. Institutions running deliberative processes largely lack the staffing, procurement discipline, legal frameworks and auditing capacity required to steward AI-enabled deliberation responsibly.
The finding is usually read as a warning about adopting AI too quickly. It is also a description of something more basic: these institutions have no deliberative infrastructure at all. There is no procurement category for it, no standard for what a process should leave behind, no audit trail that would let anyone check months later whether a recommendation reflected the deliberation or the facilitator's summary of it. The AI question arrived at an institution that had not yet solved the record-keeping question.
That ordering matters, because it suggests the useful intervention is not primarily about AI. It is about whether a deliberation produces something that persists, can be interrogated by people who were not present, and can be built on by the next process.
What infrastructure would change
A deliberation that runs on a living structure rather than a facilitator's notes produces a different kind of artefact. Perspectives are preserved with their grounds attached, whether a contribution came from research, professional expertise, lived experience or a values position, so that a reader months later can tell not just what was argued but from where. Connections between contributions are explicit, so the reasoning is navigable rather than reconstructed. Minority positions remain findable and attributed. And where alignment is genuine rather than nominal becomes visible rather than assumed.
The compounding effect is the real prize. An assembly that inherits the structured understanding of three comparable assemblies does not start at zero. It starts by interrogating what was already understood, and spends its expensive deliberative capacity on what is genuinely unresolved. That is the difference between a field that runs a thousand events and a field that builds.
Why this is addressed to practitioners
The people who convene and facilitate these processes are, by some distance, the most qualified group in existence at the actual work of collective sensemaking. They understand what a room needs, why sortition matters, how to hold a disagreement open long enough to be useful. None of that is in question.
What they have been asked to do, repeatedly, is build something remarkable and then let it disperse. A thousand times over, that is an extraordinary quantity of understanding that a society paid to produce and then did not keep. The practice is mature. The infrastructure is missing, and it is the thing standing between a thousand impressive events and a democratic capability that accumulates.