An AI mediator was placed between people who disagreed about immigration, Brexit and climate policy, and the groups it worked with moved measurably closer together. The finding is real, it held up under scrutiny, and it deserves better than dismissal. If a language model can take a room full of people who disagree and produce a statement more of them endorse than the one a skilled human mediator produced, what exactly is left for people to do together?
The answer is not that the study was wrong. The study was careful. The answer is that it measured the thing it set out to measure, and that thing is not deliberation. Reading the result correctly is more useful than disputing it, because what it demonstrates, precisely, is the boundary of what mediation can reach.
What the study did
The design was straightforward. Participants gave their individual views on a contested question. A language model read those views and drafted a group statement. Participants critiqued the draft, the model revised, and the loop repeated. At the end, participants rated the statements, and they endorsed the machine-drafted one more highly than the version produced by a human mediator working from the same inputs. Groups also reported positions closer to each other after the process than before.
That is a genuine result about a genuine capability. A system that can read many partial positions and synthesise language that a majority will accept is useful, and in some settings it is exactly the right tool: drafting a communiqué, finding acceptable wording under time pressure, surfacing where a group is already nearer than it thinks.
What common ground measured
The outcome variable was endorsement of a statement. Participants were asked whether they accepted a form of words. That is a clean, honest measure of acceptability. Acceptability is a property of an output, not of a group's understanding.
This is the same distinction that separates collaboration from collective sensemaking, arriving in a new form. Collaboration produces a document; the question it answers is what did we agree. An AI mediator is an extremely efficient instrument for answering that question. It does not touch the other one: what do we now understand that none of us understood before.
A group that agrees has a position. A group that understands has something to work with when the position stops working.
What compression removes
A mediator's task, human or machine, is to find phrasing the largest possible number of people will accept. That is an optimisation, and it has a direction: toward the centre of what was already said. The reasoning that makes a position distinctive is friction in that process. It is what has to be smoothed away for the sentence to land.
Most of the time the smoothing is harmless. Occasionally it removes the only thing that mattered. The participant whose objection came from having watched the same policy fail somewhere else, the one whose discomfort was a values position rather than an empirical claim, the one who was simply right and outnumbered: each of them contributes a sentence fragment to a consensus statement and loses the ground the fragment stood on. Nothing in the endorsement score can detect this, because the statement scores well precisely when the friction is gone.
This is why every Spark on Hunome carries its knowtype: the contributor's own declaration of how they know what they are contributing: research, expert fact, lived experience, values, gut feel, etc. Two people can write an almost identical sentence from completely different ground, and the difference between them is frequently the most decision-relevant information in the room. A consensus statement records the sentence and discards the ground. A SparkMap keeps both, along with the connections between them, which is where the understanding actually lives.
The question the study could not ask
Six months after a mediated session, the statement still exists. What does not exist is any way for the group to reconstruct why it says what it says: which objections were raised and answered, which were raised and absorbed without answer, where the agreement is load-bearing and where it is a form of words that everyone could live with. When circumstances shift, and they do, a group holding only the conclusion has to start again. A group holding the reasoning can move on and adapt.
That is the difference between retrieval and emergence. An AI mediator recombines what was already expressed into a more acceptable arrangement. Deliberation produces understanding that nobody brought into the room: the connection one person draws from another's experience, the reframe that makes an inherited assumption visible, the pattern that forms across contributions from people who have never spoken. No optimisation over existing statements reaches it, because it was not in the statements.
Where AI genuinely belongs
None of this argues for keeping AI out. It argues for putting it where it compounds rather than where it compresses.
On Hunome, The Lens shows what happens on the SparkMap. The lead and the contributors see and can navigate to the areas of interest. Clusters show what is sizable and what is emerging and where edge thinking resides. Trains of thought show where contributors have experienced an Aha moment. SparkMap network view shows what characterisation the Sparks have and enables viewing spots of interest. It also shows where perspectives are structurally absent. Hunome AI overall does not write the deliberation’s conclusions, even if it suggests summaries. This makes the deliberation navigable, so that people can deliberate over more than they could otherwise hold in view or mind. AI in that position increases what a set of contributors can understand.
The distinction to hold
Can AI find common ground? On the evidence, yes: reliably, and better than we might have expected. Hold that finding at exactly its size. Finding common ground is a service performed on a set of positions. Deliberation is the work that produces positions worth holding in the first place, and it leaves behind a structure a group can keep using.
An organisation that adopts AI mediation will reach agreement faster. Whether it understands more is a separate question, and the endorsement score was never going to answer it.