34.1The iterative mediation task
The procedure at issue is iterative in a specific sense. Participants offer opinions and critiques; those responses are used to revise the common-ground statements; the revised statements then return for further response. Each pass changes what the next round evaluates, so the object of discussion is under revision throughout rather than fixed at the outset.
Tessler et al. (2024) report AI refining common-ground statements through participant opinions and critiques, alongside preferred statements, converging views, and incorporation of dissent (Tessler et al., 2024).
Three outcomes should be kept apart. A preferred statement records preference: participants judge a formulation better than its alternatives. A converging view records change: positions that differed move closer across the exchange. Neither amounts to a binding decision. Preference for a wording is not sufficient for a mandate, and movement toward agreement is possible without being inevitable. Iteration provides occasions for revision; it does not guarantee convergence, nor does convergence guarantee lasting stability.
Convergence, as reported, compares participants' views with one another across the exchange; it does not compare the statements against an independent standard of accuracy. That comparison supports claims about the acceptability of formulations under repeated discussion. It does not by itself show that the resulting statements deserve authority over anyone.
Incorporation of dissent alongside majority positions is evidence that minority input can enter the statements. Whether every objection is preserved is a separate question, and the report does not settle it. Nor does it establish how preferences were elicited or how convergence was assessed, since those trial procedures have not been examined; this is an editorial knowledge limit rather than evidence that the source fails to establish its methods. The favorable findings therefore support a bounded conclusion: iterative mediation can yield statements that participants prefer together with a convergence of views, while the authority of those statements remains unestablished.
The reported use of the procedure depends on iteration: participants offer opinions and critiques, statements are revised, and later judgments are made on outputs shaped by that feedback (Tessler et al., 2024). Any comparison that substitutes a one-shot summarizer for this loop therefore changes the task, not merely the encoding of an unchanged task. It removes the return path through which participant input can alter what is presented, and it removes the repeated exposure by which views are reported to converge. The comparison may still be informative, but it should identify that change: the two procedures differ in whether participants' contributions are inputs to the statement at the moment of evaluation. Holding the task and inputs fixed would require giving each method the same material and the same opportunity to revise; without that, a difference in preference tracks procedure change as well as drafting quality.
Two further distinctions bound what the comparison can support. First, participant preference for an AI-mediated statement over a human-written alternative compares those two statements under that elicitation. It shows which was chosen, not that any statement is correct, inclusive of every dissenting position, or fit to bind anyone. Converging views likewise show reported movement among participants; they do not by themselves establish agreement about grounds or warrant the label of a legitimate decision. Second, evidence that successful statements incorporated dissent while respecting majority positions (Tessler et al., 2024) concerns the drafting of those statements. Incorporating dissent is a property that some outputs can display, and one the iterative feedback makes possible; it is not a necessary condition for preference, nor is it sufficient for authority. A one-shot system might produce a statement with the same apparent balance. The point is that its success would then be contingent, whereas iteration offers opportunities for such balance rather than a procedural guarantee of it.
The scope question follows. If the intended output is a statement a group prefers and can discuss further, iteration is a means of refining that statement. If the intended output is a binding decision, preference, convergence and dissent-inclusion remain preparatory evidence only.
34.2A statement the group prefers
Tessler and colleagues report participant preferences, converging views and incorporation of dissent (Tessler et al., 2024). Those results give the favorable case an evidential basis within that study's setting. They warrant examining what the mediation achieved and which further claims would require different evidence.
Preference, however, is a specific outcome, and its meaning should be preserved precisely. A favorable rating indicates that a participant regarded a statement as preferable to a relevant alternative; it does not by itself show that the participant revised any prior view. Nor does it show that the participant agreed with every clause. These distinctions matter because a mediated statement can succeed at coordination while leaving substantive disagreement intact. Preference alone is evidence of reception; the separate observations about changed views address a further outcome.
Two further claims would require additional evidence. A claim about binding decisions would need to establish that participants had conferred decision authority on the process—for instance through prior consent to delegation, an explicit ratification step, or a recognized procedure by which the resulting statement becomes operative. Preference ratings establish none of these: liking a formulation is not the same as authorizing it to settle a matter. A claim about complete preservation of dissent would need a record showing that each dissenting argument was captured, retained in an accessible form, and open to later challenge, together with some indication from participants that their objections remained visible rather than absorbed into smooth phrasing. The reported embedding-based evidence about dissent is relevant, but it is not an exhaustive argument-by-argument demonstration of this stronger preservation claim.
The study therefore offers positive findings about preference, changed views and the incorporation of dissent within its reported setting. The stronger claims considered here concern decision authority and exhaustive argument preservation. Those require evidence matched to those claims; acknowledging that requirement should not erase the favorable results already reported.
A favorable rating establishes only that, under the elicitation presented, the participant ranked one statement above the alternatives offered. In the hypothetical case already introduced, the participant favors the common statement yet keeps one conditional objection. There is no contradiction here: a conditional objection may concern circumstances that have not arisen, so it need not affect the present ranking. The rating records the ranking; it does not index which considerations remain open.
Preference is therefore neither necessary nor sufficient for authorization to make a binding decision. It is not sufficient because authorization raises a different question: who is entitled to settle the matter and on whose responsibility. Nothing in the ordering of statements answers that question. Nor is it strictly necessary, since a participant might accept an outcome for procedural reasons while preferring another wording. The two judgments can converge without being the same judgment. What the rating compares is candidate statements as favored under that elicitation; it does not compare who may decide.
To see what a rating can recover, hold the task and the candidate statements fixed and vary only how participant input is recorded. A single preference score and a record that preserves objections are different encodings of input on the same task. It is possible that the conditional objection can later be located and revisited; that outcome depends on the encoding, not on the fact of preference, and is not inevitable. A second limitation concerns elicitation: if the participant was never asked to state reservation conditions, their absence from the record shows nothing about the participant's position.
The further question is one of responsibility at transition — who carries an unresolved condition forward when a favored statement is adopted, and who is answerable if the triggering circumstance arrives. That question is left open by the preference evidence.
34.3Dissent inside common ground
The favorable findings extend to dissent. In the checked abstract, AI used participant opinions and critiques to refine statements; the report gives preferred statements, convergence of views, and incorporation of dissent alongside majority positions (Tessler et al., 2024). Read at face value, incorporation is stronger than a tolerance claim: dissenting input is said to have affected the wording, while majority positions co-occurred with that incorporation in the preferred statements. The reported success rests on a combination rather than a trade-off, in which respecting majority positions did not require erasing minority objections from the text.
That combination is a property of the statements produced, and it should be kept distinct from the other reported outcomes. Convergence of views is a change in what participants hold across iterations; statement preference compares candidate wordings under the elicitation used. Neither comparison is with an authorized decision, and convergence in particular records movement among views rather than a transfer of authority. Incorporation of dissent likewise shows that an objection left a trace in the wording. What remains outside the reported evidence is the transition by which such wording would authorize action, and with it the assignment of responsibility for carrying it out.
Finally, the evidential force of the finding needs sorting. It is reported for successful statements, which supports an association: dissent incorporation accompanied statements that were preferred. It does not show that incorporation was necessary for success or sufficient to produce it. As a labelled supposition, a preferred statement built without any dissent recorded would leave that association untested rather than refute it; a statement incorporating dissent that participants did not prefer would show the feature is not sufficient by itself. The supported claim stays modest: in this account, majority respect and dissent incorporation co-occurred in statements participants preferred.
The account already described should therefore not be dismissed as showing no dissent preservation. If some preferred statements incorporated a minority objection while still respecting the majority position, then common-ground statements can, in some cases, carry dissent into the shared text. That is enough to establish possibility, and enough to defeat a universal claim that synthesis necessarily erases disagreement. It does not establish that every material objection was preserved, or that preservation would be inevitable in a new design.
Two stronger claims require separate tests. Exhaustive retention would require every objection capable of altering the statement to appear, explicitly or by implication, in the result. Evidence that some objections were incorporated does not supply that comparison: it shows presence of particular dissent, not coverage of the whole set of objections raised across iterations. Nor is incorporation sufficient for more than inclusion. A minority reason can appear in a statement as a consideration while the majority’s conclusion still governs. Preserving a reason is therefore not the same as giving that reason binding force.
A further distinction concerns what “converging views” reports. It may describe participants’ successive opinions moving toward one another, which is a comparison across rounds of discussion. It need not compare the final statement with the complete record of earlier critiques. A statement can be preferred by participants, and can reflect some convergence, while still omitting grounds that were voiced but not carried forward.
The defensible inference is conditional rather than universal: where a process elicits critiques, revises candidate statements, and checks the result against participants’ views, dissent may survive into the common-ground text. Where those checks are absent, omission remains possible. On this scope, the account supports preservation as an available achievement, not as a guarantee of exhaustive retention.
34.4Majority position and minority reasons
A common statement can carry a minority reason alongside the majority position without making that reason a condition of the statement's authority. The two propositions differ. One says that the reason is preserved within the shared record; the other says that the statement stands only if the reason is satisfied. Conflating them yields two symmetric errors: it may hand a single objector a veto that no one agreed to grant, or it may treat recorded dissent as decoration to be discarded once the wording is settled. Giving a reason veto weight would make its satisfaction necessary for the statement to stand. Merely including the reason leaves satisfaction unnecessary while keeping the disagreement visible.
The hypothetical test makes the separation concrete. A participant favors the common statement yet maintains one conditional objection. Recording that objection inside the statement would show that it had been heard; it would not show that the objection constrains later use. The preference rating is still less informative. A rating expresses an overall stance toward the statement as a whole, and it does not establish whether that exact conditional objection can later be located in whatever representation is relied on. Traceability of a particular reason is therefore a distinct property from acceptance of the whole, and neither by itself is sufficient for the other.
Consider next how strong the case for inclusion really is. Recording the objection makes later retrieval possible, but not inevitable: retrieval depends on the representation being kept, and on someone looking for the conditional qualification when the statement is reused. Verbatim preservation would narrow that gap without closing it, since a reason written into a record can still be passed over. What inclusion does support is the narrower claim that a later dispute can be checked against something the participants actually shared. That gives the minority reason a place to be found; it does not, by itself, give the reason any power over what follows from the majority position.
Once a minority reason is carried into a common statement, the workflow owes an explicit account of what follows. It must name the operative consequence: whether the reason becomes a reservation attached to the text, whether it opens a specified route to revision, or whether it governs only the later interpretation of the passage it qualifies. Each of these is a decision rule; none is entailed by the reason's presence in the document. To leave the choice unstated is to let the format of the statement silently assign authority.
The hypothetical test makes the problem concrete without settling it. The participant favors the common statement yet keeps one conditional objection. The rating answers one question: how the respondent stands toward the statement as presented. It compares assent with that text; it does not compare the objection's wording, its preservation, or its later recoverability. A favorable rating is therefore compatible with an objection recorded in full, with one preserved only in outline, and with one never captured. The rating cannot tell these apart, so it cannot by itself tell whether the objection can later be located.
That is why the stated consequence must be checked against retrieval. Recording the objection is necessary if anyone is later to find and answer it, but recording is not sufficient: the condition must appear in its own terms and be linked to the claim it qualifies. If the rule says only that minority reasons will be noted, the outcome may be a record that cannot be resolved back to any proposition someone could accept or contest. What inclusion changes is the content of the record; what it decides about that record's later force remains a separate rule the workflow must supply.
34.5What preference ratings leave unanswered
A preference rating answers a question fixed in advance: whether the respondent favors the statement as presented, under the conditions supplied in that elicitation. That is a genuine finding, and it is bounded. It compares the respondent's assessment of one formulated proposition with the scale or alternatives given at that moment. It does not compare the statement with every variant the respondent might accept, nor does it record which conditions the respondent took to be load-bearing.
Take the hypothetical test already in view: a participant favors a common statement while maintaining one conditional objection. The rating registers the favoring. It is compatible with the objection being trivial, essential, or already resolved; those possibilities are not ranked by the number alone. A favorable rating is sufficient evidence of preference under that elicitation, but it is not sufficient evidence that no qualification remains, and it is not evidence that the objection has been withdrawn. Conversely, preference is not necessary for agreement, since a participant might accept a statement for coordination while rating it modestly.
Nor does the rating settle whether a view changed. One elicitation records a position; a transition claim requires either two comparable ratings or an explicit account of revision. The rating also speaks to the respondent's attitude, not to the authority of the statement. Binding force depends on a rule specifying whose favorable assessment counts, at what stage, and with what consequence. That rule is separate from the elicitation; importing it silently turns a reported preference into a decision procedure it never carried.
Finally, the rating leaves open what the record preserves. In the hypothetical, it does not tell whether the participant's exact conditional objection can later be located, since the rating's inputs are the presented statement and the scale, not the identity of the objection. Holding the task and inputs fixed, different encodings of the same encounter may therefore support different later claims about what the participant meant.
A preference rating reports how a participant evaluated the statement placed before them at the moment of evaluation. What it leaves unanswered is whether the qualification attached to that evaluation survives in a form a later reader can find. Consider the hypothetical test already introduced: the participant endorses a common statement while keeping one conditional objection. The endorsement is recorded; the objection is the condition on which it depends. Nothing in a rating value by itself establishes where that condition went.
Two distinct questions therefore need separating. The first is what the participant preferred. The second is whether the condition can be recovered and contested. Answering the second requires a different task with its own inputs: a retrieval or correction exercise in which the reader is asked to locate the condition, state it, and propose a change if it has been misrecorded. Holding that task fixed while varying only how the objection was recorded lets a comparison isolate the effect of the encoding rather than mixing it with changes in the material presented.
The resulting test is also asymmetric in what it can show. Successful retrieval of the objection is necessary for a later reader to challenge it, but it is not sufficient: the reader must also be able to register a correction and have the record respond. Failure to retrieve, meanwhile, makes challenge impossible in that instance, but it does not make loss inevitable in every arrangement; a condition that is preserved and indexed differently may remain locatable. So the hypothetical outcome is possible under a rating-only record, not guaranteed by it.
Thus the rating answers a bounded question about the presented statement, while the separate task asks whether its conditions remain inspectable.
34.6A comparison the prototype still needs
The favorable ratings reported so far show that participants preferred the mediated statements; they do not show that source links or correction paths hold up when challenged. A fair test should therefore keep the deliberation task, the available inputs, and the interactive exchange fixed, varying only the provenance layer: one version presents the settled statement alone, the other adds linked sources and a route for lodging corrections. If the question being decided or the information offered also changes, any difference may be owed to a different task rather than to the encoding.
What the comparison measures is the fate of a fixed objection, not overall approval. Confront both conditions with the same challenge that a qualifying condition has been dropped from the statement. Then record whether the objection can be attached to the claim it contests, whether the underlying basis can be retrieved from the record itself, and whether a resulting correction is visible to later readers. These endpoints differ: finding a source is not amending the record, and amending the record is not binding the group to the amendment.
Two qualifications shape the test. A source link is not sufficient for accountability, since a link to an inapposite or unretrievable basis leaves the objection standing. Nor is a link strictly necessary if the basis can be recovered reliably by other means. Equally, the affordances make correction possible rather than inevitable: they may create more points at which a dropped condition becomes visible, or simply more material through which it disappears. Which effect dominates is exactly what the comparison is meant to determine. Because no result exists yet, the design supplies a question rather than a verdict, while leaving the mediator's interactive strengths in place and requiring source inspection and correction to demonstrate what they add.
The proposed source-linked correction task extends the mediator's process with records: a claim now carries a source link, and a correction is entered rather than merely voiced. That extension is a prospective design; no result of the comparison exists. The issue is therefore incremental benefit — what the records add over what the simpler mediator already achieves — rather than whether a richer prototype can in principle produce better statements.
To isolate that increment, task and inputs must be held fixed: the same dispute, the same material, the same obligation to reach a statement. Only the encoding varies, with linked records in one condition and an unlinked exchange in the other. If the assigned work differs across conditions, any difference in the statements reflects the task as much as the encoding, and the comparison says nothing about recordkeeping.
Two qualifications follow. First, improved correction accuracy would be welcome but not sufficient to justify the extension: the simpler mediator might reach comparable corrections without records, so the added burden must be weighed against that possibility rather than against no correction at all. Second, the quantities need naming. Burden means the effort a joint statement costs under each encoding, not the number of entries created; usefulness means whether the output supports the work actually assigned afterward, not whether it looks more complete.
A conditional countercondition marks where the richer version would win. Supposing a later participant must revisit a disputed claim without having taken part in the original exchange, a source link and a recorded correction could carry information the unlinked mediator cannot supply, and that could repay the added burden even if current statements are equally accurate. Absent such a downstream demand, the richer prototype may lose to the simpler one on burden and usefulness. Its loss is possible, not inevitable: it turns on whether records are consulted, and on for whom. The comparison should therefore name the downstream use before the encodings are judged.
34.7A bounded account of the benefit
The favorable evidence earns a clear credit. On the reported comparison, the AI-mediated common-ground statements were preferred to the alternatives participants were given, and that preference is a real benefit: the mediator produced formulations that people were readier to accept. Nothing in the measured comparison licenses a stronger claim, and the strength of the claim is what matters here.
The comparison holds the task and its inputs fixed and varies only the way the common-ground statement is encoded and presented. Any measured gain therefore attaches to that encoding. It does not compare one mediation arrangement with another, and it cannot speak to the records a richer mediator might carry.
Three outcomes must be kept apart. A preferred statement is one a participant endorses over an alternative. A changed view is a revision in what a participant holds. A legitimate binding decision settles a matter for those it governs. Preference is evidence of acceptability; it is not sufficient for either of the others. Endorsement can be produced without any authority to commit, and a formulation that reads well may be agreed to without a view being revised. Convergence of views is a possible effect of mediation, not an inevitable one, and even genuine convergence still falls short of a decision, which requires a further condition — responsibility to commit — that preference does not supply. Nor is prior preference always necessary for a decision to bind; some arrangements bind without it.
That defines the narrower issue pursued here: not whether the mediator helped, but which of these three outcomes the favorable result establishes, and what each further outcome would additionally require. The proposed source-linked correction task bears on that issue. It adds records to the mediator's process, so its extra burden must be justified by an incremental benefit, and no comparison yet shows one. The case accordingly remains a prospective design.
A research boundary marks the limits of what has been shown; it does not negate what lies within them. If the reported benefit is that mediation produces preferred statements, the warranted conclusion is likewise bounded: selection at elicitation is evidence of a judgment at that moment, not of a changed view, and still less of a decision that binds anyone.
Three outcomes therefore need separating. A preferred statement is one a participant selects among offered alternatives. A changed view is a revision of judgment attributed to the encounter. A binding decision is one whose authority survives a demand for justification. Selection does not entail revision: it may simply record a stable prior judgment. Neither entails authorization, since preference and persuasion do not by themselves supply the standing to commit others. Eliciting a choice is necessary for a claim of endorsement but not sufficient for legitimacy; conversely, a legitimate decision need not rest on anyone changing their mind.
Separating preference, changed judgment and authorization leaves open that current mediation already suffices for some purposes. For coordination around a contested point — settling which statement a party will stand behind in later exchange — elicited preferences may be adequate, and additional machinery would then add nothing. Suppose instead that the object is a decision whose legitimacy can be challenged. Then the account of mediation has not supplied the record that would answer such a challenge. The proposed source-linked correction task targets exactly that gap, by adding records to the mediator's process. It remains a prospective design with no comparison result: its superiority is possible rather than inevitable, and it must justify the extra record-keeping it imposes.
Any such comparison must fix the task and inputs — the same correction question, the same participants, the same underlying material — so that only the records differ. Otherwise a difference in outcome cannot be attributed to the encoding. The point is not that existing mediation fails, but that its adequacy is purpose-relative, and untested for the harder purposes.