Complete underlying reasoning (97 claims and questions)
The main steps above are a reading path. These are all assessed contributions to this argument, including challenges and questions.
Answered · How did the AI mediator choose a group statement?
Selection is part of the reported procedure, but does not itself establish consensus or causal effects.
How did the AI mediator choose a group statement?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did majority-leaning statements predict movement toward the majority?
This bears on whether possible bias explains measured convergence, while not establishing causation.
Did majority-leaning statements predict movement toward the majority?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the study measure group agreement, and did it increase with AI-produced group statements compared with simply viewing others’ opinions?
Measured convergence and its comparison condition relate to qualifications about consensus and AI-specific causation.
How did the study measure group agreement, and did it increase with AI-produced group statements compared with simply viewing others’ opinions?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did group members become more aligned after AI-mediated discussion?
Measured alignment is qualified against equating it with consensus or causation.
Did group members become more aligned after AI-mediated discussion?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What kind of mediator did the researchers develop?
The bounded mediator role is context for interpreting findings without overreading them as delegated governing authority.
What kind of mediator did the researchers develop?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the AI system select among candidate statements?
Selection is procedural context and does not establish consensus or an AI-specific causal effect.
How did the AI system select among candidate statements?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What information went into revisions, and how long did a round take?
This asks about the procedure underlying the reported deliberation evidence, not its general effectiveness.
What information went into revisions, and how long did a round take?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did the findings also appear in a group of citizens discussing public issues?
The evidence answers whether findings appeared in the assembly sample, with the reported setting limiting the inference.
Did the findings also appear in a citizens’ assembly sample?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesPartly answered · Can people discuss an issue carefully and openly to help different groups move toward a common position?
The narrative directly concerns measured alignment and its limits; evidence reports closer positions and agreement after AI mediation, with comparison and self-report qualifications.
Can people discuss an issue carefully and openly to help different groups move toward a common position?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What did using a language model let the researchers examine?
The question concerns what the researchers measured in their deliberation study.
What did using a language model let the researchers examine?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesPartly answered · How did the authors test the mediator in different settings?
The question concerns study design underlying the empirical evidence, though the passage does not report outcomes.
How did the authors test the mediator in different settings?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the authors define possible majority and agreement biases?
It asks about the study’s defined measures rather than an unrestricted conclusion about bias.
How did the authors define possible majority and agreement biases?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did participants describe their own views during the discussions?
The answer supplies a specific reported view-change measure and shows that most rounds involved no reported change.
How did participants describe their own views during the discussions?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What study plans did the authors register in advance?
Preregistration is relevant context for interpreting the empirical evaluation, without independently confirming its findings.
What study plans did the authors register in advance?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did AI-assisted discussion change how divided participants were?
The question directly concerns measured alignment, answered by the reported comparison with unmediated exchanges.
Did AI-assisted discussion change how divided participants were?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Was acceptance of the final statement related to shifts toward other members’ views?
The evidence answers the association question and explicitly limits causal interpretation.
Was acceptance of the final statement related to shifts toward other members’ views?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did the group statements move toward a middle ground?
The question asks about the study’s measured movement toward compromise, which the narrative treats as evidence distinct from consensus.
Did the group statements move toward a middle ground?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did the text measure reflect participants’ stated positions?
The question concerns how the study’s text-based measure relates to stated positions, a qualification relevant to interpreting its results.
Did the text measure reflect participants’ stated positions?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the AI’s training goal relate to the changes in participants’ views?
This distinction qualifies the interpretation of measured view changes and prevents conflating training objectives with deliberative outcomes.
How did the AI’s training goal relate to the changes in participants’ views?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How was the AI helper designed and trained to produce group statements?
The system design and training help define what the study tested and qualify interpretation of its reported outcomes.
How was the AI helper designed and trained to produce group statements?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How were the system’s two models set up?
The model setup is part of the reported study method and bounds what its results establish.
How were the system’s two models set up?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How were initial statements assessed and selected?
These procedures clarify the study’s measures and selection mechanism, which inform how its results should be interpreted.
How were initial statements assessed and selected?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What research materials can readers access?
Access to these materials can aid inspection of the study’s evidence, while the stated limits prevent inferring complete research-material availability.
What research materials can readers access?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the authors measure the position expressed in a statement?
The measurement procedure is relevant to what the study’s position and compromise results do—and do not—show.
How did the authors measure the position expressed in a statement?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the AI helper select a group statement?
The selection procedure is part of the study’s method and does not, by itself, establish that the chosen statement represents everyone fairly.
How did the AI helper select a group statement?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How much did views shift on the issues where participants moved in a common direction?
The reported issue-specific results directly describe measured view changes, with the exploratory and uneven nature of the findings retained.
How much did views shift on the issues where participants moved in a common direction?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesOpen · What additional outcomes are mentioned beyond the measures described here?
The narrative concerns limits on what the reported evidence establishes; the missing supplementary results leave these additional outcomes unresolved.
What additional outcomes are mentioned beyond the measures described here?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did AI mediation reduce disagreement more than simply showing people one another’s views?
This is a study comparison relevant to interpreting the reported alignment, while leaving causal attribution qualified.
Did AI mediation reduce disagreement more than simply showing people one another’s views?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did the group statements fall between members’ positions?
The result is a measured compromise indicator, relevant to the narrative’s caution that measured alignment does not itself establish consensus or broader deliberative success.
Did the group statements fall between members’ positions?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the authors analyze participants’ ratings and other results?
The analysis methods contextualize reported deliberation evidence but do not resolve the narrative’s limits on causal or consensus claims.
How did the authors analyze participants’ ratings and other results?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did participants prefer revised statements?
This is a study-specific preference and rating result; it is relevant evidence but does not by itself establish consensus, causation, or readiness for wider deployment.
Did participants prefer revised statements?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Could the discussion process, rather than the AI itself, explain why views moved closer together?
The exploratory human-mediator comparison qualifies attribution of convergence to AI specifically; it leaves the causal role of the mediation process uncertain.
Could the discussion process, rather than the AI itself, explain why views moved closer together?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the researchers estimate positions from written opinions?
The measurement method and imperfect correspondence qualify interpretation of the study’s text-based position and deliberation measures.
How did the researchers estimate positions from written opinions?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did people move toward the same view, or did support simply shift toward the majority?
The result distinguishes directional movement from consensus, qualifying what group alignment implies.
Did people move toward the same view, or did support simply shift toward the majority?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · How did the study compare group statements and show uncertainty in its results?
The comparison measures and confidence intervals contextualize the reported evidence without resolving broader causal or consensus claims.
How did the study compare group statements and show uncertainty in its results?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Were the opinion shifts consistent across all issues?
Issue-specific variation and a nonsignificant result qualify broad inferences from reported shifts or alignment.
Were the opinion shifts consistent across all issues?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Could simply seeing more majority views explain the shift?
The result weighs against a simple exposure account but does not establish the alternative explanation or settle causation.
Could simply seeing more majority views explain the shift?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Who took part, and what did the study examine?
The sample and topics define the scope of the deliberation evidence.
Who took part, and what did the study examine?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What research gap did the authors identify?
The reported research gap contextualizes the study’s empirical claims without establishing their strength.
What research gap did the authors identify?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What topics did participants discuss?
The topics delimit the subject matter and setting of the reported study.
What topics did participants discuss?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesPartly answered · Could AI bias explain the changes seen across groups?
This bears on causal interpretation, but the reported result does not exclude the explanation conclusively.
Could AI bias explain the changes seen across groups?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What did participants do in the main deliberation task?
The task and measurement procedure contextualize evidence about view change and agreement.
What did participants do in the main deliberation task?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What was the system designed to produce?
The output’s nonbinding nature limits what reported agreement results imply.
What was the system designed to produce?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · Did adding participants’ critiques make group statements more acceptable?
This is direct evidence about statement acceptability, not proof of consensus or general effectiveness.
Did adding participants’ critiques make group statements more acceptable?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesAnswered · What happened in the virtual assembly, and how did the authors interpret it?
The assembly result and authors’ interpretation are relevant; the causal caveat limits inference without withholding the reported answer.
What happened in the virtual assembly, and how did the authors interpret it?
AI browsing question drawn from this source idea; not a quotation or a claim that the authors asked this question.
Inspect the contribution and sourcesImportant limit · Plans set in advance
Preregistration is methodological context, but the missing plan details limit further inference.
The authors say they preregistered, or recorded in advance, sample sizes, rules for deciding which participants or results to leave out, and main analyses for all evaluation experiments.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage reports preregistration but does not provide the details of the registered plan. Addresses: The authors’ reported study procedures.
Inspect the contribution and sourcesStarting point · The AI revised statements using feedback
This process underlies the reported outcomes, while those outcomes do not by themselves establish consensus or an AI-specific cause.
The AI helper drafted a group statement from participants’ opinions, then revised it using their critiques. The design drew on Jürgen Habermas’s theory of communicative action, a theory about how people communicate and reach understanding. Across successive rounds, researchers used supervised fine-tuning, or further training with examples, and reward modeling, or training a model to score outputs, to improve the system. Group approval was the stated goal.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article says the system used individual views to generate statements and participants’ critiques to refine them. Across successive rounds, researchers used supervised fine-tuning, or further training with examples, and reward modeling, or training a model to score outputs, to improve its ability to capture shared views and maximize group approval ratings. Addresses: The Habermas Machine’s process and training
Inspect the contribution and sourcesStarting point · The study examined groups in the United Kingdom
The sample and topics delimit the scope of the reported evidence.
The experiments involved 5,734 participants from the United Kingdom and examined common ground, changes in participants’ views, how statements handled minority and majority views, and possible bias. The researchers also tested the approach in an online citizens’ assembly, a meeting where residents discuss public issues, with a sample described as reflecting the UK population’s demographics.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports 5,734 UK participants and an online citizens’ assembly involving a sample described as reflecting the UK population’s demographics. It lists several questions the experiments examined, including view changes and possible bias in group statements. Addresses: The participants, settings, and research questions
Inspect the contribution and sourcesSupporting point · Statements moved toward compromise
This is measured alignment evidence, not by itself consensus or an AI-specific causal finding.
As the process went from individual opinions to initial and then revised group statements, the statements moved toward a compromise between participants’ positions.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The figure caption describes the measured distribution of positions moving toward compromise across these stages. Addresses: Changes in group statements
Inspect the contribution and sourcesImportant limit · Additional groups tested parts of the system and new topics
These tests bear on causal attribution and generalization, but their stated aims alone do not resolve either.
The researchers collected data from three additional groups to test the importance of parts of the AI system, including its reward model, which helps guide the system’s outputs, and supervised fine-tuning, a method of training it with examples. They also checked whether results carried over to discussion topics not used in training.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The supplied passage describes the goals of these analyses, but does not provide their results. Addresses: Additional cohorts and system checks
Inspect the contribution and sourcesStarting point · The study tested AI support for group discussion
The study aim frames empirical questions about common ground but does not prove consensus.
The study asked whether AI could help people with different views produce a group statement they would accept together.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors describe group discussion as difficult to scale and potentially unequal in whose voices are heard. They tested AI mediation as a possible way to help people find common ground. Addresses: The study’s purpose
Inspect the contribution and sourcesImportant limit · The researchers used a question-specific scale to score written opinions
The measurement method and imperfect correspondence limit what text-derived positions establish about alignment.
The researchers represented each written opinion as a set of 768 numbers. They arranged these number sets along a line for each question, from agreement to disagreement. They checked the resulting scores against participants’ earlier ratings. The scores were related, but they did not match exactly.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The article describes how the researchers represented opinions with 768 numbers and placed them on an agreement-to-disagreement line for each question. It also reports how closely those scores matched earlier ratings. Addresses: The method used to interpret participants’ text and compare it with their position ratings.
Inspect the contribution and sourcesSupporting point · Similar findings also appeared in a group of citizens discussing public issues
This adds evidence of endorsement and increased agreement in the studied UK setting, not proof of consensus or AI-specific causation.
The authors report that participants supported the statements at a high rate and that agreement increased in a virtual UK citizens’ assembly, a group of citizens who discuss public issues. They describe its participants as more representative of the UK than those in their earlier groups recruited through crowdsourcing.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The discussion says the findings were replicated in a virtual assembly using a sample described as demographically representative of the UK. Addresses: Replication in a UK citizens’ assembly
Inspect the contribution and sourcesImportant limit · Measuring text positions
The study-specific position measure qualifies interpretation of reported alignment and compromise.
The authors used Sentence T5, an AI model that represents text as numbers, to represent opinions, group statements, and statements expressing agreement or disagreement with each question. They defined the question’s position line using the number representations of the statements expressing agreement and disagreement. They then measured where opinions and group statements fell along that line.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The axis was built from representations of affirming and negating position statements for each question, not from unspecified examples. The authors also note that other text models and axis definitions are possible. Addresses: The text-representation method described in the article.
Inspect the contribution and sourcesImportant limit · Did seeing more majority-leaning statements predict a shift?
This non-causal analysis qualifies explanations for observed shifts; the proposed content explanation is not established.
The researchers measured exposure by counting the share of group statements whose position scores fell on the majority side of the median opinion. They found no reported relationship between this measure and how much participants’ views later shifted toward the majority. The authors infer that what the statements said, rather than how many majority-leaning statements participants saw, may explain the shift.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article counted a statement as majority-leaning when its position score fell on the majority side of the middle opinion. It then tested whether the share of such statements was related to participants’ change in position. Addresses: Whether seeing more majority-leaning statements was linked to later movement toward the majority.
Inspect the contribution and sourcesSupporting point · Overall, support shifted toward the majority view
It reports measured change while distinguishing parallel movement from full agreement or consensus.
Overall, participants’ ratings shifted toward the majority view after discussion, even though minority views could still gain support. The authors call this pattern “parallel belief updating”: people may move in the same direction without all ending up with the same view.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The article reports an overall shift toward the majority view and distinguishes this from full agreement or people reaching the same view. Addresses: The difference between an overall shift toward the majority and everyone reaching the same position.
Inspect the contribution and sourcesStarting point · Two models with separate tasks
The model setup is relevant background for interpreting results but does not itself establish quality, fairness, or causation.
The system has two AI systems, called models. One model was trained further using examples that show the statements it should produce. It generates possible statements. A second model, trained to estimate each person’s preferences, predicts which statements they would prefer. Both were trained further from Chinchilla, an AI system trained to work with language.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This describes the method, not the accuracy of the preference predictions or the quality of the generated statements. Addresses: The model components and training
Inspect the contribution and sourcesImportant limit · Analysis methods
The analytic choices are relevant to assessing reported findings but do not establish that all measurement or inference concerns are resolved.
The authors report registering their procedures for excluding data in advance, so those plans could be checked against the later work. They used statistical models that account for variation among participants and groups. For each analysis, they used the most detailed random-effects structure the available data allowed. They used ordered-response models for rating-scale answers and two-sided tests for reported significance tests.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage identifies analysis choices but does not provide estimates, test results, or evidence that these choices resolve every source of uncertainty. Addresses: The statistical procedures reported in the article.
Inspect the contribution and sourcesImportant limit · Participants reported some movement
The figures qualify broad claims of view change: most rounds were reported unchanged, and the measure does not establish causation.
In reports after discussion rounds, participants said their views moved closer to the group opinion in about 30% of rounds. They said their views moved in some other way in 4.6% of rounds, and reported no change in 60.3%. For their current perspective, they chose the final group opinion in 66.6% of rounds and their original opinion in 33.3%.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These reports give participants’ accounts of how their views related to the group’s view. They are not the same as a direct measure of why views changed, and most rounds were reported as involving no change.
Inspect the contribution and sourcesStarting point · Participants rated and commented on statements
Procedure and view measures are context for interpreting reported agreement and change, not proof of consensus or causation.
Participants privately wrote a short paragraph about their opinion; they typically gave both their position and a reason. The AI helper generated candidate group statements for participants to rate and rank. After the top-ranked statement was critiqued, the AI generated revised candidates for another ranking. Participants compared the initial and revised winners and answered a brief survey about whether the process affected their views. They also rated their positions on the questions immediately before and after the full set of deliberation rounds.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This describes the steps in each round and the separate before-and-after measure of participants’ positions across the full set of rounds. Participants knew an algorithm generated the group statements. Addresses: Study procedure and measurement
Inspect the contribution and sourcesImportant limit · Measuring statements that lean toward the majority
It clarifies how possible bias was measured and limits the inference: the measures alone do not establish absence of bias.
The authors defined majority bias by checking whether generated statements leaned toward the majority or minority side of the median opinion. Its score was the proportion of statements leaning toward the majority. For the virtual assembly, they also assessed agreement bias by comparing a statement’s position with the agree and disagree sides relative to the median opinion.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These are the authors’ stated measures; the passage does not report their values or show that they capture all forms of bias. Addresses: The bias measures described in the article.
Inspect the contribution and sourcesStarting point · Training material
The training setup is methodological context for assessing study results, not performance evidence by itself.
The authors further trained Chinchilla 70B, the name of the language model, an AI system that generates text, on opinions and critiques from earlier rounds, using highly rated group statements as targets. The material came from 578 groups, 2,473 participants, and 1,630 questions.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The counts and model name describe the reported training setup; they do not show how well the model learned or worked. Addresses: The model-training method reported in the article.
Inspect the contribution and sourcesPartly answered · An open question about online assemblies
The virtual-assembly agreement result bears directly on the question, but the omitted fact-finding and expert-testimony stages limit the answer’s scope.
The authors say it was unknown whether AI could help people find agreement in an online citizens’ assembly, a group of citizens who meet to discuss public issues.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This is an unresolved question in the introduction, not a conclusion from the supplied passages. Addresses: The open question motivating the study
Inspect the contribution and sourcesStarting point · Revisions used critiques and earlier views
The revision procedure is context for interpreting preferences for revised statements; it does not by itself establish an effect.
The AI helper created revised group statements using participants’ critiques together with their initial opinions and the initial winning statement. A deliberation round on one question lasted about 15 minutes.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The inputs to revision and the approximate round length describe how the process operated. Addresses: Revision procedure and duration
Inspect the contribution and sourcesImportant limit · Support for the group statement was linked to changes in views
The association is relevant to view change but does not establish that accepting a statement caused participants to change their views.
Participants who more strongly endorsed the final group statement tended to show more influence from other group members’ starting views in their later positions. Participants’ own starting positions also significantly predicted their later positions, meaning a statistical test found evidence of a link.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors interpret the link between statement endorsement and influence from others’ starting views as suggesting that a well-accepted shared statement may help people shift their perspectives. Participants’ own starting positions also had a significant relationship with their later positions. These results show associations, not that statement endorsement or starting positions caused later views.
Inspect the contribution and sourcesImportant limit · Some rounds with a human mediator showed convergence like rounds with an AI mediator
It limits AI-specific causal attribution: convergence may arise from mediation more generally, and the analysis is exploratory.
A preliminary analysis explored the results and found similar movement toward agreement in rounds when a human mediator did better than the AI mediator. This suggests that the mediation process, rather than a feature unique to AI, may help explain the result.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors describe this as an exploratory result: a preliminary analysis that does not settle the question, rather than a firm comparison of human and AI mediation. It leaves open whether the two mediators work in the same way or are equally effective.
Inspect the contribution and sourcesStarting point · Participants rated and ranked statements
The procedure is context for interpreting study results and preferences, not evidence that the process guarantees consensus.
Participants rated each initial statement’s argument quality even if they disagreed with it. Separately, they ranked statements by how much they agreed with them. The winning statement was chosen by ranked choice, a method that uses participants’ rankings, and the same rule for combining group preferences as the AI helper. Participants were told that an algorithm generated the statements.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These details clarify how the researchers gathered participants’ judgments and chose the statement for critique. Addresses: Study procedure
Inspect the contribution and sourcesPartly answered · Does artificial intelligence represent different views fairly?
The disagreement frequency contextualizes the study, but the question of fair representation remains broader than this statistic.
The article raises, but does not answer in these passages, whether the artificial intelligence (AI) system represents majority and minority views fairly. About 77% of the discussion rounds began with disagreement.
Source coverage: The full readable text of the 10-page Science article, including references. Supplementary materials are not included. The authors identify balancing majority and minority views as an important issue and ask whether the AI represents viewpoints equally. The reported share of rounds that began with disagreement shows how often disagreement arose in the main study groups. These passages do not provide the answer.
Inspect the contribution and sourcesImportant limit · The AI’s training goal differed from the change seen in people
The node distinguishes the AI’s training objective from observed change. The reported self-assessments—movement toward the group opinion in about 30% of rounds and no change in 60.3%—further qualify claims about how often participants changed views, without disproving the study’s measured reduction in division.
The AI was trained to produce group statements that received strong support. But the discussion process reduced division at least partly as participants’ views moved toward one another.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The authors distinguish the goal used to train the AI from the change seen during discussion. The reported reduction in division does not mean that changing participants’ views was itself the AI’s training goal.
Inspect the contribution and sourcesSupporting point · Groups considering AI-created statements showed more agreement
The comparison supports a reported difference between mediated discussion and simple exposure, while leaving the causal contribution of individual process features unresolved.
Groups whose members considered statements made by AI from the group’s views became more aligned. Groups that instead viewed one another’s individual opinions did not show a similar change.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The comparison suggests that having AI bring views together contributed beyond simply showing people other participants’ views. It supports the authors’ conclusion that the process reduced division partly as people’s views moved toward one another. It does not identify every part of the process that mattered.
Inspect the contribution and sourcesSupporting point · Most group statements scored within the range of members’ views
The reported score range is evidence about the study’s operationalized compromise measure, not proof of consensus or correctness.
Nearly all group-statement scores fell between the lowest and highest individual scores in that group: 96% were within that range. The researchers interpret this as evidence that the statements represented a compromise among members’ views.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The article reports that 96% of group-statement scores fell between the lowest and highest scores of group members. Addresses: Whether the written group statements stayed within the range of individual views.
Inspect the contribution and sourcesSupporting point · AI-assisted groups became less divided
This is a central reported alignment result, subject to the narrative’s cautions about what measured alignment establishes.
After AI-assisted discussion, participants’ reported positions moved closer together; this did not happen when participants simply exchanged views without help from AI.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports that division decreased after AI-assisted discussion, while direct exchanges without AI did not produce the same result. Addresses: Changes in participants’ reported views
Inspect the contribution and sourcesStarting point · Revising statements from critiques
The iterative critique-and-revision procedure is relevant to interpreting the reported deliberation outcomes, but the feature alone does not demonstrate those outcomes.
The system can use participants’ written critiques to generate and select revised group statements.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This is a system feature; it does not establish that every critique appears in the final text. Addresses: The iterative revision feature
Inspect the contribution and sourcesImportant limit · A test found no clear link with AI bias
The bias analysis qualifies causal interpretations of group-level changes but does not establish that bias played no role.
The researchers tested whether AI statements that favored one side predicted people changing toward that side. They found no clear relationship. This counts against, but does not rule out, the explanation that AI bias drove the changes seen across groups.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article describes a score based on how often statements leaned toward agreement relative to the group’s middle opinion, and reports no clear link with people changing toward that position. Addresses: Possible influence of bias in the language model
Inspect the contribution and sourcesSupporting point · Written opinions matched participants’ ratings
This provides measurement context for interpreting text-based positions and reported changes; it does not establish consensus or settle measurement validity generally.
Participants’ written opinions lined up closely with their ratings of the same positions. This suggests that the AI’s measure, based on text, captured meaningful differences in viewpoints.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The figure caption reports a strong relationship between participants’ ratings and the measure of their positions drawn from their written opinions. Addresses: How the study measured positions.
Inspect the contribution and sourcesStarting point · The study asked whether groups’ views move closer together
It frames the convergence question that the reported alignment evidence addresses, without itself supplying an empirical answer.
The study asked whether people who discuss issues freely would move toward a shared position across groups, as Habermas proposed. This is a research question, not a finding that every group reached agreement.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article frames the study as testing whether discussion with AI led to views moving closer together across groups, connecting this question to Habermas’s proposal. Addresses: The study’s central research question
Inspect the contribution and sourcesImportant limit · Evaluation plans were registered before the studies
Preregistration bears on the study’s evaluation process and helps contextualize its evidence, but does not resolve the limits of its measures or generalization.
The authors report that they registered the planned sample sizes, exclusion criteria, and primary analyses for all evaluation experiments before the evaluations began.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. Passage 38 says the researchers registered these study details before conducting the evaluations. This helps readers check what the researchers planned before conducting the evaluations. Addresses: The article’s evaluation experiments
Inspect the contribution and sourcesSupporting point · Agreement increased after discussions with AI help
This is evidence of measured alignment, qualified against treating agreement as consensus, correctness, or uniform view change.
Agreement within groups increased after discussions supported by AI in all three main study groups. The share of groups with unanimous agreement also rose in the first group.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The researchers measured agreement before and after discussion as the difference between the shares supporting each side. The reported average increase was about eight percentage points across the three groups. This measure shows greater alignment, not that everyone changed their mind or that the final view was correct.
Inspect the contribution and sourcesImportant limit · Critiques may shape revisions
It offers a possible explanation of how revisions relate to common ground, while cautioning against treating that explanation as an established cause.
The authors suggest that an initial statement appealing to the majority may prompt minority members to critique it, leading to a revision that gives minority views more weight. They interpret participants’ preference for revised statements as consistent with finding common ground, not simply switching between sides.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage presents the critique-and-revision sequence as an interpretation of the observed pattern, not as a separately established cause. Addresses: The authors’ proposed explanation for the change between initial and revised statements.
Inspect the contribution and sourcesStarting point · What earlier tools had not tested
The stated research gap provides context for interpreting the study’s reported deliberation evidence; it is not itself evidence that the method succeeds.
The authors say earlier AI discussion tools had not been tested in a thorough, planned way to find out whether they work, how they affect people’s beliefs, or how they affect the meaning and content of discussion.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors distinguish existing prototypes from careful, planned testing; they do not say that no earlier tools existed. Addresses: The stated gap in earlier research
Inspect the contribution and sourcesSupporting point · Groups preferred revised statements
Preference for critique-informed revisions is evidence about the study’s deliberative process, but it measures acceptability rather than consensus or general effectiveness.
Across three main study groups, participants preferred group opinion statements revised to include their critiques over the earlier versions. A separate comparison found that groups also preferred statements from an AI that had seen their critiques over statements from one that had not.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The separate comparison tested whether people favored revised statements simply because they saw them later or had spent extra effort. Preference remained after accounting for statement length, which supports the authors’ view that incorporating critiques helped. These tests measured preferences in the study, not whether the process works in every setting.
Inspect the contribution and sourcesImportant limit · Some issues showed no clear shift
Issue-specific differences limit general claims that deliberation produced a uniform change in views.
The pattern was not uniform: views did not change on Brexit, while opposition to universal free childcare increased. However, the study could not rule out that the childcare rating shift was due to chance.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports no change on Brexit and a shift toward opposition to free childcare that could have been due to chance. Addresses: Limits on the reported shifts
Inspect the contribution and sourcesImportant limit · The analyses found no sign of favoring one political side
The absence of detected bias qualifies one possible explanation for the observed results but does not establish absence of bias generally.
The authors report that their analyses found no sign that the model’s bias caused the group statements to favor one political side. They say the statements reflected the majority view and gave dissenting voices extra weight, while stressing that fairness still needs careful checking.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The reported finding is limited to the analyses described; the authors still call for vigilance in other uses. Addresses: The study’s analysis of group statements
Inspect the contribution and sourcesImportant limit · Support sometimes grew for both majority and minority views
The result shows movement in more than one direction and does not imply unanimous consensus or uniform change.
Across discussion rounds, support grew for majority views on some issues and minority views on others. Support for the view that began with more people grew in 29–32% of rounds, while support for the view that began with fewer people grew in 20–26%.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The authors report that support grew for both majority and minority views, but more rounds saw support grow for the majority view. Addresses: Whether the process allowed both majority and minority views to gain support.
Inspect the contribution and sourcesPartly answered · Questions about disagreement and views held by fewer people
The aim of testing whether mediation reduced division is addressed by supplied reports of increased alignment and view movement; those results do not establish consensus or general effectiveness.
The researchers also planned to examine whether AI mediation, meaning AI helping to guide the discussion, reduced division, suppressed views held by fewer people, and could be used in practice.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These are research aims, not evidence that the system reduced division or protected minority views. Addresses: Additional study aims
Inspect the contribution and sourcesSupporting point · Assembly groups shifted in similar directions
The assembly finding contributes evidence about group-level movement, while the interpretation that deliberation produced shared views remains the authors’ interpretation.
In the virtual citizens’ assembly, groups’ views tended to move in a similar direction on controversial issues. The authors report that AI bias did not explain this pattern and interpret it as deliberation helping shared views emerge.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports similar shifts in groups’ views and says the shifts were not attributable to bias in the AI. The authors interpret the pattern as the emergence of shared perspectives through deliberation. Addresses: Assembly results and the authors’ explanation
Inspect the contribution and sourcesSupporting point · Agreement also rose in the virtual assembly
The assembly result adds evidence of increased agreement in this study setting, but agreement is not equivalent to consensus or universal generalization.
The researchers tested the process in a virtual citizens’ assembly, a group discussion, with 200 UK participants recruited to reflect people living in the UK across age, gender identity, ethnicity, social and economic background, and region. Assembly groups used the same discussion process as groups in the main study. Support for group opinions was similarly high, and group agreement increased significantly from before to after the discussion, as it did in the groups recruited through crowdsourcing.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports the assembly sample, its use of the same discussion process as groups in the main study, and outcomes similar to the groups recruited through crowdsourcing, including a statistically significant increase in group agreement. Addresses: Whether results from earlier participant groups were also seen in a group recruited to reflect people living in the UK across the listed characteristics, using the same discussion process.
Inspect the contribution and sourcesPartly answered · Four questions about AI-guided discussion
The common-ground and reduced-division aims are addressed by reported alignment and view movement, but the question node itself is an agenda of inquiry rather than evidence of success.
The study asks whether AI helps guide discussion so people can find common ground, reduces division, represents different views fairly, and can support a citizens’ assembly.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These are questions to investigate, not findings that the system meets these aims. Addresses: The study’s four research questions
Inspect the contribution and sourcesStarting point · The figure compared statements in several ways
The measures and uncertainty display provide methodological context for evaluating the study’s reported comparisons, without independently supporting their substantive conclusions.
The figure caption describes several ways the study compared AI-generated and human-written group statements, including rankings that favored one over the other. It also compares revised statements that included critiques with initial statements that left them out. The figure’s error bars show 95% confidence intervals, which are estimated ranges of uncertainty calculated by resampling the data.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The caption separates participant rankings and ratings from outside judges’ ratings. It also explains how the figure shows uncertainty. These methods describe the study’s measures. They do not prove that the statements would work well in real public decisions.
Inspect the contribution and sourcesImportant limit · Statements leaning toward the majority did not predict change
It qualifies interpretation of reported movement by finding no evidence that majority-leaning statements predicted movement toward the majority.
The researchers found no evidence that groups moved toward the majority simply because the AI’s statement expressed a position closer to the majority’s view.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The figure caption reports no link between how often the AI produced statements closer to the majority’s view and groups moving in that direction. Addresses: The relationship between the position of statements and changes in group views.
Inspect the contribution and sourcesSupporting point · Participants preferred revised statements
This adds assembly-specific evidence about preference for revised statements and positive ratings, bounded to the reported setting and measures.
In the online assembly, participants preferred revised statements over initial ones. They also rated the group statements positively and judged them to be high quality.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The figure caption reports that participants preferred revised statements in their final judgment. It also says they gave the group statements positive ratings and rated them as high quality. Addresses: Statement preference and quality ratings
Inspect the contribution and sourcesImportant limit · Earlier studies found mixed effects in citizens’ assemblies
Mixed prior findings qualify broad expectations about assembly effects and distinguish prior research from the current study’s results.
The authors report that earlier studies of citizens’ assemblies—groups of citizens who meet to discuss public issues—found mixed effects: views may become more alike or more divided, decisions may improve or worsen, and later political participation may rise or fall.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This summarizes prior studies, not the article’s own experiments. Addresses: The article’s summary of prior research
Inspect the contribution and sourcesSupporting point · Agreement rose with AI-produced group statements, but not with exposure alone
This supplies a comparison relevant to the reported alignment result, while the participant-presence condition and study measures bound the inference.
Group agreement increased after people used group statements produced with AI, and unanimous agreement became more common. In the comparison without AI, where people viewed one another’s opinions, agreement did not increase. The figure’s trend lines are fitted regression lines, which show a pattern in the data, with standard errors. Its error bars show 95% confidence intervals, which are estimated ranges of uncertainty calculated by resampling the data.
Source coverage: The full readable text of the 10-page Science article copy, including references, was available. Supplementary materials were not included. The article reports increased agreement and unanimity after discussion using AI-produced group statements, but no change in agreement in the group that saw others’ opinions without AI. It also describes the figure’s fitted lines and uncertainty ranges. The figure includes only rounds with all five group members present before and after discussion. Addresses: The reported change in agreement, how it was measured, and the comparison with viewing others’ opinions without AI.
Inspect the contribution and sourcesImportant limit · Opinion shifts differed by issue
Issue-level variation and exploratory status limit general claims about consistent view change or consensus.
Exploratory analyses, which the researchers used to examine possible patterns, found shifts in a common direction on five of nine issues. Support for reducing the number of people in prison rose from 60% to 75%, and support for making it easier for asylum seekers to enter the UK rose from 39% to 51%. Smaller reported shifts were 42% to 57% for national pride, 77% to 85% for the minimum wage, and 75% to 80% for net zero.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports the percentage changes for five issues and describes the analyses as exploratory, meaning they examined possible patterns. It reports stronger statistical evidence for the prison and asylum issues, and weaker, less clear evidence for the other three. Addresses: Changes in individual ratings across issues.
Inspect the contribution and sourcesSupporting point · Participants supported AI-mediated statements more often
Adds a study-reported comparison of AI-mediated statements with statements from human mediators and reports less division after AI mediation.
The authors report that participants accepted statements produced through an AI-mediated process more often than statements written by human mediators who were offered an incentive and were not professionals. They also say groups were often less divided after using the AI mediator.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The discussion reports that participants accepted statements produced through an AI-mediated process more often than statements written by human mediators who were offered an incentive and were not professionals. It also says groups were often less divided after interacting with the AI mediator. Addresses: Participant responses to statements produced through an AI-mediated process
Inspect the contribution and sourcesSupporting point · Experiments tested parts of the AI system
Provides context about the study’s evaluation design while the incomplete excerpt does not establish what every experiment found.
The authors say they ran seven evaluation experiments. The main task and ablation experiments, which test a system after removing parts of it, assessed the importance of different components of the Habermas Machine, the system being evaluated. The visible list also names tests of the critique phase, mediation’s effect on belief change, and a human-mediator experiment. The passage cuts off before describing that last experiment or listing all seven.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage states the purpose of the main task and tests that removed system parts to assess their importance, and names other evaluation categories, but cuts off before giving the full list and details. Addresses: The study’s evaluation experiments
Inspect the contribution and sourcesSupporting point · Statements included disagreement and majority views
Adds a reported finding about how group statements reflected majority and minority views.
Support for the majority position increased after discussion, while revised AI statements included critiques from minority participants. An analysis that compared patterns of meaning in the text found that successful statements included disagreement while respecting the majority position.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports increased support for the majority position, the inclusion of minority critiques, and a text-embedding analysis finding that successful statements incorporated dissent while respecting the majority position. Addresses: The content of revised group statements
Inspect the contribution and sources