Complete underlying reasoning (86 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 participants rate AI-written statements compared with human-written ones?
It asks about a bounded comparative study measure; the process difference qualifies what the participant comparison establishes and does not establish general readiness or superiority.
How did participants rate AI-written statements compared with human-written ones?
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 under conditions · What evidence do the authors offer about whether the findings may also apply to other groups and settings?
The question asks what evidence bears on generalization; the UK-based evidence is relevant but does not establish results elsewhere.
What evidence do the authors offer about whether the findings may also apply to other groups and 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 sourcesPartly answered · What conditions do the authors propose for using the system in practice?
The question concerns conditions for use, aligned with the distinction between promising results and demonstrated readiness.
What conditions do the authors propose for using the system in practice?
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 under conditions · Could the method support much larger groups?
The evidence answers the scale question while distinguishing tested group sizes from a conditional, untested possibility.
Could the method support much larger 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 · How did the group discussion work, and how did the study measure participants’ views?
The procedure and measurement provide study context, but do not establish broader deliberative readiness.
How did the group discussion work, and how did the study measure 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 · Did the findings also appear in a UK citizens’ assembly?
The assembly result is relevant evidence, while the separate qualification that fact-finding and expert testimony were omitted limits claims that it was equivalent to a full citizens’ assembly.
Did the findings also appear in a UK citizens’ assembly?
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 practical value do the authors see in the approach, and how fast did it produce statements?
Efficiency does not settle broader quality, scale, or readiness.
What practical value do the authors see in the approach, and how fast did it produce 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 problem was the study designed to address?
The study rationale is relevant context for distinguishing practical constraints from broader readiness.
What problem was the study designed to address?
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 far might the findings apply beyond the people and issues studied?
This concerns limits on generalizing study results and broader readiness.
How far might the findings apply beyond the people and issues studied?
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 outside raters judge the statements?
The ratings concern some quality dimensions, not all quality or readiness.
How did outside raters judge the 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 risks follow from the system’s current limits?
These limits qualify claims that the method is ready or sufficient as a stand-alone process.
What risks follow from the system’s current limits?
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 study did the researchers conduct?
Study type and scope are relevant context for limits on inferring broad performance or readiness.
What kind of study did the researchers conduct?
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 test whether the language model could handle topics unlike its training topics?
The held-out-topic procedure bears on the scope of the study’s evaluation, while the result is not supplied.
How did the authors test whether the language model could handle topics unlike its training topics?
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 do the authors distinguish this work from earlier AI research?
The comparison with earlier work is relevant to the study’s stated contribution and its evidential scope.
How do the authors distinguish this work from earlier AI research?
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 assembly evidence bears on study scope but does not establish equivalence to a full assembly or broad generalization.
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 sourcesAnswered · What earlier AI work motivated this study?
Prior tools provide context for the study’s claimed contribution without establishing broad readiness.
What earlier AI work motivated this study?
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 problems with careful, focused discussion in person might the method reduce?
The evidence distinguishes proposed practical advantages from demonstrated effectiveness.
What problems with careful, focused discussion in person might the method reduce?
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 might be lost when people do not deliberate, or discuss an issue carefully, in person?
This qualification limits the potential advantages of efficient or scalable online deliberation.
What might be lost when people do not deliberate, or discuss an issue carefully, in person?
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?
Evaluation settings help bound what the study’s evidence can establish; this passage identifies settings but omits results.
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 · What should readers know about the human comparison?
Comparison details qualify what can be inferred from ratings of AI versus human statements.
What should readers know about the human comparison?
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 material was used to train the AI mediator’s language model?
The training description helps characterize the studied system and the limits of performance claims.
What material was used to train the AI mediator’s language model?
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 human-mediator procedures differ from the main task?
These task differences qualify comparisons and limit inferences from the evaluation design.
How did the human-mediator procedures differ from the main 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 wider benefits and uses do the authors propose?
The answer distinguishes proposed benefits and applications from demonstrated readiness across settings.
What wider benefits and uses do the authors propose?
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 benefits do the authors hope technology could bring to group deliberation?
The answer describes potential practical benefits, not broad applicability or readiness.
What benefits do the authors hope technology could bring to group deliberation?
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 were participants recruited, and how representative were they?
The different recruitment approaches qualify how broadly the study’s results can be generalized.
How were participants recruited, and how representative were they?
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 steps can reduce risks in careful group discussion without deciding its result in advance?
The cited safeguards bear on limits and conditions for using the system in a broader deliberative process.
What steps can reduce risks in careful group discussion without deciding its result 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 · How did participants respond to statements produced with AI compared with statements from human mediators?
The comparison is evidence about a particular study measure and mediator comparison, not general readiness or equivalence to a full assembly.
How did participants respond to statements produced with AI compared with statements from human mediators?
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 idea inspired the system’s name and approach?
The intellectual inspiration is relevant context for the system’s design, but does not establish its practical performance or equivalence to deliberation.
What idea inspired the system’s name and approach?
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 closely did the virtual assembly match a full citizens’ assembly?
The assembly’s limited scope directly qualifies claims that the study demonstrates equivalence to a full citizens’ assembly.
How closely did the virtual assembly match a full citizens’ assembly?
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 in the study, and how and when were they recruited?
The sample and data-collection details delimit the setting to which the study’s findings pertain.
Who took part in the study, and how and when were they recruited?
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 and outside judges rate the statements?
The ratings are particular study measures and do not establish all dimensions of deliberative quality or broad readiness.
How did participants and outside judges rate the 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 · Which questions did the assembly use, and how were they chosen?
The specified topic-selection and assembly design help define the tested scope; they do not establish readiness or generalization beyond that design.
Which questions did the assembly use, and how were they chosen?
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 mediator estimate which group statements people would prefer?
The reward-model description clarifies one component of the deliberation system while not establishing its quality or performance in real-world use.
How did the AI mediator estimate which group statements people would prefer?
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 researchers test the system beyond the initial comparison?
The additional tests indicate the kinds of robustness and generalization checks undertaken, but absent results they do not establish generalization or readiness.
How did the researchers test the system beyond the initial comparison?
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 statement length explain the preference for AI-written statements?
The length analysis narrows interpretation of a preference comparison and illustrates why favorable ratings do not alone establish all quality dimensions.
Could statement length explain the preference for AI-written 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 did the tested models compare, and what improvement did the authors propose?
The comparison and proposed next step are bounded model-performance evidence, not proof of broad deliberation quality or superiority.
How did the tested models compare, and what improvement did the authors propose?
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 · Where do the authors see possible uses for group statements?
The listed applications illustrate proposed reach but do not establish performance or equivalence across these distinct settings.
Where do the authors see possible uses for 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 · What does earlier research say about groups of citizens who meet to discuss public issues?
The mixed prior findings caution against treating assembly or deliberation outcomes as uniformly positive, relevant to claims of readiness or generalization.
What does earlier research say about groups of citizens who meet to discuss public 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 · How does the system select among candidate statements?
These are concrete system-design details; neither the selection procedure nor its stated purpose establishes fairness or broader performance.
How does the 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 · How did researchers prepare discussion questions?
Question generation and screening are design details relevant to the scope of the tested process.
How did researchers prepare discussion questions?
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 · An ideal of public discussion
The theoretical inspiration does not establish that the implemented process realizes ideal deliberative conditions.
The system is named the Habermas Machine after Jürgen Habermas. The authors draw on his proposal that agreement can emerge when people who use reason discuss matters under ideal conditions.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This describes the inspiration, not evidence that real-world discussion meets those ideal conditions. Addresses: The authors’ stated intellectual inspiration
Inspect the contribution and sourcesStarting point · The AI revised statements using feedback
The described implementation is narrower than evidence of broad readiness or equivalence to a full assembly.
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 · Choosing a group statement
The selection procedure specifies what was tested, without establishing readiness beyond the studied design.
The mediator generated 16 candidate statements. For each participant, a reward model—a model that uses their opinion and critique to predict how they would rank the statements—predicted a ranking. The system combined the predicted rankings using the Schulze method, a ranked-choice voting method that compares how often each statement is ranked ahead of another, and returned the winning statement; the authors repeated the process to provide multiple samples in the main evaluations.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage reports a simulated election among generated candidates. It does not establish that the winner represented every participant fairly or accurately. Addresses: The statement-selection procedure described in the article.
Inspect the contribution and sourcesImportant limit · The study examined groups in the United Kingdom
The UK setting and virtual assembly do not establish equivalence to a full assembly or broader generalization.
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 sourcesImportant limit · Additional groups tested parts of the system and new topics
A test purpose concerning unfamiliar topics does not demonstrate performance beyond the reported setting.
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 stated test does not establish broad readiness, large-scale performance, or equivalence to a full assembly.
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 sourcesSupporting point · Participants favored statements written by the AI helper
Preference ratings inform practical evaluation but do not establish all quality dimensions or readiness.
In 75 groups of six, participants chose statements from the AI helper more often than statements written by trained human discussion leaders. This happened in 56% of comparisons. Participants also rated the AI statements as having more group support and better arguments.
The reported comparison involved 439 participants and trained human discussion leaders who were paid to take part. The human leaders did not have a stage where participants criticized the statement, so the two processes were not identical. Addresses: Participant preferences and ratings
Inspect the contribution and sourcesImportant limit · Similar findings also appeared in a group of citizens discussing public issues
A virtual assembly is not thereby equivalent to a full citizens’ assembly or evidence of generalization elsewhere.
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 sourcesObjection · The system cannot check claims or guide discussion
These limits challenge any inference that speed or favorable ratings establish readiness for broader deliberative use.
In its current form, the Habermas Machine cannot check whether claims are true, keep discussion on topic, or guide what people say. If participants share poorly informed or harmful opinions, the system may produce a poorly informed or harmful statement.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors tie the risk of harmful output to the opinions people provide and to functions the system currently lacks. Addresses: The system’s ability to manage discussion content
Inspect the contribution and sourcesImportant limit · The authors see promise for wider use
The claims about scalability and wider use describe potential applications, not readiness or performance in every listed setting.
The authors describe the approach as time-efficient, fair, able to serve larger groups, and better than human mediators on key measures. They say it may help people find agreement and support collective action in areas such as negotiations, conflict resolution, political discussion, and citizens’ assemblies.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The conclusion describes these benefits and possible uses, while presenting the approach as potential for collective decision-making rather than reporting tests in every listed setting. Addresses: The authors’ assessment and suggested applications
Inspect the contribution and sourcesImportant limit · The assembly left out fact-finding stages
The assembly’s omissions limit claims that results establish performance or equivalence in a full citizens’ assembly.
The assembly mainly copied the discussion phase of a citizens’ assembly. It did not include the usual stages for gathering facts or hearing expert testimony.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article explicitly describes the assembly as mainly mimicking the discussion phase and notes which usual stages it left out. Addresses: The limits of the assembly design compared with a typical citizens’ assembly.
Inspect the contribution and sourcesSupporting point · Learning preferences
The model-training description supplies methodological context for assessing the system, not evidence that it is ready or generalizes broadly.
The authors trained a 1.4-billion-parameter Chinchilla reward model. A parameter is a value the model adjusts as it learns, and a reward model gives a score to show how much a person may agree with a statement. A linear layer, a simple calculation added to the model, produced a score for each candidate statement. The model learned from participants’ statement rankings and quality ratings; its scores were designed to reflect how much each participant was expected to agree with a statement.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage describes the model design and training signals, but does not show how accurately it predicted people’s preferences. Addresses: The reward-model training method reported in the article.
Inspect the contribution and sourcesImportant limit · Possible gains over in-person discussion
The proposed advantages are qualified by their status as possibilities rather than demonstrated performance across real-world settings.
The authors argue that the method may avoid some drawbacks of group discussion, including cost, limited scale, possible bias by a human discussion leader, pressure to give socially acceptable answers, and unequal contributions.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These are proposed advantages, not proof that the method removes these problems in practice. Addresses: The authors’ comparison with group discussion
Inspect the contribution and sourcesImportant limit · Similar results in another UK sample
The additional sample and adjustment support cautious generalization within the studied context, while the source explicitly leaves other settings unestablished.
The authors report similar results in a UK online citizens’ assembly with a sample they describe as demographically representative, meaning its demographic makeup reflected that of the people it represented. They also report similar results after adjusting findings from convenience samples—groups recruited because their members are easy to reach—for possible differences between the sample’s characteristics and those of the people it represents. They say this makes it seem likely that the findings may apply more widely, though it does not establish results in other countries or settings.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The added checks concern UK participants and do not remove the uncertainty about other contexts. Addresses: The additional UK group and adjustment for differences in participant characteristics
Inspect the contribution and sourcesStarting point · Testing whether results apply to a broader group
The named evaluation settings are relevant context for judging study scope; the statement itself supplies no outcomes or broad generalization.
The authors compared the AI mediator with human mediation and used a virtual citizens’ assembly with a sample designed to reflect the demographic makeup of a wider group of people. They used it to test whether results extended to a broader group of people.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage names these evaluation settings but does not give their results or specify here which parts of the wider group the sample was designed to reflect. Addresses: The study design described in the article.
Inspect the contribution and sourcesStarting point · The system predicted rankings to choose a statement
The selection procedure is relevant methodological context for evaluating the deliberation design, but the supplied evidence does not establish readiness for wider use.
The study used a process like an election to choose a group statement. A generative model, which creates new text, produced possible statements. A reward model tailored to each group member predicted how that member would rank them. The system combined those predicted rankings to select a statement.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The method description explains how candidates were created and ranked. It describes the system used in the study, not proof that this selection method is best for public decisions. Addresses: Statement selection method
Inspect the contribution and sourcesImportant limit · Experiments with 5,734 participants
The sample size describes study scope but does not establish broad generalization or large-scale performance.
The researchers report structured experiments involving 5,734 human participants in total.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The total describes the experiments’ scope, not proof that findings apply to everyone. Addresses: The overall method and participant count
Inspect the contribution and sourcesImportant limit · AI statements were longer on average
The length effect qualifies interpretation of statement comparisons; the remaining endorsement result does not establish every other quality dimension or practical readiness.
AI-written statements averaged 125.9 words (95% confidence interval: a range that reflects uncertainty in the estimate, from 121.4 to 130.1), compared with 105 words for human-written statements (100.2 to 108.5). Length affected ratings and rankings. However, participants still supported the AI statements more in their endorsement ratings after the researchers accounted for length.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage identifies length as a possible explanation, reports the average lengths, and says the endorsement difference remained after controlling for length. It does not say length had no effect on other ratings or rankings. Addresses: Effect of statement length
Inspect the contribution and sourcesImportant limit · Outside raters saw strengths and limits
The ratings are bounded assessments of particular statement qualities and do not demonstrate broad readiness or performance in real-world assemblies.
A separate group of 155 outside raters judged AI-generated statements clearer and more informative, less illogical, better at capturing the majority perspective, and less reflective of the individual writer’s own opinion than human-written statements.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The same raters found no statistical difference on whether the statements fairly summarized opinions, included minority views, or were seen as likely to divide opinion. Addresses: External ratings of statement quality
Inspect the contribution and sourcesSupporting point · People rated AI statements more highly
The favorable ratings are relevant to practical promise, but do not establish all dimensions of deliberative quality or readiness for broader use.
Participants preferred statements written by AI to statements written by human mediators. They rated the AI statements as more unbiased, while outside reviewers rated them higher for quality, clarity, informativeness, and perceived fairness.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article reports participant preferences and ratings, as well as separate ratings by external judges on several stated qualities. Addresses: Ratings of AI and human statements
Inspect the contribution and sourcesImportant limit · Larger groups remain untested
The small tested group size limits extrapolation from practical performance to large-scale deliberation; scaling remains untested.
The authors say the method could, in theory, support groups of hundreds if paired with language models that can handle much longer texts. They tested groups only up to five people.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The larger-group claim is a theoretical possibility, not a tested result. Addresses: The method’s possible scale
Inspect the contribution and sourcesImportant limit · Different task procedures
Differences between task procedures qualify comparisons across conditions and help define the scope of the reported evidence.
The human-mediator tasks and the tasks where participants saw other people’s opinions left out the critique stage; the human-mediator task instead used four rounds and had outside raters assess transcript quality.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These differences apply to the named tasks and should not be treated as features of every study condition. Addresses: The human-task procedure described in the article.
Inspect the contribution and sourcesPartly answered · Questions about disagreement and views held by fewer people
The practical-use aim relates to readiness, while the supplied evidence leaves scale and broader real-world performance limited or untested.
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 sourcesStarting point · Groups shared topics chosen as important
The selected topics and three-session schedule specify the study design and delimit the setting to which its results directly apply.
All groups in the virtual assembly discussed the same nine questions, chosen because participants had rated the topics as highly important during training. The assembly took place over three weekly sessions, each lasting one hour.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The article says every group discussed the same nine questions, gives the reason for selecting the topics, and states that the assembly had three weekly one-hour sessions. Addresses: The shared topic design, stated basis for selecting the assembly’s questions, and session schedule.
Inspect the contribution and sourcesImportant limit · A shared risk in deliberation
The comparison places a limitation of the system in the context of democratic deliberation more broadly, without resolving how to mitigate it.
The authors say the possibility of producing an ill-informed or harmful output from ill-informed or harmful human opinions is shared with other democratic processes. They connect this comparison to the principle that an agreement reached through free and fair discussion among citizens can be considered acceptable or justified.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage explicitly says this feature is shared with other democratic processes and links it to free and fair citizen deliberation. Addresses: The authors’ comparison with wider democratic deliberation
Inspect the contribution and sourcesImportant limit · Potential benefits of the approach
The proposed benefits of efficiency, fairness, and scale are relevant to design, but the node does not establish all are demonstrated.
The authors argue that technology helping groups find agreement could be valuable if it is time-efficient, fair, and able to serve larger groups. This is their view of its potential, not a finding that all three benefits have been demonstrated.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors present these qualities as reasons such a technology could be beneficial, rather than as a demonstrated result in every setting. Addresses: The authors’ description of the potential benefits of technology for finding agreement
Inspect the contribution and sourcesImportant limit · Wider relevance remains uncertain
This directly limits generalization beyond the studied UK participants and issues, without asserting that findings cannot transfer.
The study mainly involved UK participants discussing issues important to the UK. The authors say they do not know whether the findings apply elsewhere, though they had no reason to think the findings arose from features unique to the UK context.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors state both a limit on what their study of UK participants can establish and their reason for not attributing the results to features unique to the UK context. Addresses: The people studied and the study setting
Inspect the contribution and sourcesStarting point · Policy questions about voting age and the National Health Service
The topics define the particular policy setting for the reported deliberation evidence; they do not establish effects on other topics.
Participants discussed potentially divisive UK policy questions, including whether to lower the voting age to 16 and whether to transfer the National Health Service from public to private ownership or operation.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These examples show the study’s policy setting, not the views participants held. Addresses: The deliberation topics
Inspect the contribution and sourcesStarting point · Choosing among similar statements
The selection rule is a relevant design feature, but its intended function is not evidence that vote splitting was prevented in practice.
The system combines participants’ rankings using a voting rule intended to stop support for one option from being split between very similar statements.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This is a design choice intended to stop very similar options from dividing support; its effects are not evaluated in the supplied text. Addresses: The selection rule used by the system
Inspect the contribution and sourcesImportant limit · Expert input and carefully chosen questions may reduce discussion risks
These safeguards qualify the system’s limitations, but the statement presents them as possible measures rather than proving they eliminate risk.
The authors say discussions that inform democratic decisions may also use information from experts to reduce risks without deciding the debate in advance. They describe using questions chosen to reduce harmful or unproductive discussion.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage describes information from experts as a possible safeguard and the question design as one measure used in this project. Addresses: Information from experts and the study’s discussion questions
Inspect the contribution and sourcesImportant limit · Two ways of selecting UK participants
The recruitment distinction constrains what the total sample size establishes about representativeness and generalization.
Most participants were adult UK residents recruited online from people who were readily available to take part. A separate online citizens’ assembly used a UK sample selected at random in groups based on demographic traits. This method was intended to produce a sample 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 two ways of recruiting participants differ. The assembly’s selection procedure was intended to produce a sample reflecting the UK population’s demographic traits, but the passage does not establish whether the final sample did so. Addresses: Study samples and limits on generalizing
Inspect the contribution and sourcesStarting point · Human deliberation task
The task procedure helps specify what the study tested and measured; the node itself makes no outcome claim.
Groups of up to five people rated their initial positions and how important each question was to them. They discussed three questions, then rated their positions and question importance again.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage describes the task stages and activities, but does not establish how the task affected participants’ views. Addresses: The human task procedure described in the article.
Inspect the contribution and sourcesImportant limit · Agreement also rose in the virtual assembly
A virtual assembly extends the tested setting beyond small crowd-sourced groups but does not establish performance in a full citizens’ assembly or other contexts.
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 sourcesSupporting point · The system produced statements quickly
Fast statement production supports a limited claim of practical efficiency, but does not establish large-scale performance or overall readiness.
The authors describe the approach’s practical value as helping groups discuss issues efficiently, fairly, and at a larger scale, rather than making the AI a better mediator than people. In the study, the system produced high-quality group statements within seconds, while human mediators took several minutes.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The discussion identifies time savings, fairness, and use with larger groups as possible benefits. It reports that the system produced high-quality statements within seconds, compared with several minutes for human mediators. Addresses: The stated practical value and timing of statement production
Inspect the contribution and sourcesImportant limit · UK adults took part in 2023
The study’s participant groups and UK setting delimit the scope of evidence about the deliberation system.
The study included adult UK residents recruited through an online platform where people take part in research tasks, and a virtual citizens’ assembly recruited by the Sortition Foundation. For the assembly, the researchers selected people at random from groups based on factors such as age and gender, to match the UK by age, gender, income, ethnicity, and region. The authors report 2,473 training participants and 2,673 evaluation participants, with data collected from January to August 2023.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These details describe the people and recruitment methods covered by the study; the passage does not provide the full demographic results. Addresses: The study’s participant groups and data collection
Inspect the contribution and sourcesImportant limit · Many settings need shared statements
The listed applications extend beyond the study’s tested setting, so they should be read as possible uses rather than established performance.
The authors say many situations require groups to agree on written statements, including contracts, conflict resolution, jury discussions, diplomacy, constitutional work, artistic collaboration, political debate, and citizens’ assemblies.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. These are possible applications offered by the authors, not settings in which this study tested the system. Addresses: The authors’ examples of settings that need shared written statements
Inspect the contribution and sourcesImportant limit · The system used a ranked voting method to choose a statement
This is a procedural design detail, with the supplied account explicitly incomplete; it does not demonstrate quality or readiness for use.
The system chose a group statement from several options by simulating ranked-choice voting, a method in which voters rank options.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage says the system chose a statement from several options by simulating ranked-choice voting, a method in which voters rank options, but the supplied text cuts off before giving further details. Addresses: Choosing a candidate statement
Inspect the contribution and sourcesImportant limit · Human mediators had training and incentives
The comparison’s human task and incentives are relevant context for interpreting performance comparisons; they caution against reading them as a comparison of identical processes.
In the human comparison, one participant in each six-person group was randomly assigned to be the mediator. Mediators received training and were offered a financial bonus when their statement was the group’s most preferred. The study’s authors said the results showed that writing a statement the group prefers is a challenging task for humans.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The assignment and incentive help describe the comparison, while the authors’ conclusion identifies the task being compared as difficult for human mediators. Addresses: Human mediator comparison
Inspect the contribution and sourcesStarting point · A typical discussion session
The reported typical setup helps delimit the scale and format actually used in the study.
Groups typically had five people and discussed three questions in sequence over about an hour.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage describes the usual setup; it does not say every group had exactly five participants or the same schedule. Addresses: The group discussion format
Inspect the contribution and sourcesImportant limit · Authors claim a first demonstration of AI-guided discussion with many participants
The qualified contribution claim is relevant to the study’s scope, but is not evidence that the system is equivalent to a full assembly or tested across the proposed settings.
The authors say their approach builds on earlier AI systems whose methods were more manually designed. Those systems could debate, summarize arguments, or support discussion about decisions. They say their study is the first demonstration they know of in which AI guides groups through such discussion with many participants.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The authors compare their approach with earlier systems whose methods were more manually designed. Their claim to be first concerns a demonstration of AI guiding group discussion with many participants, and is limited to what they know.
Inspect the contribution and sourcesImportant limit · Testing new topics
The reported test design is relevant to scope and generalization, but without its result it cannot support a claim of performance on unfamiliar topics.
To test performance on unfamiliar subject areas, the authors held 12 topics out of the language-model fine-tuning data.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The passage says the topics were held out for this test but does not report the test’s results. Addresses: The reported test of generalization to new topics.
Inspect the contribution and sourcesStarting point · A private group mediator
Specifies the study’s mediation design, relevant to assessing what the deliberative system does and what claims about it can establish.
The researchers developed an AI mediator that privately collects each participant’s views before proposing a statement intended to be acceptable to the group. They call this private, individual exchange caucus mediation.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. The AI is described as helping write a group statement, not choosing policy or governing the group. Addresses: The proposed mediation approach
Inspect the contribution and sourcesImportant limit · Some research materials are public
Material availability can help readers inspect the study, but the passage does not say all supplementary materials are included or establish readiness for broader use.
The authors report that the question and human-feedback dataset used to train and evaluate the Habermas Machine is publicly available, and that the full set of seed questions is in the supplementary materials.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. Passage 38 identifies where readers can find the dataset and seed questions. This makes it easier to inspect or reuse those materials, but the passage does not say that every supplementary research material is included in the dataset. Addresses: The article’s training and evaluation materials
Inspect the contribution and sourcesImportant limit · Some benefits may be lost
Limits the inference that efficient or scalable online deliberation can simply replace in-person discussion.
The authors note that the method may lose benefits of meeting in person, such as nonverbal cues and the chance to build relationships among participants.
Source coverage: Full readable text of the 10-page Science article copy, including references. Supplementary materials are not included. This qualification limits the authors’ case for replacing or avoiding in-person meetings. Addresses: The potential advantages of face-to-face discussion
Inspect the contribution and sourcesAnswered · Is it easy to understand and accept how a group discussion leads to a decision?
The process-transparency concern limits inference from efficiency or reach to readiness; the answer does not assess a particular system’s performance.
Is it easy to understand and accept how a group discussion leads to a decision?
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 · Discussion can take too long
It highlights a practical cost of deliberation, without establishing that faster alternatives preserve deliberative quality.
Discussion takes time, and decisions may need to be made before it can finish, especially when people are socially far apart.
Source coverage: Complete Augmented Deliberation chapter from the book that is updated over time. Other chapters are not included. The chapter says social distance can slow discussion and that other ways of working together may be needed to make decisions in time. Addresses: Relying on discussion to finish every decision
Inspect the contribution and sources