Contribution to the handbook
Priority · Assessed 2026-10-28
The question brings together comparisons and uncertainty measures that help distinguish participant judgments from outside ratings and assess how much confidence to place in the findings. This is important for evaluating claims about deliberative systems.
Does this challenge the goal?
In-scope gap
Comparisons, alternative measurements, and uncertainty are relevant evaluation details within the handbook’s goal of testing proposed systems. The question does not contest the goal.
The packet describes study methods and some results but does not establish how these measures would perform in governance settings.
Why this category?
- Outside raters found strengths on some statement qualities but no statistical difference on fair summary, minority views, or polarization, preserving material qualifications to favorable ratings (s-habermas-machine-91582003f91ea4b6).
- Participant judgments included separate ratings and rankings, while statement length affected ratings and rankings; endorsement differences remained after controlling for length (s-habermas-machine-532fb8ce5a171497; s-habermas-machine-843a8b534697c360).
- The reported figure captions describe bootstrapped 95% confidence intervals, and the analysis account identifies modeling choices while noting that supplied model outputs are incomplete (s-habermas-machine-b5bd813c4ccc255f; s-habermas-machine-31d2bacac15652f3).
- Text-position measures depend on model and axis choices, and the supplied evidence reports an imperfect relationship with earlier ratings (s-habermas-machine-1ff26c006bc24bba; s-habermas-machine-16cc1d48f1a0c035).
The records are AI interpretations of the same article; they are not independent confirmations. Some uncertainty intervals are described, but the packet does not provide all outputs and details needed to evaluate every comparison.
This is an AI assessment of contribution to the goal, separate from whether the question has an answer.
- How did the study compare group statements and show uncertainty in its results?
- How did outside raters judge the statements?
- How did the authors measure the position expressed in a statement?
- How did participants and outside judges rate the statements?
- Did the group statements fall between members’ positions?
- How did the authors analyze participants’ ratings and other results?
- Could statement length explain the preference for AI-written statements?
- Measuring text positions
- Analysis methods
- Participants rated and ranked statements
- AI statements were longer on average
- Outside raters saw strengths and limits
- The figure compared statements in several ways
- People rated AI statements more highly
Tests, stop conditions and unresolved work →
Authority, accountability and resilience →
Read the exploration goalNew evidence is being assessed. Any answer below is provisional; parts whose evidence changed or was withdrawn are withheld.
Identify which kinds of statements and judgments were compared, including participant and outside-judge measures where applicable.
What we can say so far
The study compared participants’ judgments of initial AI-generated group statements, revised statements incorporating their critiques, and statements generated without access to those critiques. Participants rated argument quality separately from ranking statements by agreement; they also expressed preferences and rated endorsement, quality, and perceived bias. Another comparison set AI-generated statements against human-mediator statements. A separate group of 155 outside raters compared AI- and human-written statements for clarity, informativeness, logic, quality, majority-perspective capture, and how much they reflected the writer’s own opinion. They also assessed fair summary, inclusion of minority views, and polarization: the study reported no statistical difference on those three measures, rather than an AI advantage. The researchers additionally compared numerical position scores for group statements with the range of members’ individual scores.
What remains open
- The supplied caption account does not give detailed outcomes for every rating measure.
Conditions and limits
- These are reported study comparisons, not evidence that the statements would perform similarly in public decisions.
- The supplied accounts come from the same article and are not independent corroboration.
- AI-written statements were longer on average; length affected ratings and rankings, although the reported participant endorsement preference remained after controlling for length.
- A reported absence of a statistical difference does not establish that statements were equivalent.
Objections to consider
- Length affected ratings and rankings. The reported length-adjusted endorsement preference does not establish how adjustment affected every other measure; the answer therefore describes the judgments without attributing all differences to statement authorship.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
Outside raters saw strengths and limits 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.
Identifies outside-rater measures, their reported directions, and the three no-difference findings.
Describe how uncertainty was displayed for the reported comparisons.
What we can say so far
The figure displayed uncertainty in its group-statement comparisons with error bars representing bootstrapped 95% confidence intervals. The reported AI-versus-human statement-length averages also had 95% confidence intervals. In a separate agreement figure, fitted regression trend lines were shown with standard errors.
What remains open
No material gap is recorded within this part’s current scope.
Conditions and limits
- This describes the article’s reported figures and estimates; the supplied coverage excludes supplementary materials.
- The regression trend lines concern agreement, not the group-statement comparison figure.
Evidence and related ideas
Supporting evidence, qualifications and context play different roles in this answer.
supports
The figure compared statements in several ways 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 bootstrapped 95% confidence intervals.
Identifies the statement-comparison figure’s error bars as bootstrapped 95% confidence intervals.
AI can help humans find common ground in democratic deliberation
Read full source: AI can help humans find common ground in democratic deliberation (opens in a new tab) ↗Read source discussion: AI can help humans find common ground in democratic deliberation →
supports
Agreement rose with AI mediation, not exposure alone Group agreement increased after people used AI-produced group statements, and unanimous agreement became more common. In the unmediated comparison, where people viewed one another’s opinions, agreement did not increase. The figure’s trend lines are fitted regressions with standard errors, and its error bars show bootstrapped 95% confidence intervals.
Describes standard errors on fitted regression trend lines in the separate agreement figure.
AI can help humans find common ground in democratic deliberation