AI as a deliberation aid: evidence from Polis to the Habermas Machine

Deliberative democracy has always faced a problem of scale: careful discussion works in rooms of a few dozen people, while political communities number in millions. Software and, since 2022, large language models (LLMs) promise to ease that constraint by clustering opinions, moderating discussions, summarising arguments and drafting statements that many people can accept. This study reviews the main real-world deployments and controlled experiments: the Polis platform and Taiwan's vTaiwan process; the Alignment Assemblies run by the Collective Intelligence Project, including Collective Constitutional AI with Anthropic; Stanford's Online Deliberation Platform and its automated moderator; and Google DeepMind's "Habermas Machine", reported in Science in 2024. It then sets these against experimental evidence on the risks: that AI shifts opinions, persuades more effectively than people, and narrows the range of ideas. The overall picture is of tools that measurably help people find and state common ground, used so far mostly in consultative settings, with strong evidence that the same capabilities could be used to steer as well as to aid. The design question is therefore not whether AI can support deliberation but who controls the tool, what it is allowed to do, and how its influence is audited.

Key findings

• AI can draft statements that groups prefer. In experiments with 5,734 participants, statements produced by an AI mediator were preferred to those written by human mediators and rated as more informative, clear and unbiased; participants often updated their views towards a shared position, and the result replicated in a virtual citizens' assembly drawn from a demographically representative UK sample (Tessler and colleagues, 2024).
• Opinion-mapping tools have shaped real policy, mainly as consultation. Polis, used in Taiwan's vTaiwan process since 2014, clusters participants by how they vote on short statements and highlights statements that win support across clusters. It has informed legislation and regulation, notably in the UberX case, but its outputs remain advisory.
• Public input can change model behaviour without hurting capability. About 1,000 Americans contributed 1,127 statements and 38,252 votes to a "public constitution" for an Anthropic model; the resulting model showed lower bias on nine social dimensions than a model trained on Anthropic's own constitution, with equivalent performance on capability benchmarks (Huang and colleagues, 2024).
• Automated moderation is already used at scale. Stanford's Online Deliberation Platform manages speaking queues, timed agendas and incivility, and reports more than 50,000 unique users in more than 40 countries; it hosted, among others, Meta's 2023 Community Forum (1,545 participants in four countries) and a 2024 Taiwan deliberation on AI and information integrity (447 participants).
• The same tools can steer. Writing with an opinionated language model shifted participants' own views (N=1,506); personalised GPT-4 debaters were more persuasive than human ones in 64.4% of non-tied cases (N=900); and access to AI ideas made creative work individually better but collectively more similar.
• Evidence on legitimacy is thin. Almost all studies measure participants' ratings and attitude change in experiments or one-off consultations. There is little evidence on whether AI-aided processes are seen as legitimate by non-participants, or on their effect on actual decisions.

Research question and scope

The study asks: what does the evidence show about AI tools as aids to democratic deliberation, and what risks does it identify? "Deliberation aid" covers four functions: (1) mapping opinions in a large population (Polis); (2) facilitating discussion (automated moderation); (3) synthesising views into summaries or group statements (LLM summarisation, the Habermas Machine); and (4) aggregating public input into rules for AI systems themselves (Collective Constitutional AI, Alignment Assemblies). AI used to make decisions on citizens' behalf, AI candidates and AI in election administration are outside the scope.

Background

Deliberative theory, from Jürgen Habermas's account of communicative action to James Fishkin's deliberative polling, holds that the legitimacy of collective decisions depends on the quality of the reasoning that precedes them: inclusion, equal voice, reasons rather than power, and openness to revision. Its institutional forms (citizens' assemblies, juries and deliberative polls) rely on random selection and trained human facilitators, and are expensive per participant. Online consultation, by contrast, is cheap but tends to reward the loudest voices and attract trolling.

Computational tools entered this gap in stages. Polis, an open-source "wikisurvey" developed in the early 2010s, avoided threaded replies entirely: participants submit short statements and vote agree, disagree or pass on others' statements, and the software uses dimensionality reduction and clustering to show the structure of opinion. Taiwan's civic-technology community adopted it for vTaiwan, a government-linked consultation process. Since 2022, LLMs have added the ability to read and write natural language at scale, which allows summarisation, moderation by text, and the drafting of candidate consensus statements.

Method and sources

This is a structured review of peer-reviewed experiments, preprints and primary documentation from the organisations running the processes. The main sources are: Small and colleagues on Polis (2021) and on LLMs in Polis (2023); Hsiao and colleagues on vTaiwan (2018); the Collective Intelligence Project's description of its 2023 Alignment Assemblies; Huang and colleagues on Collective Constitutional AI (FAccT 2024) and the accompanying Anthropic report; Stanford's Deliberative Democracy Lab and Human-Centered AI Institute on the Online Deliberation Platform; Tessler and colleagues in Science (2024); Argyle and colleagues in PNAS (2023); and, on risks, Jakesch and colleagues (CHI 2023), Costello, Pennycook and Rand (Science 2024), Hackenburg and Margetts (PNAS 2024), Salvi and colleagues (Nature Human Behaviour 2025) and Doshi and Hauser (Science Advances 2024). Where the organisations running a process are also its evaluators, the text notes it.

Findings

Mapping opinion: Polis and vTaiwan

Small and colleagues describe Polis as an open-source web application "capable of collecting and synthesizing feedback from people in a scalable and distributed fashion" and report that it has built shared understanding, discouraged trolling and surfaced points of consensus, in journalism, academic research and decision-making bodies at local and national level. The design choice that matters most is the absence of replies: participants cannot attack one another directly, and the software rewards statements that attract agreement across opinion groups rather than within one.

vTaiwan, launched in 2014, combined Polis with open meetings, livestreamed consultations and officials' participation. Hsiao, Lin, Tang, Narayanan and Sarahe documented its process, from proposal and opinion-gathering through reflection to legislation, and used the 2015 UberX consultation as a case study in which a "rough consensus" among drivers, taxi operators, passengers and officials fed into regulation. Audrey Tang, a participant in vTaiwan and later Taiwan's digital minister, became the best-known advocate of the approach.

Two limits are well documented in the literature on vTaiwan. Its outcomes depended on ministries choosing to take part and to act; the process had no binding force. And participation was self-selected: it was open to anyone, which is different from representative of everyone.

Facilitating discussion: automated moderation

Stanford's Deliberative Democracy Lab and Crowdsourced Democracy Team built the Online Deliberation Platform for video discussions in small groups. As described by Stanford HAI in 2020, its automated moderator manages a queue of participants who want to speak, limits each turn to 45 seconds, runs a timed agenda and asks the group whether to continue or move on, prompts quiet participants, and transcribes speech, scoring it for toxicity with Google's Perspective API; if a speaker crosses a threshold, the other participants are polled, and if most agree, the speaker is muted for a few minutes. The developers acknowledged that it cannot read body language or sense sensitive topics as a human moderator can.

The platform's own page reported in 2026 that it operates in more than 40 languages and has been used in more than 40 countries, with more than 50,000 unique users and more than 100,000 hours of deliberation. Examples include:

• Meta Community Forum on generative AI, 28-29 October 2023: 1,545 participants from scientifically sampled groups in Brazil, Germany, Spain and the United States deliberated on generative AI; a further forum in October 2024 addressed AI agents.
• Taiwan Deliberative Poll on using AI to enhance information integrity, 23 March 2024: 447 Taiwanese citizens, organised by the Ministry of Digital Affairs, Stanford's Deliberative Democracy Lab and National Yang Ming Chiao Tung University, discussed platform self-regulation versus government supervision and how large platforms should use AI against false information and deepfakes.

The platform's developers have published one evaluation. Gelauff and colleagues analysed surveys and participation metrics from a Japanese Deliberative Poll run on the platform and report that it performed on par with an earlier in-person, human-moderated deliberation on a very similar topic, with preliminary evidence of more equal participation by gender and similar equity across income and education. The comparison is between two separate events rather than a randomised trial, and it was carried out by the team that built the tool.

A related experiment by Argyle and colleagues used an LLM to suggest, in real time, rephrasings of messages in online conversations about a divisive topic, intended to make the partner feel understood. The interventions improved reported conversation quality and "democratic reciprocity" and improved tone, "without systematically changing the content of the conversation or moving people's policy attitudes". This is a narrower aid than a moderator, but it shows that light-touch AI interventions can change how people talk without changing what they think.

Synthesising views: the Habermas Machine

Tessler and colleagues at Google DeepMind trained an LLM-based system, which they called the Habermas Machine, to mediate group deliberation in the United Kingdom. Participants wrote their opinions on a social or political question; the AI drafted a group statement; participants critiqued it; the AI produced revised statements; and participants ranked the results. In experiments with 5,734 participants, people preferred the AI-generated statements to those written by human mediators, rating them as more informative, clear and unbiased. Discussants often updated their views after deliberation, converging on a shared perspective. Analysis of text embeddings showed that successful statements incorporated dissenting voices while respecting the majority position. The findings were replicated in a virtual citizens' assembly with a demographically representative sample of the UK population.

The study is the strongest controlled evidence that AI can perform the "common ground" function of a facilitator. Its limits are also important. The process is mediated in writing rather than spoken; participants did not deliberate with each other directly; the system was optimised for statements groups would endorse, which is not the same as statements that are true or wise; and the convergence it produces could, in other hands, become a tool for manufacturing consent.

Aggregating input for AI itself: Alignment Assemblies and Collective Constitutional AI

The Collective Intelligence Project (CIP) ran a series of "Alignment Assemblies" in 2023 to bring public input into decisions about AI: a pilot around the Summit for Democracy (March 2023); a risk-ranking exercise with OpenAI (June 2023); a process with Taiwan's Ministry of Digital Affairs that adapted vTaiwan to ask how government policy should respond to generative AI, on topics including copyright, compensation, bias and public services (July 2023); Collective Constitutional AI with Anthropic (October 2023); and a consultation for Creative Commons on the use of licensed works in AI training.

Collective Constitutional AI is the best-documented. About 1,000 US adults, recruited to reflect the population by age, gender, income and geography and screened for basic familiarity with AI, used Polis to submit and vote on principles for an AI assistant. They contributed 1,127 statements and 38,252 votes. Statements with consensus in both of the two opinion groups Polis identified were kept; CIP removed hateful, duplicate or irrelevant statements, leaving 275, which were merged and rephrased into the format used in Anthropic's Constitutional AI training. The public constitution overlapped by roughly half with Anthropic's own in concepts and values, and placed more weight on objectivity and accessibility, and tended to promote desired behaviour rather than prohibit undesired behaviour.

Source: Huang and colleagues (2024) and the Anthropic/CIP report (October 2023).

The authors list the caveats themselves: a small, US-only sample; subjective judgement calls in converting statements into principles; limited and preliminary evaluations; and training that proved more complex than expected. The process was also consultative in a strong sense: the company decided to run it, chose the question, moderated the statements and decided what to do with the result.

What the deployments have in common

None of these processes was binding on a legislature or company. The evidence is therefore about whether AI improves the deliberative experience and the quality of outputs, not about whether it changes how power is exercised.

The risks: steering, persuasion and homogenisation

The capabilities that make AI useful in deliberation make it potentially dangerous to it.

• Latent persuasion. Jakesch and colleagues asked 1,506 participants to write a post on whether social media is good for society, some with an AI writing assistant configured to argue one side. Using the opinionated assistant shifted both the opinions expressed in the writing and the participants' attitudes in a later survey. A facilitator or summariser with a built-in slant could move a group without anyone noticing.
• Persuasive power. Costello, Pennycook and Rand engaged 2,190 people who believed in conspiracy theories in personalised, evidence-based dialogues with GPT-4 Turbo; belief fell by about 20% and the effect lasted at least two months. Salvi and colleagues found that GPT-4 debaters with basic sociodemographic information about their opponent were more persuasive than human debaters in 64.4% of the cases where the two were not equally persuasive (N=900). Hackenburg and Margetts, in a pre-registered experiment with 8,587 participants, found GPT-4 messages broadly persuasive on political issues but no statistically significant advantage from microtargeting. These results cut both ways: AI can correct false beliefs, and it can shift true ones.
• Homogenisation. Doshi and Hauser found that writers given story ideas by an LLM produced stories rated as more creative and better written, but more similar to one another than stories written without AI. In deliberation, the analogous risk is that AI-drafted statements converge on a bland centre and crowd out minority framings, even while each participant feels better represented.
• Opacity and control. Small and colleagues' pilot work with LLMs in Polis found that summarisation and moderation could make conversations more efficient, but that the limited amount of text a model could process "significantly" affected the depth and quality of results, and identified risks requiring principles for characterising and mitigating them. Whoever chooses the model, its instructions and its training data shapes the output.

What the evidence does not show

• Effects on decisions. No study reviewed here measures whether AI-aided deliberation produced better policy, or changed what a government or company did, compared with a non-AI process.
• Legitimacy for non-participants. The Habermas Machine and Collective Constitutional AI measured participants' preferences. Whether the wider public would accept a statement because an AI helped draft it is untested; public-attitude surveys suggest scepticism about AI as a decision-maker in government.
• Automated versus human moderation. The only published comparison of Stanford's automated moderator with human moderation, by its developers, contrasts one online Deliberative Poll with an earlier in-person event; randomised comparisons have not been published.
• Durability. Attitude changes were measured over hours or, at most, weeks; whether convergence persists is unknown.
• Generalisation across cultures. Most experiments were run in the United States and the United Kingdom, in English, with online samples.

Implications for institution design

1. Keep AI in the facilitator's role, not the decider's. Mapping, moderating and drafting are aids; endorsement should remain with participants, and decisions with accountable bodies.
2. Separate the tool's owner from the issue's stakeholders. A company running a deliberation about its own product, or a government about its own policy, should not also control the model that drafts the consensus.
3. Publish the configuration. Instructions, model version, moderation rules and the transformation from raw statements to outputs should be open and auditable, as Polis's open-source code and CIP's documentation partly are.
4. Protect minority framings. Require that synthesis outputs show dissenting positions and the size of the groups holding them, rather than a single "common ground" statement.
5. Test for steering. Run AI-aided processes against randomised controls (human facilitation, no summarisation) and measure attitude change, not only satisfaction.
6. Combine with sortition. AI lowers the cost per participant; random selection supplies representativeness. The 2023 Meta forum and the 2024 Taiwan deliberative poll show the two can be combined.

Open questions

• Would a binding citizens' assembly use an AI drafting tool, and how would its members check what the tool had done?
• Can "bridging" algorithms that reward cross-group agreement be gamed by coordinated groups, as other ranking systems have been?
• Should AI-aided deliberative outputs carry a disclosure, and does disclosure change how they are received?
• Who should certify a deliberation model as neutral, and against what standard?
• If AI can reduce false beliefs through dialogue, should public institutions deploy it, and with what safeguards against its use for advocacy?

Sources

1. Small, C., Bjorkegren, M., Erkkilä, T., Shaw, L. and Megill, C. (2021). Polis: Scaling Deliberation by Mapping High Dimensional Opinion Spaces. Recerca. Revista de Pensament i Anàlisi 26(2).
2. Hsiao, Y. T., Lin, S.-Y., Tang, A., Narayanan, D. and Sarahe, C. (2018). vTaiwan: An Empirical Study of Open Consultation Process in Taiwan. SocArXiv.
3. Small, C. T., Vendrov, I., Durmus, E., Homaei, H., Barry, E., Cornebise, J., Suzman, T., Ganguli, D. and Megill, C. (2023). Opportunities and Risks of LLMs for Scalable Deliberation with Polis. arXiv:2306.11932.
4. Tessler, M. H., Bakker, M. A., Jarrett, D., Sheahan, H., Chadwick, M. J., Koster, R. and colleagues (2024). AI can help humans find common ground in democratic deliberation. Science 386(6719), eadq2852.
5. Huang, S., Siddarth, D., Lovitt, L., Liao, T. I., Durmus, E., Tamkin, A. and Ganguli, D. (2024). Collective Constitutional AI: Aligning a Language Model with Public Input. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 1395-1417.
6. Anthropic and Collective Intelligence Project (2023). Collective Constitutional AI: Aligning a Language Model with Public Input, 17 October 2023.
7. Collective Intelligence Project. Alignment Assemblies.
8. Stanford Deliberative Democracy Lab. Online Deliberation Platform.
9. Stanford HAI (2020). A Moderator ChatBot for Civic Discourse, 6 February 2020.
10. Stanford Deliberative Democracy Lab. Meta Community Forum on Generative AI: results (October 2023 forum).
11. Stanford Deliberative Democracy Lab. Taiwan Deliberative Poll on Utilizing AI to Enhance Information Integrity (23 March 2024).
12. Argyle, L. P., Bail, C. A., Busby, E. C., Gubler, J. R., Howe, T., Rytting, C., Sorensen, T. and Wingate, D. (2023). Leveraging AI for democratic discourse: Chat interventions can improve online political conversations at scale. PNAS 120(41), e2311627120.
13. Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L. and Naaman, M. (2023). Co-Writing with Opinionated Language Models Affects Users' Views. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems.
14. Costello, T. H., Pennycook, G. and Rand, D. G. (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science 385(6714), eadq1814.
15. Hackenburg, K. and Margetts, H. (2024). Evaluating the persuasive influence of political microtargeting with large language models. PNAS 121(24), e2403116121.
16. Salvi, F., Horta Ribeiro, M., Gallotti, R. and West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour 9(8), 1645-1653.
17. Doshi, A. R. and Hauser, O. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28), eadn5290.
18. Gelauff, L., Nikolenko, L., Sakshuwong, S., Fishkin, J., Goel, A., Munagala, K. and Siu, A. (2023). Achieving parity with human moderators: A self-moderating platform for online deliberation. In The Routledge Handbook of Collective Intelligence for Democracy and Governance. Routledge.