Can AI mediation improve democratic deliberation?

The strength of democracy lies in the free and equal exchange of diverse viewpoints. Living up to this ideal at scale faces inherent tensions: broad participation, meaningful deliberation, and political equality often trade off with one another (Fishkin, 2011). We ask whether and how artificial intelligence (AI) could help navigate this “trilemma” by engaging with a recent example of a large language model (LLM)-based system designed to help people with diverse viewpoints find common ground (Tessler, Bakker, et al., 2024). Here, we explore the implications of the introduction of LLMs into deliberation augmentation tools, examining their potential to enhance participation through scalability, improve political equality via fair mediation, and foster meaningful deliberation by, for example, surfacing trustworthy information. We also point to key challenges that remain. Ultimately, a range of empirical, technical, and theoretical advancements are needed to fully realize the promise of AI-mediated deliberation for enhancing citizen engagement and strengthening democratic deliberation.

Introduction. In the early 2000s, concern arose in the United States about the fairness and efficacy of the allocation process for kidney transplants. The system at the time pursued a seemingly robust objective: namely, it respected the time spent on the waiting list and the expected number of additional years of life for the recipient. However, there were valid objections: Did it discriminate unfairly against the elderly by virtue of them leading healthy lives? Did it recognize that the medical urgency of situations differed? Did it treat different demographic groups fairly? The experience of people placed on the waiting list suggested not. To address these concerns, public consultations took place in which everyone affected was encouraged to share their viewpoint and engage in deliberation. After numerous rounds of feedback, a new algorithm was created that weighs ten different variables, a change widely praised by medical practitioners and patient groups alike. Commenting on the process, scholar David Robinson notes that “[E]xperts, patients, and advocates balanced hard trade-offs to remake the kidney allocation algorithm, leading to better health for more people and a fairer allocation of organs...People raised their voices. People were heard” (Robinson, 2022). The only serious limitation was that the process took 20 years to complete.

Imagine if we could channel the power of public deliberation effectively and efficiently, providing timely guidance on a range of issues. Traditional methods for public consultation have well-known limitations. Focus groups can provide detailed feedback but are too small scale to be representative and are susceptible to biases affecting group discussion (MacDougall and Baum, 1997; O. Nyumba et al., 2018). Canvassing opinions door-to-door to elicit large-scale input can be costly and timeconsuming. Running a poll can help reveal the balance of opinion but typically is restricted to a small number of judgments, does not allow for discussions of different points of view, and can be sensitive to the wording of the questions (Desaint and Varbanova, 2013; Gallup, 1941). More structured forms of public input, like deliberative polls or citizens’ assemblies, involve meaningful deliberation but also require significant time commitments from participants (Fishkin, 2003). Consequently, well-resourced groups—who have more time, better coordination, or greater knowledge about the process—can dominate and be over-represented in these settings (Birhane et al., 2022), and these offline methods are not scalable to very large groups of people.

Recent efforts to overcome the scalability and logistical challenges of traditional consultation methods take shape in digital deliberation technologies (Coleman and Shane, 2012; Goñi, 2025). Systems like Pol.is and Remesh facilitate large-scale input by allowing participants to contribute statements and react to others, typically through voting (Small et al., 2021). These platforms process the interaction data (e.g., votes) with machine learning techniques in order to uncover collective sentiments by grouping participants based on shared opinions and highlighting points of consensus or division. Despite the significant expansion of scale of participation and efficient ways to map general agreement, these tools are limited in their capacity to process the rich semantic content of participant contributions. Further, the deliberation they support is structured around discrete statements and voting and may miss out on deeper engagement with the reasoning and substance of diverse perspectives, critical for finding agreement and consensus in a population (Habermas, 1985).

Recent advances in artificial intelligence (AI), specifically the emergence of increasingly capable large language models (LLMs), have introduced new opportunities to further support deliberative and democratic functions (Landemore, 2023; Lazar and Manuali, 2024; Small et al., 2023). LLMs can process vast amounts of text and can be prompted or trained further to perform arbitrary tasks involving text (Brown et al., 2020).2 Researchers and practitioners have explored how LLMs can support democratic deliberation by summarizing content (Small et al., 2023), aggregating opinions in a manner to increase collective support (Fish et al., 2023), representing individual judgments in decision-making processes (Gudiño et al., 2024; Jarrett et al., 2023), facilitating public deliberation (Ma et al., 2025), and implementing principles derived from democratic methods (Huang et al., 2024). Despite lots of speculation and demonstrations of the strengths and shortcomings of LLMs to support various deliberative tasks, there has been relatively little empirical work evaluating their efficacy.

Recently, Tessler, Bakker, et al. (2024) built and investigated an LLM system—which they call the Habermas Machine (HM)—designed for the task of finding common ground between people with different viewpoints (Figure 1).

Related work. Introducing new technology into deliberation can be seen as an attempt at an answer to Fishkin’s “trilemma” (Fishkin, 2011). Fishkin’s trilemma highlights a fundamental tension between three democratic goals: political equality (everyone’s voice or vote is counted equally), inclusion (as many people as possible exercise their voice or right to vote), and deliberation (citizens thoughtfully consider issues). Large-scale voting captures equality and inclusion, but not deliberation; online discussion fora may have deliberation and inclusion, but unequal access or organized interests can lead to a lack of equality; structured deliberation exercises (such as Stanford’s Deliberative Polls) prioritize equality of voice and deliberation but are difficult to scale to large groups. Addressing these trade-offs is therefore a key driver behind the exploration of technological solutions, and especially those involving AI (Landemore, 2023).

The field of AI-enabled deliberation is growing quickly (Argyle et al., 2023; Kim et al., 2020; Ma et al., 2025; Michael et al., 2023; Shin et al., 2022; Small et al., 2023). In order to understand the potential of AI in this field we focus on one particular model—the Habermas Machine—to ground the discussion in a specific example that has been evaluated for its efficacy in experimental settings (Tessler, Bakker, et al., 2024).3 The HM also serves as a useful example of a large language model technology being inserted into human deliberation, raising technical and ethical questions.

Tessler, Bakker, et al. (2024) offer a noteworthy experimental demonstration that AI-mediated deliberation can help diverse small groups find common ground on potentially divisive topics. These results naturally raise the possibility that AI mediation could resolve the tension between mass participation, equality of contribution, and quality of deliberation that Fishkin (1991) introduced. While the trilemma forms the core of the following analysis, we will explore the broader implications, including the role of AI in the democratic process and its comparison to traditional deliberation, in the Discussion.

Efforts to improve deliberative quality are, in fact, central to many existing deliberative democratic practices. For example, citizens’ assemblies (such as Ireland’s Citizens’ Assembly established in 2016) include expert testimony phases drawing on a diverse range of experts in order to ground discussions in evidence (Müller et al., 2023; Suiter, 2018). Deliberative polls provide participants with balanced briefing materials to create a shared, informed foundation (Fishkin and Luskin, 2005).

Method. The HM, as described in Tessler, Bakker, et al. (2024), is a system of LLMs that was designed with the goal of finding common ground among people with diverse perspectives (Figure 2B). The HM is an LLM system which handles a list of inputs, coinciding with the opinions, critiques, or comments from a group of individuals. Interaction with the HM begins by group members submitting their personal opinions on a topic (e.g., in response to ‘Should we lower the voting age to 16?’) as a text statement (e.g., a short paragraph outlining their position and justification). The HM then generates a set of ‘group opinion statement(s)’—statements that try to reflect the common ground among the group of individual opinion writers. Participants then vote on which one they prefer. The group members then are invited to write comments or critiques on the (winning) statement (e.g., expressing approval or disapproval, ways to improve the statement, etc.). These critiques are sent to the HM, which then generates a new set of revised group opinion statements. Another round of voting occurs, and the process could continue for more rounds of critique and revision. In Tessler, Bakker, et al. (2024), the process ended after two rounds (i.e., opinions and one round of critiques).

The HM has two LLM components: a generative model and a reward model. The generative model proposes a number of candidates (e.g., 32) for a ‘group opinion statement’ based on the individual opinion statements submitted. These statements explore different wordings and ways of combining the opinions, ideally spanning a diverse set of possible group opinions. The reward model predicts how much each group member would agree with each of the generative model’s candidates, based on the group member’s individual opinion (i.e., would the person with opinion X like statement Y?).4 The reward model predictions are converted into rankings such that each group member has a predicted ranking over the candidate statements. The rankings are aggregated using social choice theory (e.g., the Schulze method, a way of implementing a ranked-choice voting scheme), and the winning candidate is shown to the group members (or some number of the top candidates could be shown). The HM also has the capacity to revise a group statement by incorporating written critiques from the individual group members through the same generation and selection process.

In other words, the Habermas Machine implements a simulated election. Candidates are proposed from a generative model, and votes are determined through a reward model. Votes (rankings) are aggregated in a way that the outcome of a ranked choice election would be determined, and the winning statement is returned to participants for them to write comments or critiques. The process iterates until some predetermined stopping point (e.g., after a certain number of rounds of critiques).

The HM architecture has three core components: a generative model for sampling candidate group statements, a PRM to rank these statements according to the predicted preferences of each participant, and an aggregation mechanism to choose a final statement based on these PRM-predicted rankings. We discuss how fairness can be built into each component.

Fair aggregation Aggregation is the process of combining a set of individual preferences into a collective ranking. It is the process of counting votes in an election. Both social welfare and social choice frameworks could be used for aggregating preferences over a set of candidates. In a predecessor of the Habermas Machine, Bakker et al. (2022) explored different social welfare functions ranging from a utilitarian welfare function (weighing all participants’ preferences equally) to a Rawlsian welfare function (only taking the most dissenting participant into account). While this social welfare framework is computationally simpler, this approach relies on the questionable assumption that the strength of one person’s preference can be directly weighed against another’s.

Discussion. Numerous principles can contribute to a fair democratic process, such as transparency, inclusivity, equality in decision-making, and safeguarding minority rights against majority rule. While a normative discussion of which principles are essential is beyond scope, it is worth asking whether an AI system could uphold any reasonable interpretation of democratic fairness (Barocas et al., 2023). How can we design and implement AI mediators like the Habermas Machine to ensure they operate in a manner that is demonstrably fair and equitable to all participants?

The fact that language models do not adhere to prescribed, deterministic rules in generating outputs is precisely what empowers them to effectively handle a vast and diverse array of inputs. But this opacity and unpredictability can also undermine the legitimacy of the system’s output, clashing with the democratic norms of transparency and mutual understanding of the process. How can we ensure that the statements generated by the Habermas Machine, or a similar technology, are derived fairly and faithfully from participants, and that participants can trust the system to do so? We argue that principles of democratic fairness need to be explicitly incorporated into the design of the system, and then combined with interpretability tools that continuously verify and monitor that the AI mediator is aligned with those principles.

Creating a high quality PRM is nontrivial. Tessler, Bakker, et al. (2024) report that the relatively small and dated Chinchilla-based reward model (which was used in the HM) was actually more performant than a large, state-of-the-art, out-of-the-box large language model (SM 3.5). The primary reason is that the reward model had been specifically fine-tuned from human preference data for the task. Such a result highlights the importance of careful consideration and curation of training data:

Scalability is a key promise of AI-mediated deliberation, offering the potential to engage far larger and more diverse groups of people in democratic processes. Can the HM effectively facilitate deliberation among increasingly large groups of people?

One of the foundational ideals for deliberative democracy is the principle of full inclusion, that “no one capable of making a relevant contribution has been excluded” (Habermas, 2003). Traditional deliberation methods struggle with large-scale participation, owing to the time or cost incurred by participants and organizers of deliberative exercises.5 What if we didn’t have to make choices about who to include? What if we could include everyone? In Tessler, Bakker, et al. (2024), small groups (up to five people) were studied, a scale where human mediators are competitive with the AI. If a human had to aggregate twenty, one hundred, or one thousand opinions, the task would not be possible.

Even if an LLM can technically aggregate thousands of opinions, how can we be confident it does so effectively? Simply expanding the HM protocol to large groups and optimizing for endorsement might lead to short, bland statements that say little of substance. While Tessler, Bakker, et al. (2024) found that the model produced informative statements in small groups, this may not generalize to much larger, more diverse populations.

Two approaches from the world of scalable oversight provide possible ways out of this quagmire: AI assistance and task decomposition. Capitalizing on the ability for LLMs to perform arbitrary languagebased tasks, one can further train an LLM to issue critiques (natural language critical comments) in a given context. Saunders et al.

Conclusion. In summary, the Habermas Machine demonstrates considerable promise for scaling deliberative democracy, although important questions remain. While the original training data was limited in The convergence of AI and democratic deliberation offers a potent avenue for enhancing citizen engagement and fostering more representative governance. While systems like the Habermas Machine are not a panacea for all democratic challenges, they can represent a significant stride towards augmenting traditional deliberative processes. Further research and development, guided by the principles of fairness, transparency, and inclusivity, promise to unlock the full potential of AI-mediated deliberation, creating a future where technology empowers citizens to shape the policies that affect their lives.

Limitations. Finally, the authors replicated their findings in a demographically representative sample of the UK, finding convergent shifts in position across groups on certain topics. An example final group statement is shown in Figure 2C, and many examples can be found in the Supplementary Materials (SM 6) of Tessler, Bakker, et al. (2024). The deliberation questions and the data collected are publicly available at: github.com/google-deepmind/habermas_machine.

Despite certain benefits of digital and caucus mediated deliberation, important variables may be lost compared to face-to-face settings (Sætra, 2024). Specifically, fully digital approaches may struggle to foster the positive social connections—trust, empathy, and reciprocity—that often characterize inperson dialogue. Furthermore, the very nature of algorithmic decision-making can trigger algorithmic aversion (Dawes, 1979), wherein individuals distrust or resist outcomes generated by algorithms, even when these outcomes are demonstrably superior to those achieved by humans.