Democracy is a muscular affair: if it is not exercised, it atrophies.
This is the premise of the latest report published by DemocracyNext, with the telling title Deliberative Muscles & AI. The “research and action” institute seeks to address the “trilemma” of distrust governing relations among citizens, between citizens and institutions, and vice versa. Its work builds on what has been described as a very recent “deliberative wave”: a surge of more than 1,000 deliberative initiatives. Essentially, groups of people selected as statistically representative samples have been asked to make decisions on matters of public interest.
The pillars of the institute’s work, and more broadly of the concept of deliberative democracy, can be summarised as an expansion of the idea of representation. Through initiatives such as citizens’ assemblies or citizens’ juries, people do not have to delegate the representation of their interests entirely. Instead, they can become advocates for shared interests. Through a structured dialogic process, participants build sufficient consensus to pursue a common decision, developing a form of collective intelligence in which value-based perspectives converge with informed, rationally developed decisions.
The question posed by the report is whether this trend in “DelibTech” can genuinely serve the aims of deliberative democracy by strengthening individual and collective “muscles”, or whether it risks allowing these abilities to deteriorate.
The report sets out to provide a very preliminary assessment of these experiences, while also offering a useful framework for those who facilitate deliberative processes. Its starting point is a distinction made by David Krakauer, director of the Santa Fe Institute, concerning “cognitive artefacts”. Although it is difficult to summarise the researcher’s thinking, for the purposes of this discussion it is enough to divide these material or cultural tools into two categories: complementary artefacts and competitive artefacts.
The report, written by Claudia Chwalisz, Sammy McKinney, Jorim Theuns and Eugene Yi, combines this distinction with the identification of seven “muscle groups”, asking in each case how artificial intelligence tools might strengthen or weaken them.
These reflections now draw on a substantial history of platforms and experiments integrating AI into assembly-based processes, whether institutional or organised by research institutions and third-sector organisations. A paper by Simone Vagnogni and Monica Palmirani of the University of Bologna is entitled Fostering Deliberative Democracy in the Digital Age: An AI-Powered Platform for Enhanced Citizen Engagement in Legislative Process. It is evidence of academic interest in developing AI-enhanced platforms designed to strengthen citizens’ political engagement.
In the authors’ proposal, machine learning, natural language processing and argument mining are essential mechanisms enabling the platform to mitigate logical fallacies in argumentation, thereby improving the quality of discourse aimed at producing legislative proposals. The theoretical basis of this proposal lies entirely within the theories of deliberative democracy, but it must contend with a series of trade-offs created by the integration of AI: bias, data protection and inclusivity, as well as the digital divide and misinformation. These factors threaten the conditions identified by Jürgen Habermas and James Fishkin as necessary for the legitimacy of a process of “collective reasoning”.
Drawing on constructive examples such as Taiwan’s vTaiwan, Kialo and IBM’s Project Debater, the researchers propose an AI-Assisted Argument Construction Wizard that would provide users with real-time suggestions, together with AI agents acting as facilitators. These agents would have clearly defined and, at times, almost Goldonian roles, such as “Devil’s Advocate”, “Mediator” and “Fact-Checker”.
There is therefore both interest and optimism surrounding the development of deliberative platforms integrated with artificial intelligence. If the aim is to maximise citizen participation in the management of public affairs, digital tools become crucial. In DemocracyNext’s previous report, the research team examined the difficulty of scaling deliberative processes, calling for scaling out, involving more participants; scaling up, reaching higher levels of government; scaling across, creating more processes; scaling deep, making them more profound; and scaling in, improving their overall quality.
According to the authors, the construction of “artificial societies” affects only the quantitative dimension. AI can collect vast amounts of data, but it treats deliberation as a machine for distributing opinions. The population to which it refers does not exist: it is synthetic. Deliberation becomes the product of a massive collection of behavioural data, but what actually takes place is merely a process of information processing in which nobody learns anything and nobody listens.
Moreover, the fundamental difference between platforms and deliberative processes lies precisely in their spatial nature: one is intangible, while the other is deeply physical. This brings us back to the decisive question: if an AI tool improves and accelerates certain functions, should it be regarded as an added value?
In reality, it risks being both ineffective and inefficient, because it leaves us with a cognitive and behavioural debt. Once the tool is removed, we are no longer able to complete the deliberative process, wherever it may take place: in parliament, in schools, in museums or within families.
Returning to the muscular metaphor, the report identifies seven areas: self-reflection, reasoning, dialogue, vulnerability, collaboration, imagination and facilitation. For each area, the authors divide AI applications into two categories: those that provide complementary support and those that act antagonistically, depriving the human element of the practice required to give value not merely to the outcome, but to the process itself.
Self-reflection, for example, is an essential part of the deliberative process. Delegating it entirely to AI means giving up a valuable set of cognitive tools. “Plural Reality” is a tool that has been used in numerous deliberative processes in Japan, but it has both strengths and weaknesses. Participants interact with a Cartographer that takes their disorganised ideas and systematises their position on the issue to be discussed.
However, as its own founder, Shutaro Aoyama, has observed, “if the system produces coherence on your behalf, that is not empowerment. It is ventriloquism with a good user experience”, because “people do not know what they want”.
The same considerations apply to problem-solving and shared reasoning. Cognitive offloading, entrusting AI with the effort of processing our thoughts and generating our priorities on our behalf, is one thing. Asking AI to filter the most relevant information, or using it to present that information in more effective visual and multimedia formats, such as podcasts or live scribing, is quite another.
These may appear to be merely formal distinctions, but the difference becomes substantial in areas such as vulnerability, imagination, dialogue and cooperation. In an article for The New York Times entitled No Shy Person Left Behind, which begins with a reconsideration of Martha Nussbaum’s concept of fragility, Hélène Landemore argues that emotions are the true compass of deliberative processes.
Imagine AI agents mechanically allocating speaking time, as envisaged by Stanford’s Deliberation Platform. Such disciplined allocation fails to account for a range of factors. What happens if someone bursts into tears? What if someone needs more time to articulate their thoughts? What if someone feels intimidated during their turn? All these elements, absent from a platform, cannot be subjected to the automatic mechanisms of an AI facilitator.
Imagining alternatives to the socio-economic order of our time is not something that can be entirely delegated to artificial intelligence. AI can create images from our prompts, but doing so relieves us of the necessary effort involved in collectively weaving together a vision of the future.
The technology is undoubtedly useful in many forms. In Democracy R&D’s report The Year in Deliberation 2025, produced by a network bringing together organisations working on deliberation around the world, the first chapter is, unsurprisingly, devoted to artificial intelligence in deliberative spaces.
Sammy McKinney, a Cambridge researcher who also co-authored the first report, is optimistic. AI can help summarise information, identify missing perspectives that may not have emerged during the debate, and support translation and different communication formats, as tools such as Dembrane and deliberAIde already do. Yet just as it can generate inclusive possibilities, it can also become a potential source of inequality linked to digital literacy. Furthermore, the entire supply chain required for its development demands a democratic discussion of its own.
“I do not want to introduce an intermediary into the relationship between the state and its citizens: I want to create an interpersonal experience. […] But it makes sense to use AI for all the preparatory and follow-up work: simultaneous translation, data synthesis and similar operations,” argues Matthew Byrne, Senior Director of Deliberation at Unify America.
There are forms of labour from which AI can relieve us, and areas in which it can act as an ally in deliberative processes. But this is possible only on the condition that it does not become a “plastic membrane” enveloping the process and stripping it of its tactile, bodily and distinctly human qualities.
Platforms such as POL.IS can provide valuable training, but there is a friction inherent in the fullness of the deliberative experience. Shortness of breath and aching muscles are the true signs of a mature democracy.
Ludovica Taurisano