AI and Deliberation. Moving from Sophisticated Polling to the Order of Discourse

Bernard Reber and members of the projects Democratic Commons, AI for Augmented Deliberation. Innovations in the subfield of AI and Democracy resemble sophisticated surveys more than genuine deliberations. This is a paradox, because when we think of LLMs, we assume we are dealing with language. This is typically the case with the equally famous and misnamed “Habermas Machine” (Tessler et al., 2024), published in the prestigious journal Science, for example. This is undoubtedly one of the last misinterpretations Habermas was subjected to before his death. It proposes an approach based on language models to accelerate the search for consensus among a plurality of political opinions by modeling how humans evaluate a variety of algorithm-generated consensus statements. I will discuss this in the first part of my talk. Bearing in mind the communicative skills required in deliberation – literary techniques that have been known and developed for Antiquity– we will see how, and whether, philosophy is useful for practical experiments, such as real-life deliberative experiments. Rawls and Habermas, did not believe so, even though some consider them to be the theoretical foundations of the deliberative democracy theory. On the contrary, it seems to me that philosophy can indeed aid deliberation (AI for Augmented Deliberation: https://ai4ad.proj.lip6.fr/). The emergence of AI might challenge even more philosophers among others to do so; but not under just any conditions. We must shape it, not be shaped by it, knowing that the whole of human history depends on a co-evolution with the technologies it has produced. Without them, human beings would not be what they are; for better or for worse. We will then turn to underdeveloped components that ensure the quality of deliberation: communicational capacities. These fall within the ‘the order of discourse’, which structures linguistic exchanges and intra- and inter-institutional design choices. AI is certainly misnamed by its comparison with embodied intelligence, but its generative conversational form (LLM or LM) focuses precisely on all linguistic capabilities, whether simple, such as translating, moderating or writing, or more complex, such as facilitating or synthesising. In any case, it offers an opportunity to clarify these linguistic capabilities in order to link them to the communication skills deployed in deliberation. Depending on one’s view of AI, one will react differently. If one sees all its limitations, one will argue that AI-assisted deliberation will not be any more intelligent. On the contrary, as I shall do, one can recognise that AI tools are already here, imposed, or worse, widely adopted, and that they accompany established platforms that enable new participatory and deliberative uses. But all technology and its uses are subject not only to evaluation, but also to correction. This is one of the challenges of the ongoing project, Democratic Commons. It explores the idea of AI-assisted and simulated deliberation in an interdisciplinary manner. It proposes, as a starting point, democratic criteria to be applied to the entire socio-technical chain of participatory platforms throughout the political cycle, from election campaigns to the co-construction of public policies.