AI reflectors

As AI gets better at processing free-form text, it’s tempting to point it at a population and ask it to output what the people want. Manon Revel (now at Google DeepMind) and political philosopher Philippe Pénigaud argue that this temptation rests on a flawed assumption, and propose a more modest and more useful role: the AI reflector.

Three stages of inferring the collective will

Their starting point is that today’s AI methods are the latest in a century-long lineage of trying to compute the collective will:

1. Polling — from Gallup’s stratified sampling onward. Cheap and scalable, but it forces people into constrained multiple-choice questions, and “humans don’t think in tick-boxes.”
2. Latent methods — “collective response systems” like Polis and Remesh, where people write free-form statements and others vote agree/disagree, building an interaction (or “will”) matrix. Bridging aggregation then surfaces statements that span opposed groups.
3. LLM methods — the Habermas Machine and generative social choice, which use language models to turn raw opinions into representative statements (usually still reconstructing a “will matrix” in the middle, to keep some interpretability).

The flaw: the calculability hypothesis

Underneath all three sits what Revel and Pénigaud call the calculability hypothesis: the belief that preferences exist as standalone, pre-given quantities, and the only question is how best to compute them. They argue the collective will is instead indeterminate at three levels:

• Empirically — preferences are unstable; they shift with framing, context, and mood.
• Aggregatively — even with identical inputs, different aggregation methods yield different “winners” (the classic social-choice problem).
• Normatively — there is no people’s will floating in the world waiting to be found. It is co-created through the process of deliberating, opting in, and feeling heard (a point they draw from the historian Pierre Rosanvallon).

Because of this, they’re wary of using AI to make binding decisions. In Habermas’s terms, AI doesn’t belong in the context of justification (where opinion becomes binding law) — the legitimacy cost is too high. Its proper home is the context of discovery.

AI as a mirror, not an oracle

So they reframe these tools as AI reflectors: a mirror that helps a community understand itself, rather than an oracle that hands down an answer. Two functions:

• Reflective elicitation — you write open-ended thoughts privately, and the AI offers counterpoints, testimonies, or contradictory facts on request, so you reconsider your view in light of others before any group decision (the introspective “deliberation within” moment).
• Synthesis — generating an inclusive picture of where the group actually stands, without burying the individual reasons underneath.

This is the same instinct as the complementary-vs-competitive test: build tools that strengthen collective sense-making, and keep the binding decisions human. See also synthetic participation for the failure mode of skipping the humans entirely.

Sources

• Manon Revel & Philippe Pénigaud, “AI-Facilitated Collective Judgements” / “AI-Enhanced Deliberative Democracy and the Future of Collective Will,” 2025: arxiv.org/abs/2503.05830
• Manon Revel, talk to the Cooperative AI seminar series, 2025: youtube.com/watch?v=u_azC4tgpRU