Skip to content
Citizen Infrastructure

Can AI scale deliberation?

Good deliberation is slow, structured, and usually in-person — which raises an obvious temptation: use technology to scale it to millions. Hélène Landemore’s answer is a careful maybe, and it comes with a sharp test for telling helpful tools from harmful ones.

First, what deliberation actually is

Section titled “First, what deliberation actually is”

Deliberation is structured exchange in which everyone is roughly equally exposed to the arguments and can respond to them. By that definition, a lot of what gets called “scaling deliberation” isn’t:

  • Mass online text platforms, where millions thumbs-up fragments of each other’s opinions, are aggregation — a wisdom-of-crowds process. Genuinely useful, but a different thing: too fragmented to be deliberation.
  • Agentic AI “deliberating” on your behalf is furthest of all. Outsource it entirely and you lose the democratic muscle — the capacity to understand the issue and make the conclusion your own. What’s left is blind deference.

A test: complementary vs competitive tools

Section titled “A test: complementary vs competitive tools”

David Krakauer (Santa Fe Institute) distinguishes complementary cognitive artifacts, which build your capacity — like an abacus, which makes you better at arithmetic even after you put it down — from competitive ones, which replace it — like a calculator, which leaves you helpless without it. Good deliberation tech is complementary: it strengthens people’s ability to deliberate. The danger is “efficiency” that smooths away the productive friction — the trust-building, and the slow collective work of sense-making (one practitioner draws the line at letting AI cluster people’s ideas, because doing that work together is where understanding and ownership are born). The friction often is the point; see civic love.

Worth testing, carefully

Section titled “Worth testing, carefully”

The boundaries are genuinely unknown, so experiment — while keeping the human core. Some promising uses are clearly assistive: AI helped Israeli and Palestinian peace activists find consensus framings (via Remesh) when face-to-face talks had stalled; “values warm-up” tools help people articulate what they care about before a deliberation. The question to ask of any tool is Krakauer’s: does this build the democratic muscle, or replace it?

A framework for judging each use

Section titled “A framework for judging each use”

For a more systematic version of the same instinct, Sammy McKinney’s study of AI in citizens’ assemblies maps 17 possible applications across an assembly’s whole life cycle (translation, facilitation, aggregation, clustering public input, generating consensus statements, inclusive learning materials, and more) and scores each against three kinds of good. Democratic goods: does it help inclusiveness, popular control, considered judgment, and transparency? Institutional goods: does it improve efficiency and scalability? Ethical integrity: does it respect privacy, avoid imposition of an off-the-shelf tool on a local context, and mitigate bias? His conclusion mirrors Krakauer’s test: AI can raise both democratic quality and institutional capacity if the right safeguards are kept — human oversight, ethical data governance, co-design with participants, and hybrid human-plus-AI designs. For the bigger question of which direction to scale, see the five dimensions of scaling deliberation.

Approximating mass deliberation

Section titled “Approximating mass deliberation”

Landemore’s deeper worry is a legitimacy one: a democracy’s laws are fully legitimate only if they could have issued from inclusive deliberation among everyone, yet real deliberation breaks down past a few hundred people. Her wager is that AI might let us approximate mass deliberation well enough to count. She floats two models: mass online deliberation (a single shared space, à la Wikipedia, where an algorithm clusters everyone’s proposals into a manageable bird’s-eye view — proposed by engineer Cyril Velikanov), and many rotating mini-publics (enrol the whole population in randomly-selected assemblies and rotate them until, in effect, everyone has deliberated with everyone). Neither needs all citizens: she speculates that enrolling 10–15% — still millions of people, and representative if truly random — might be a “good enough” threshold for legitimacy. France’s Great National Debate was a low-tech gesture in this direction. For the fuller map of what “scaling” can mean, see five dimensions of scaling deliberation.

This is the practical, tool-level companion to AI for participation and synthetic participation. For platforms in this space, see Decide & make sense together.

Sources

Section titled “Sources”