Synthetic participation

Synthetic participation is the idea that citizens’ input could be artificially generated — a model of the populace standing in for the people themselves. Instead of consulting real residents (which is messy and expensive), an institution could just “ask” an AI trained to imitate them. It’s worth naming clearly, because it’s a tempting shortcut and a dangerous one.

Why it’s risky

Faking public input isn’t new — long before ChatGPT, “astroturf” campaigns flooded official channels with manufactured comments (the US FCC’s net-neutrality comment period is a notorious case). AI makes convincing fakes vastly easier. If policymakers start treating a model of the electorate as a substitute for the electorate, it can erode the legitimacy of the very channels meant to give people a say. Opinion polling is a mild, analog version of this; AI could accelerate it into something corrosive.

The line: who owns the twin

There’s a meaningful distinction. A person-owned, person-controlled digital twin — one you tune to argue for your views, that you send into the world and remain responsible for — is a different proposition from an institution modelling a whole constituency without anyone’s consent. The first keeps a human accountable for what they “hit submit” on; the second replaces people with a guess about them.

The principle to hold

Use AI to support participation, not to simulate the public. Where AI helps people express and understand one another, it strengthens democracy; where it manufactures a synthetic public, it hollows it out.

A harder case: agents built from real positions

Between the two poles sits a subtler use. Tools like Gemot build synthetic agents from people’s own published positions — each grounded in direct quotes — and have those agents deliberate to surface the cruxes that divide a debate, with anti-sycophancy checks that reject a revision when it drifts from what the source actually said. That is more honest than modelling a whole constituency from thin air: it stays anchored to real, attributable views and can be audited against them. But it is still a map of a debate, not a substitute for the people in it. The safe use is the same as any argument map — to understand the shape of a disagreement before real people deliberate, never to let the agents’ “agreement” stand in for a public that was never asked.

Sources

• Matt Stempeck, with audience discussion (including the person-owned-twin counterpoint), “Civic AI” — Metagov seminar, 2025: youtube.com/watch?v=SonfdR0_h2w
• Gemot — an open-source multi-agent deliberation tool that builds agents from published positions and surfaces cruxes rather than summaries