Explore Policy

Most of the field is still arguing about the wrong ceiling — or, more precisely, running two conversations that barely speak to each other.

Conversation A (analysis): Talk to the City, Polis, the archaeology papers. AI listens to what people already said. Excellent at hidden consensus at scale. Blind at what happens before deliberation and what breaks in implementation.

Conversation B (generation): Argyle, social simulacra, multi-agent debate. AI generates what people would say. Excellent at stress-testing in silicon. Blind at whether the utterance is grounded in real stakeholder residue or just plausible-sounding.

What’s missing — what Explore is building as Silico — is a platform that does both, seamlessly.

Colin Megill and Anders Sandberg already named the two experiments — listening to the record versus branching the futures. Our claim is sharper: those are not two features of a product. They are two halves of a new democratic object.

In silico deliberation, for us, is continuous institutional rehearsal under adversarial and multi-epistemic pressure — grounded in real vote geometry and media residue, allergic to fake consensus, and unfinished until a human room can refuse the machine’s map.

Not a chatbot town hall. Not Hidalgo’s digital twins voting all night. Not Ion whispering “the people agree” into a prime minister’s ear. Something stranger and more useful: a wind tunnel for governance.

The ceiling Colin hit — and the floor Anders opened

Colin: the archaeology of disagreement

Megill’s 2023 thread treated in silico as extracting structure from noise. Polis showed the trick: vote on contested statements; cluster the votes; discover that polarised fights often hide a large region of shared stakes with different weights. Talk to the City pushed the same instinct into unstructured text. For civil society: you can read the record before you convene.

That insight is non-negotiable. It is also incomplete. Archaeology tells you what the bones already are. It does not tell you whether the building falls when you hang a new roof on them. And if you stop at “make the 70% consensus visible,” you inherit a quiet bias: consensus as the prize. Sometimes the democratic task is to keep a minority un-smoothed. Sometimes the task is to show that two publics do not share a world, and any bridging sentence is a lie with good typography.

Anders: refuse the single model

Sandberg’s move is to treat policy talk as unfinished until it has been projected into futures that disagree with each other. His method, in plain terms:

Parallel models, not a turf war — run several pictures of the world at once; argue about which policy still holds across them.
Crude simulations as agenda weapons — Club of Rome–style foresight: even a simple run can force a room to admit a trajectory it was refusing to see.
Scenarios that outrun the bureaucracy’s present — narcotics to 2030 via maker labs and platform delivery, against ministries still fighting last decade’s street corner.

That is not “AI helping people agree.” It is AI making disagreement about the future expensive to ignore.

Aviv: simulate the process, keep humans in the driver’s seat

Aviv Ovadya’s cut is about designing democratic systems themselves. Simulations, in his framing, let you try process variants before you burn a city or a parliament on them — different convening rules, different information diets, different mediation stacks. He is also unusually honest about the present:

Test approaches in silicon first — cheap dead ends beat expensive institutional ones.
They are not yet trustworthy enough to be definitive — today’s runs are research instruments, not mandates.
Use sims to locate the human boundary — where a representative agent might be enough for a subtask, and where a human must stay in the driver’s seat.
Accuracy is a research programme, not a shipping claim — improve the wind tunnel so the field stops repeating costly mistakes.

That third point is load-bearing for Explore. Silico is not “agents instead of citizens.” It is a boundary-finding machine: which parts of prep and stress-test can be automated without stealing the decision, and which parts must remain stubbornly human.

Colin excavates the record. Anders branches the futures. Aviv insists the simulation of process stay provisional — useful, improvable, never confused with the demos.

Explore starts where those three sentences collide: listen until the geometry is honest; attack the proposal until the geometry breaks; never let the run replace the room.

What the last three years killed

Megill’s catalogue was prophetic. The follow-up literature is a slaughter of the naive reading.

Habermas Machine (Science, 2024) shows AI can mediate real discussants into statements they prefer — Colin’s lane, adult form. Mediation ≠ mandate.
Bisbee et al. (Political Analysis, 2024) and the homogenisation papers gut silicon samples as survey replacements. Persona agents are a statistical Das Man, not a constituency.
Sycophancy / deliberative illusion work shows multi-agent “debate” often converges by deleting facts and dissent. Agreeable agents are anti-deliberative technology wearing a toga.
Ion remains the political warning label: captured inputs + consensus theatre = algorithmic authoritarianism with a dashboard.

So we do not “fix” agent town halls by making agents ruder. We change the object of simulation.

We simulate institutions, coalitions, adversaries, and time — not souls.

Push the definition: five theses for Explore

1. The product is a disagreement engine, not a consensus engine

Polis already knew this mathematically: the interesting output is the map, including the clusters that refuse to merge. Explore treats non-convergence as a first-class result. If every branch ends in a tidy “shared values” card, the system has failed. A healthy in silico run should be able to return: these groups do not share a decision-space; here is the minimal conflict set; here is what each would have to give up.

Consensus is sometimes excavated. Sometimes it is manufactured. The machine must be able to tell the difference — and to prefer an ugly true map over a beautiful false bridge.

2. Votes are sacred; vibes are not

When people say they “used Polis data,” they often mean they used comment text. That is archaeology with the skeleton removed. The vote matrix — who agreed, disagreed, or passed on which statement — is the democratic instrument. Flatten it into embeddings and you can still make pretty clusters; you can no longer recover unexpected overlap.

That is why jonaskg/open-deliberation ships as two configs on purpose:

Config	Object	Vote-preserving?
`text_corpus`	Multi-source utterances for NLP / mediation / scenario prompts	No
`polis_votes`	`comments` + `votes` + `summary` per conversation	Yes

If your pipeline cannot say which one it is using, it is already lying about its epistemology.

3. Time is a deliberative stakeholder

Ordinary consultation freezes a public in the present tense. Sandberg’s scenarios smuggle future constituencies into the room: the people who inherit drought, platformised drugs, automated layoffs, collapsed trust. Explore pushes this further — not as mystical “rights of future generations” rhetoric, but as adversarial timeline agents: five-year drought, ten-year patronage adaptation, twenty-year institutional decay. They do not vote. They prosecute the proposal.

A policy that survives today’s stakeholder workshop and dies in year four was never deliberated. It was marketed.

4. Multi-epistemic pressure, or it is just Silicon Valley cosplay

Habermasian deliberation — reasons exchanged among equals — is one technology of decision among many. Elder authority, kinship obligation, religious jurisprudence, harmony norms, and post-colonial suspicion of “neutral” procedure are not edge cases. If your agents only know how to argue like seminar liberals, your wind tunnel only certifies policies for seminar liberals.

In silico, for us, means running the same proposal through incompatible legitimate decision cultures and logging where it is unintelligible, offensive, or unenforceable — not “translating” everything into a single rationalist dialect until the conflict disappears.

5. The loop does not end in the model

Hidalgo’s Augmented Democracy dreams of delegates that never sleep. Explore’s nightmare is the same dream with better UX. Ovadya’s caution is the antidote: simulations are not yet reliable enough to guide as destiny. The terminal of an in silico run is not a recommendation. It is a brief for a human convening — and a map of where the agent was sufficient versus where a human must remain in the driver’s seat: fragile points, false overlaps, adversary exploits, minority lines that algorithms would sand off, scenario branches that refuse to die.

The machine’s highest virtue is to make the subsequent human meeting shorter and harder — fewer ritual speeches, more confrontation with the actual trade-offs.

What “agentic deliberation” actually is (when pushed)

Not agents pretending to be voters. Not a jury of GPTs asked to “roleplay a farmer.”

Agents as prosecutors of a proposal under constraints drawn from real residue.

Ingest geometry — vote matrices, consultation dumps, media diets, institutional rules. Prefer structure over persona fanfic.
Spawn lenses, not souls — pastoral water rights, urban utility, patronage broker, climate court, future drought, hostile capture. Lenses can be stubborn by design; sycophancy is a bug we measure, not a feature we prompt around.
Branch in parallel — Sandberg’s multi-model rule: never one forecast. Utilitarian / libertarian / ecological / extractive-state / community-harmony forks of the same clause.
Attack — how does a bad actor use this rule? Where does enforcement fail? Which coalition collapses first?
Emit a fragility map — not “approve / reject,” but a cartography of break points for the humans who still have to sit in a room.

That is Colin’s archaeology feeding Anders’s wind tunnel until the proposal either earns a meeting or dies in silicon, cheaply.

A Nairobi water-allocation draft should be able to fail overnight against pastoral record + drought branch + patronage capture — before anyone spends three months performing consultation. That is not anti-democratic. It is anti-theatre.

The political risk, named without euphemism

Every listening machine can become Ion. Every scenario machine can become a technocratic alibi (“the model said”). Explore’s design constraint is therefore moral, not cosmetic:

No single “public score.”
No agent whose output is framed as a vote.
No consensus card without a dissent card of equal visual weight.
Provenance: which corpus, which vote matrix, which scenario assumptions.
A human refusal path that is cheaper than compliance with the dashboard.

If those constraints feel like they slow the product down, good. Democracy’s failure mode is speed with a halo.

What Explore is building

Explore In Silico Deliberation is not a chatbot with a civic skin. It is infrastructure for disagreement-preserving rehearsal:

open data that refuses to collapse votes into vibes (jonaskg/open-deliberation — text_corpus vs polis_votes);
scenario canvases that run forks in parallel (the policy work already gesturing this way);
agentic stress layers that prosecute clauses instead of impersonating citizens;
and a public surface at /silico — a wind tunnel you can point a proposal into.

Launch note: Launching Explore In Silico Deliberation (Julia Park & Jonas Kgomo).

The field asked whether AI could scale deliberation. The timid answer was: summarise better. The reckless answer was: replace the demos.

Our answer: scale the rehearsal until theatre becomes expensive and fragility becomes visible — then put humans back in a room that can finally afford to tell the truth.

In silico does not mean “democracy, but digital.”
It means democracy, with a right to fail in silicon first.

Explore In Silico Deliberation