Launching Explore In Silico Deliberation

The field of AI × democracy has been running two conversations that barely speak to each other.

We are launching Explore In Silico Deliberation — and opening the data behind the first cut of /silico — because that gap is not a literature gap. It is an infrastructure gap between institutions and people.

Two conversations

AListen

Talk to the City. Polis. The archaeology papers.

Thesis: AI listens to what people already said.

Feed comments, interviews, and vote matrices. Cluster arguments. Surface hidden overlap on polarised topics. Publish a map so stakeholders can see that the loud 30% is not the whole space.

Good at: hidden consensus and structured dissent at scale.

Blind at: what happens before people deliberate, and what breaks during implementation. A perfect map of last year’s consultation will not tell you whether this year’s water rule collapses under drought and patronage.

BGenerate

Argyle’s silicon samples. Social simulacra. Multi-agent debate.

Thesis: AI generates what people would say.

Persona agents, town-square prototypes, policy red-teams in silicon. Stress-test a clause before you spend three months performing consultation.

Good at: rehearsal and attack surfaces — Sandberg’s parallel models, Club of Rome–style agenda-setting, futures that outrun the ministry’s present tense.

Blind at: whether the utterance is grounded in real stakeholder residue, or just plausible-sounding prose. Ungrounded generation collapses into homogenised agreeableness. A room of helpful agents is not a demos.

Aviv Ovadya adds a third cut aimed at process design: simulations let you try democratic procedures before you ship them into the real world — and he refuses to overclaim. Today’s sims are not yet trustworthy as definitive guides. Their research value is partly boundary finding: where a representative agent might be enough for a subtask, and where a human must stay in the driver’s seat.

What’s missing

A platform that does both, seamlessly — and treats Ovadya’s caution as a product constraint, not a footnote.

Silico is Explore’s name for the fused object: excavate disagreement from the record, prosecute the proposal under parallel futures, then hand a fragility map to a human room that can refuse the machine.

Not Hidalgo’s sleepless digital delegates. Not Romania’s Ion. A wind tunnel for governance — useful before it is trustworthy; honest about where humans must drive; democracy with a right to fail in silicon first.

Data we opened for building Silico

We released the dataset as two configs on purpose. If you collapse them, you are already lying about epistemology.

text_corpus — listening residue

A lean multi-source CSV (~5.8k rows, ~1.8 MB).

Columns: source · corpus · record_type · topic · question · text · stance_or_label · agreement_pct · extra

This is not vote-preserving. Use it for argument mining and grounded language — not to claim “the public voted.”

polis_votes — geometry intact

Per conversation:

polis_votes/<id>/
  comments.csv
  votes.csv
  summary.csv

Lean pack (~25 MB): Seattle $15, Bowling Green, Brexit, Canadian electoral reform, football concussions, London youth policing, Scoop Hiveminds, San Juan farmland, vTaiwan UberX, and one Austria Klimarat conversation.

Text without votes is archaeology with the skeleton removed. Silico’s listening layer must load votes.csv when the claim is geometric.

from datasets import load_dataset

text = load_dataset("jonaskg/open-deliberation", "text_corpus")
votes = load_dataset("jonaskg/open-deliberation", "polis_votes")

What builders should notice

1. Coverage skew — climate and US local assemblies dominate; Global South residue is thin. That is a product requirement, not a footnote.
2. Habermas triples — question / opinion / mediator rows help train bridge drafts without inventing voters.
3. Remesh agreement % — thin, but useful for “how contested is this?”
4. Vote matrices are heavy — Bowling Green’s votes.csv alone is ~12 MB. Stream per conversation; don’t assume one mega-table.

How Silico uses the split

1. Listen — polis_votes when geometry matters; text_corpus for argument mining.
2. Stress-test — lenses, not souls: pastoral rights, utility, patronage, future drought, hostile capture.
3. Emit a fragility map — non-convergence allowed; no consensus card without equal-weight dissent.
4. Convene humans — /silico is a dress rehearsal surface, not a synthetic electorate.

Infrastructure between institutions and people

There is an absence in the infrastructure between institutions and people. The goal is not to make people more political. It is to make the conditions of their lives more legible — organising information that already exists and arranging it around the shape of life.

Silico is one piece of that infrastructure: continuous enough to update as the world moves; humble enough not to mistake a model run for a mandate.

What we are not launching

• A dashboard of “public will” from persona agents
• A single consensus score
• A Habermasian-only epistemology that treats elder, kinship, or harmony norms as bugs

We are launching a research programme + open data + a coming demo for disagreement-preserving rehearsal.

Explore In Silico Deliberation. Listen. Branch. Then put humans back in a room that can finally afford to tell the truth.