Listening to a public at scale

What is "sensemaking"?

When thousands of people talk about an issue, nobody can read it all. The usual shortcuts — polls, hashtag counts, viral posts — either flatten opinion into multiple-choice, or amplify whoever shouts loudest.

Sensemaking is a middle path: use AI to read everything, organize it into themes and distinct opinions, keep people's actual words attached as evidence, and present it so a human can explore what a whole community thinks — including the quiet, thoughtful voices.

The research community calls the collection side of this "broadlistening" — the opposite of broadcasting: one institution listening to many people, instead of speaking to them. The term comes out of the Plurality movement around Audrey Tang (Taiwan's former digital minister) and Glen Weyl.

Whose tools are we standing on?

Project Prelude: the issue at hand

Prelude studies media propaganda targeting Iranians — how audiences talk about the influence of outlets like Iran International, Manoto, BBC Persian, state TV (صداوسیما), and social-media influencers, across 2016–2026 including the war period.

The az-democracy hashtag study (#از_دموکراسی_بگو, 4,180 tweets) was our prototype: we ran it through every stage first, end to end, so Prelude runs on a proven path.

Two ways of listening, one analysis machine

Collection has two channels. First we harvest what people already say publicly — broad, cheap, and it maps the terrain. That map is what makes good interviews possible: the themes it reveals become the questions our AI interviewer asks in depth. Both sources merge into one dataset.

That merged dataset feeds the analysis machine: organize it into themes, opinions and quotes → publish an explorable report → test which concrete statements could command broad agreement → validate with real votes. Any stage can send us back to collect more.

Harvesting the discourse

Purpose

Gather what people already say publicly, at scale, without prompting them.

Input

Keyword & phrase queries in Farsi (outlet names, propaganda terms, media-criticism phrases).

Tools

prelude/collect.py against the twitterapi.io service (X/Twitter data); categorize.py (Claude Haiku) labels each tweet's role and dimensions; export.py packages the dataset.

Output

A clean CSV — one row per statement: participant_id, survey_text — plus per-dimension slices and keyword lists (bilingual README included).

Every statement gets tagged with which of six questions it speaks to: the mechanisms of propaganda, disputed facts, its impact, the discourse around outlets, who is affected, and proposed solutions. Those six dimensions echo through every later stage — starting with the second collection channel, where they become the interviewer's question plan.

An AI interviewer that listens in Farsi

Purpose

Harvesting is broad but shallow — one post per person, skewed toward complaint. The second channel invites people into a real ~15-minute conversation with adaptive follow-ups, to capture reasons, lived experience, and the things people don't volunteer on social media (like solutions).

Input

An interview plan prepared from the harvest: the six Prelude dimensions became six questions, each with follow-up guidance, and the interviewer is "briefed" with background knowledge on the topic — read the full interview guide (bilingual, with every question and probing path).

Tool

Team Mirai's interview platform, which we fully localized: Persian interviewer persona, right-to-left Persian interface (585 translated strings), running on Claude. Safety is built into the prompt: no pressing for identity or location, any question can be skipped, anonymity by default.

Output

Per-person transcripts and structured reports (summary, stance, background, key opinions with verbatim quotes) — exported into the same format as the harvest and merged with it into one dataset that feeds everything downstream.

From 28k posts to a map of opinion

Purpose

Discover what themes exist and what distinct positions people take — with receipts.

Input

The merged collection CSV (harvest + interview statements) + a short "context file" telling the AI who's speaking and what the study is about (this framing file matters more than any setting).

Tool

Jigsaw's topic-discovery stage, running on Claude. It reads every statement, builds a topic taxonomy, writes distinct opinions under each topic, and extracts verbatim quotes as evidence for each opinion. It even grades its own extraction quality.

Output

A structured CSV: topic → opinion → quote → who said it.

Prototype benchmark: 4,180 tweets became 30 topics, 226 opinions, and 9,132 quote-to-opinion links. Rule of thumb: read the topic list before trusting anything downstream.

Finding the constructive voices

Purpose

Not all quotes are equal. This stage scores each one for constructive quality — does it give reasons, acknowledge other views, invite dialogue?

Input

The organized topic/opinion/quote table.

Tool

A cheap, fast model (Claude Haiku) scoring tens of thousands of quotes.

Output

The same table plus a quality score per quote.

This is what lets the final report surface thoughtful contributions instead of the loudest or most repeated ones — the quiet middle of a discourse becomes visible.

An explorable, bilingual report

Purpose

Make the whole discourse browsable by a human — themes, summaries, and real quotes side by side.

Input

Scored quotes + AI-written topic summaries.

Tool

Jigsaw's web report builder, which we extended with right-to-left layout, a Farsi translation layer, and an English/فارسی switcher.

Output

A static website — shareable as a link, no servers to maintain.

Prototype example (az-democracy): az-democracy-report.pages.dev — English at /, Farsi at /index-fa. Prelude gets the same treatment once its analysis run completes.

From "what people said" to "what they'd endorse"

Purpose

Turn the opinion map into concrete, votable statements and predict which could command broad agreement.

Input

The opinion map from stage 2.

Tool

Jigsaw's deliberative track: it distills opinions into short "propositions", then runs a simulated jury — the AI role-plays real participants (each persona built from that person's own posts) voting Agree/Disagree — and aggregates the votes with real election math, including a proportional method so minority viewpoints stay represented.

Output

Ranked statement slates per topic, with simulated approval rates.

From simulated votes to real ones

The simulated jury is our preliminary analysis — fast and nearly free, it tells us which propositions are even worth putting in front of people. The real test happens on Showra, our own implementation of Polis.

The model

Polis is the open-source deliberation format proven in Taiwan's vTaiwan process and Jigsaw's Bowling Green project: participants see one statement at a time, vote Agree / Disagree / Pass, and can contribute statements of their own. Clustering reveals opinion groups and — most valuably — the statements that bridge them. Showra brings that format to our community, already built and ready to use.

Input

The top-ranked propositions from the simulated jury become Showra's seed statements.

Output

Real votes from a self-selected but much wider community — anyone with the link can participate, not just the Twitter users whose posts we harvested.

So the division of labor is: simulated jury to shortlist cheaply, Showra to measure for real. Where the two disagree, that gap itself is a finding — it usually means the harvested discourse doesn't represent the wider community.

The az-democracy dry run

Before pointing this at Prelude, we ran the analysis path — harvest through simulated jury (stages 1a → 4) — on the smaller hashtag study:

See the results yourself: the explorable report — az-democracy-report.pages.dev (فارسی) — and the propositions + jury-simulation analysis — /propositions (فارسی).

Every stage ran on Claude (the toolkits were built for other providers — the adaptations are ours and reusable), and the interviewer has hosted real Persian test sessions. Not yet exercised: Showra validation and interview-fed re-analysis — Prelude will be the first project to run the full process.

What this process can't tell us

• Who's missing: harvested posts represent people who post — not lurkers, not the offline public. Interviews reach only people we can invite safely.
• AI bias flows downstream: topic discovery, scoring, translation, and jury simulation all inherit the model's blind spots. We spot-check every stage against raw quotes.
• Simulated agreement is a hypothesis: jury results shortlist statements worth testing with real people; they are never findings on their own.
• Participant safety comes first: for this population, anonymity, consent framing, and never collecting identifying details are requirements baked into the interviewer — not options.

Prelude status

• Done: 28,619-statement dataset collected, labeled across six dimensions, and exported in the pipeline's input format (bilingual documentation included).
• Done: Persian AI interviewer configured and verified for the propaganda topic — six questions mirroring the six dimensions.
• In progress: the main analysis run (topics → opinions → quotes → scores) on the full dataset.
• Next: bilingual Prelude report; proposition/jury pass; then invite-based deep interviews where the harvested data is thinnest — likely the solutions dimension, which is always underrepresented in complaint-driven discourse.

Questions → Hadjar. This deck: docs/prelude-sensemaking-deck.html in the sensemaking workspace.