AI sensemaking

Open a public consultation to a whole city and you hit a wall: tens of thousands of comments, far more than any team can read, let alone synthesise fairly. AI sensemaking is the method that has grown up to meet that wall — using large language models to take a mass of civic input and turn it into something legible: the themes people raised, where they converge, and where they split. It is the machine-led counterpart to the human, small-group craft of civic listening, and increasingly the back end of large participation platforms.

The pattern

However the input arrives — votes on statements, free-text survey answers, transcripts of many small conversations — the method runs through the same four moves:

Cluster the raw input into a manageable set of topics or opinion groups.
File every individual statement or vote under the themes it belongs to, so nothing is silently dropped.
Summarise each theme in plain language, with the summaries grounded in citations back to the original statements, so a reader can check the machine’s claim against what people actually said.
Surface the shape of the conversation: the points of broad consensus, and the places of genuine division — not averaging disagreement away but making it visible.

The pay-off is speed and scale: what once took weeks of manual coding can run in minutes, which is what makes consultations of hundreds of thousands of people tractable at all.

Where it shows up

Jigsaw’s Sensemaker takes the statements and votes from a Polis-style conversation and uses Google’s Gemini to find the topics, file every statement, and write cited summaries flagging agreement and disagreement. It was the engine behind Bowling Green, Kentucky’s 25-year planning consultation, turning more than a million votes into twelve themes, and Jigsaw has since run it at state level (a 1,000-resident statewide survey in Oklahoma) and nationally (the We the People deliberation, which by Jigsaw’s account distilled 2,400 Americans across all 435 congressional districts into 26 summary statements). In 2026 Jigsaw open-sourced the full developer library — its categorisation, predictive-agreement, quote-ranking, and recursive-summarisation pipelines — so civil-society groups can run the method on their own infrastructure.
Convoca adds an accountability layer to the same pattern: human-led AI sensemaking wired to tools that close the loop — making traceable who was heard, who wasn’t, and what changed — so a synthesis leads somewhere a resident can check rather than vanishing between planning cycles. Built by an MIT Media Lab team and used by New York City’s planners.
Polis itself does a lighter, statistical version: its opinion-group maps already surface the statements that win support across clusters — the uncommon ground — before any LLM is involved.
Open tools in the same family (such as Talk to the City) apply the pattern to free-text and interview transcripts rather than vote matrices.

The catch

AI sensemaking buys scale at the cost of the small-group, story-first texture that civic listening preserves, and it inherits every weakness of the model doing the summarising: it can flatten nuance, miss a minority framing, or assert a consensus that isn’t there. That is why even Sensemaker’s designers keep a human in the loop to check and correct the output before it travels, and why grounding every summary in citations matters: it lets people audit the machine rather than trust it. Treated that way — as scaffolding for human judgement, not a replacement for it — it is one of the few realistic answers to hearing a whole community at once.

Sources

“Making sense of large-scale online conversations” — Jigsaw (Google, 2024): medium.com/jigsaw
“One year on: the momentum behind Jigsaw’s Sensemaking AI” — Jigsaw (Google, 2026), on open-sourcing the library and the Oklahoma and We the People deployments: medium.com/jigsaw
Convoca — community-listening and sensemaking infrastructure, from an MIT Media Lab team
The Computational Democracy Project, on Polis and its opinion-mapping method