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:

1. Cluster the raw input into a manageable set of topics or opinion groups.
2. File every individual statement or vote under the themes it belongs to, so nothing is silently dropped.
3. 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.
4. 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 (open-source, 2024) 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.
• 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.

The line not to cross

The risk to watch is synthetic participation: summarising real input is sensemaking; generating opinions people never held is not.

Sources

• “Making sense of large-scale online conversations” — Jigsaw (Google, 2024): medium.com/jigsaw
• The Computational Democracy Project, on Polis and its opinion-mapping method