How to Read a Crowd Without Flattening It

Here is a problem that, until fairly recently, had no good answer.

Say you want to know what a town of a million people thinks about a new development. You can have real conversations with ten of them, an hour each, long enough to understand not just what they want but why. At the end you'll know those ten people well and have no idea whether they speak for anyone else. Or you can send all million a survey with a single box to tick, a new park: yes or no, and get back a clean percentage that tells you nothing about what kind of park, for whom, or whether a park was ever the real question. Listen deeply and you can't listen widely; listen widely and you can't listen deeply. The two have always pulled against each other, because you have only so much time and attention to give, and you can spend it on a few people in depth or on many people thinly.

Broad listening is a recent attempt to break that trade-off, and the easiest way to understand it is by its opposite.

Broadcasting is the technology of one voice reaching many. Radio, television, the front page, the post shared a hundred thousand times. We have spent a century getting very good at it, to the point where almost anyone can now address a crowd from their phone. The reverse was never built. There has never been an equally good way for a single person to hear a crowd, to take in what ten thousand people are saying and come away actually understanding it. Past a certain number, listening simply breaks. The voices exceed what any mind can hold, and you fall back on blunt instruments like the poll, the petition, and the like count, which scale only because they crush each person down into a number.

What broad listening proposes is that the missing half can finally be built, not by reducing people to numbers but by keeping their own words and using software to find the structure inside them. The basic version is simple to describe. People write short statements in their own words and react to one another's, usually with nothing more than agree, disagree, or pass. The software then reads across all of it at once. It groups together statements that mean roughly the same thing, finds where people separate into camps, and surfaces the statements that draw agreement from across those camps. What comes back is less a tally than a map.

That final step quietly reverses the logic of most of the internet. A comment thread or a feed is built to reward whatever provokes the strongest reaction, so the most divisive material rises and the broad, unexciting agreement underneath stays invisible. A listening tool is built to do the opposite. It goes looking for the things that people who agree on nothing else can both accept, and lifts those to the top. A handful of systems work on this principle, each with its own strengths: Polis, the open-source platform Showra is built on; Talk to the City; Kōchō AI; and Google Jigsaw's Sensemaker.

The idea is older than the tools, and for a long time it stayed only an idea, because reading thousands of free-text opinions and making sense of them took a room full of trained analysts and months of work. Three things changed at roughly the same moment. Software learned to convert the meaning of a sentence into numbers, so a machine could recognize that two differently worded statements were saying the same thing. It learned to take that dense numerical data and lay it out as a picture a person can actually read. And large language models arrived, the technology behind ChatGPT, able to read and summarize text in bulk. Together these collapsed the months of analyst work into an afternoon.

None of which makes the result automatically true, and this is the part to be careful about. A map of a conversation is not the conversation, and the people who turn up to one are never a clean sample of everyone. A small and motivated group can flood a discussion and emerge looking like a majority. A confident, well-drawn chart can persuade you that you've heard a public when you've heard a self-selected sliver of it. The people who wrote the Broad Listening book put it about as plainly as it can be put: "a word cloud is not analysis." A portrait of a community that quietly leaves out its minorities isn't a neutral summary. It's a flattering lie. So the honest way to use any of this is to remember what it actually is. It is a way to hear the shape of a debate, qualitative and rich, and it belongs beside polls and other methods rather than in place of them.

That caution is most of the real work, and it's the part Showra takes most seriously. The aim is to let a group of people see itself honestly, both what it agrees on and where it truly divides, without quietly erasing the ones who don't fit the majority. The clustering and the maps are the easy part. Whether the picture tells the truth about the people in it comes down to a hundred unglamorous decisions about how voices get gathered, weighted, and shown. Get those right and you have something democracy has rarely had: a way for a crowd to listen to itself. Get them wrong and you've only built a more convincing way to ignore people.

For what this looks like on a real and bitter dispute, see the story of Taiwan and Uber.

---

Sources and further reading: the Broad Listening book by Digital Democracy 2030, the most accessible guide to this topic and the source of the depth-and-scale and broadcast framings used here (https://broadlisteningbook.com/); Plurality, by Audrey Tang, E. Glen Weyl and the ⿻ community, where the idea is laid out (https://www.plurality.net/); and the Computational Democracy Project, the team behind Polis (https://compdemocracy.org/).