What is Kouchou AI?

When a city hall runs a survey, the "comments" box can fill up with thousands of answers. Reading and sorting them one by one can take staff weeks. Kouchou AI is a tool for organizing these "too many to read" voices with AI.

In short

Kouchou AI (広聴AI, "broad listening AI"; written kouchou-ai in its code) is software that turns large numbers of free-form comments into an "opinion map". Load survey answers or social media posts as a CSV file (a simple table format), and the AI places similar opinions close together, color-codes them and adds summaries [1][2].

It is free for anyone to use, and its program code is public (open-source software, or OSS). It was built by Digital Democracy 2030 (DD2030), a citizens' community. They reworked Talk to the City, a tool from the United States, for the practical needs of Japanese local governments and politicians [1][2].

Technology of this kind, for listening to many voices, is called broad listening.

For example

Utazu, Kagawa Prefecture (published 2025-04-30). In a residents' survey for the town's comprehensive plan, the town organized 396 free-form answers with Kouchou AI. The town reports that the AI finished in about 10 minutes a job of compiling results that used to take a long time [3].

Shibuya, Tokyo (published 2025-07-02). The ward analyzed 6,037 free-form answers from its residents' attitude survey. It produced an overall summary, classifications and charts by topic, and trends by generation and by area [4].

Hiroshima Prefecture (August–September 2025). The prefecture analyzed 2,953 opinions submitted by residents. Staff said they were able to classify about 3,000 free-form entries in under 30 minutes [5][15]. Done by hand, it is estimated, the work could have taken about 500 hours [5]. The prefecture's call page states that this broad listening was carried out with the cooperation of DD2030 [16].

Team Mirai (2025 House of Councillors election). The political party Team Mirai looked through proposed changes to its policies, sent in by citizens, on Kouchou AI's charts. It added a feature to jump from a dot on the chart to the original proposal in one click, and this feature was later merged back into DD2030's main version [6][7].

Examples from local governments are gathered in the article on broad listening in local governments.

How it works

How it came about

It is based on Talk to the City (TTTC), made in 2023 by the AI Objectives Institute, a US nonprofit [2].

Takahiro Anno's team used TTTC in the 2024 Tokyo governor's election, in a TV network's election coverage and in the Tokyo government's call for public input. Along the way, one problem became clear: a system was needed that people who are not engineers could use just by uploading a file [2].

So development began as a branch (fork) of TTTC's scatter-plot version, and it was released as OSS on 2025-03-16 [1][2].

Outside records say the same. The Tokyo Shimbun reported, as a plan at DD2030's launch (January 2025), that TTTC would be made usable without engineers and its analysis algorithms improved [17]. A Hiroshima Prefecture report also describes Kouchou AI as a tool DD2030 improved and developed from TTTC for use in Japan [15].

What changed from TTTC

According to the book Broad Listening (written by the DD2030 community), the main differences are these [6]:

• It runs in a browser. The original was operated by typing commands into a text-only screen, and needed a programming environment to be set up. With Kouchou AI, once a server is set up, you upload a file in the browser, run the analysis, and share the result just by sending a URL.
• Large and small groups can be seen at once. Opinions are organized in nested form: small groups within large groups.
• Only the tight groups can be pulled out. Groups that mix in weakly related opinions can be excluded.
• It favors visual clarity. In the original method, one group could end up scattered across distant parts of the chart. Kouchou AI switched to a method in which nearby dots belong to the same group. The book writes that it put visual coherence ahead of precision [6].

The processing steps

Roughly, it works in this order [8]:

1. A large language model (LLM, an AI that handles text) splits each post into individual points
2. Each point is converted into a list of numbers (a vector)
3. These are laid out on a two-dimensional chart
4. Nearby points are grouped, and groups are bundled into larger groups
5. The LLM names the groups and writes an overall summary

The book Broad Listening estimates that processing 10,000 comments with gpt-4o-mini (an inexpensive AI model from OpenAI) costs about 480 yen. More than 90% of that goes to the first step, splitting posts into points [8].

Versions and development

Version 2.0.0 came out on 2025-04-29 [9]. Version 4.0.0, on 2025-12-28, added support for more kinds of AI models and kept reports with limited access from appearing in search engines [10].

According to the book Broad Listening, 807 changes (pull requests) were merged in about a year after development began. Of 2,756 recorded commits in total, 210 (about 7.6%) were made by Devin, an AI that writes code [6].

Development also went through a rough patch. When Anno left DD2030 in May 2025 to found Team Mirai, many of the developers moved there too. Reviews of changes went unanswered, and there were gaps of as long as two and a half months before changes were merged. The book records that a small group kept holding regular meetings so that development would not stop [6].

Reuse by other organizations

Because it is open source, groups other than DD2030 also adapt it. In August 2026 the association Maniken released "Kiku AI" ("listening AI"), built on Kouchou AI with added features, such as organizing the issues in a planning document's PDF to present to residents, and asking the writers of aggressive posts to reconsider (Maniken's own announcement) [18]. On October 1, 2026, Tagawa City in Fukuoka Prefecture began using "broad listening" to collect residents' requests online, with the cooperation of a Waseda University institute on democracy (デモクラシー創造研究所) and Maniken (Yomiuri Shimbun; the article does not name the tool) [19]. According to Maniken, the tool is Kiku AI [20].

What is debated

• The makers' own warnings. DD2030's official page asks users not to rely too heavily on the analysis. Its point, in substance, is that online opinion is easily skewed, so the results are not meant to be turned directly into policy, and decisions should be made only after users have checked them thoroughly [2]. General cautions about bias and misreading are in the article on broad listening.
• Support and opposition end up in the same group. Supporting and opposing opinions on the same topic tend to look like "similar text" to the AI, and can be lumped together. The book Broad Listening treats this as an unsolved problem and describes several workarounds [8].
• An assessment by a local government that used it (Hiroshima Prefecture). The report by the prefecture's Hiroshima AI Lab says the tool can turn large amounts of free text into an overview in a short time, and that processing could be adjusted flexibly through prompts, but also that Kouchou AI is not all-powerful. It notes that AI processing can change the original form of residents' opinions and that dissimilar topics can end up in one cluster, and it recommends checking the original texts rather than taking results at face value, and publishing the prompts and the AI used so that third parties can verify them. This evaluation was carried out with DD2030's support [15].
• It is slow to spread among local governments. DD2030 reached out to local governments, but weak incentives, gaps in enthusiasm between mayors and frontline staff, staff transfers and a lack of engineers made it hard to move from trials to full use. In many local governments, the problem came before analysis: not enough opinions were coming in [6].
• Criticism that responsibility is unclear. A note article by a member of the public argues that OSS-based citizen participation, such as Kouchou AI and Idobata, leaves quality, accountability and guarantees of maintenance unclear [11].
• Not necessarily Team Mirai's tool. In February 2026, Shinta Nakayama, the lead author of the book Broad Listening, wrote on X that Team Mirai no longer uses either TTTC or Kouchou AI [12].

Where it fits in Plurality

The book Plurality does not mention Kouchou AI by name. But in chapter 5-4, "Augmented Deliberation", it gives Talk to the City, on which Kouchou AI is based, and its use in Anno's governor campaign, as examples of broad listening [13]. Kouchou AI can be seen as that TTTC turned into a "component anyone can use" in Japan.

In the book Broad Listening's classification, it is a leading example of "broad listening as a tool". Its role differs from discussion spaces such as Idobata or Polis [14]. Its developer, DD2030, builds it as a public good, offered neutrally to political parties and local governments [6].

Further reading

1. Broad Listening, 10-0 "Development Work on 広聴AI (Kouchou AI, Broad Listening AI) by DD2030" (Japanese and English, book, free to read) — https://broadlisteningbook.com/
2. Broad Listening, chapter 13 "Reading the Implementation of 広聴AI (Kouchou AI, Broad Listening AI)" (Japanese and English, book, technical) — https://broadlisteningbook.com/
3. DD2030, "広聴AI" introduction page (in Japanese, web) — https://dd2030.org/kouchou-ai/
4. The kouchou-ai repository and manual (in Japanese, software) — https://github.com/digitaldemocracy2030/kouchou-ai
5. Documentation site (in Japanese, software guide) — https://digitaldemocracy2030.github.io/kouchou-ai/

Sources

1. kouchou-ai README (in Japanese) — https://github.com/digitaldemocracy2030/kouchou-ai
2. DD2030, "広聴AI" introduction page (in Japanese) — https://dd2030.org/kouchou-ai/
3. Utazu Town, "ブロードリスニングの試行" (A trial of broad listening) (in Japanese) — https://www.town.utazu.lg.jp/page/4114.html
4. Shibuya City, press release (2025-07-02, in Japanese) — https://www.city.shibuya.tokyo.jp/kusei/hodo/hodo-2025/hodo_20250702.html
5. Broad Listening, chapter 8, section 2, on Hiroshima Prefecture (in Japanese; English translation in the same repository) — https://github.com/digitaldemocracy2030/broad-listening-book/blob/f69c60c838a91b91668abf109d7093b240610139/en/08_02_hiroshima_case.md ［CC BY-NC 4.0］
6. Broad Listening, 10-0 "Development Work on 広聴AI (Kouchou AI, Broad Listening AI) by DD2030" (in Japanese; English translation in the same repository) — https://github.com/digitaldemocracy2030/broad-listening-book/blob/f69c60c838a91b91668abf109d7093b240610139/en/10_00_dd2030_kocho_ai_development.md ［CC BY-NC 4.0］
7. Broad Listening, chapter 6, section 1 "Team Mirai's Efforts in the 2025 House of Councillors Election" (in Japanese; English translation in the same repository) — https://github.com/digitaldemocracy2030/broad-listening-book/blob/f69c60c838a91b91668abf109d7093b240610139/en/06_01_team_mirai.md ［CC BY-NC 4.0］
8. Broad Listening, chapter 13 "Reading the Implementation of 広聴AI (Kouchou AI, Broad Listening AI)" (in Japanese; English translation in the same repository) — https://github.com/digitaldemocracy2030/broad-listening-book/blob/f69c60c838a91b91668abf109d7093b240610139/en/13_kocho_ai_tech_stack.md ［CC BY-NC 4.0］
9. DD2030, announcement "広聴AI v2.0.0" (in Japanese) — https://dd2030.org/
10. DD2030, announcement "広聴AI v4.0.0" (in Japanese) — https://dd2030.org/
11. "『チームみらい』が唱える『Plurality』は看板だけにならないか" (Will the "Plurality" that Team Mirai advocates end up as just a signboard?) (note, in Japanese) — https://note.com/ippanjin_2025/n/n299555fb4908
12. Shinta Nakayama, post on X (2026-02-07, in Japanese) — https://x.com/tokoroten/status/2020004238718570604
13. Plurality, 5-4 "Augmented Deliberation" — https://www.plurality.net/v/chapters/5-4/eng/ ［CC0］
14. Broad Listening, chapter 1 "What Is Broad Listening?" (in Japanese; English translation in the same repository) — https://github.com/digitaldemocracy2030/broad-listening-book/blob/f69c60c838a91b91668abf109d7093b240610139/en/01_what_is_broad_listening.md ［CC BY-NC 4.0］
15. Shota Emori (Hiroshima AI Lab), chapter on broad listening with Kouchou AI, in the FY2025 AI Lab annual report (2026-03, in Japanese) — https://www.pref.hiroshima.lg.jp/uploaded/attachment/658524.pdf
16. Hiroshima Prefecture, "デジタル化で描く未来の広島は？" (call for opinions page; in Japanese) — https://www.pref.hiroshima.lg.jp/site/hiroshima-dx-torikumi/hiroshimanomirai.html
17. Tokyo Shimbun (2025-01-30), on the launch of Takahiro Anno's "Digital Democracy 2030" (in Japanese) — https://www.tokyo-np.co.jp/article/382701
18. Maniken, announcement of "Kiku AI" for local governments and assemblies (2026-08-01, in Japanese) — https://maniken.jp/blog/3658/
19. Yomiuri Shimbun (2026-10-01), Tagawa City starts "broad listening" (in Japanese) — https://www.yomiuri.co.jp/local/kyushu/news/20261001-GYS1T00051/
20. Maniken, "Kiku AI" pilot study starts in Tagawa City (2026-10-01, in Japanese) — https://maniken.jp/blog/3876/