What are Alignment Assemblies?

An effort to decide "how AI should behave" not only among companies and experts, but by bringing ordinary citizens together to discuss it and feeding the results into the decisions of AI companies and governments. A US nonprofit started it in 2023, and Taiwan's government joined in.

For now, it is mainly the companies that build AI that decide what it may answer and what it should refuse. Yet AI touches the lives of everyone who uses it. So shouldn't ordinary people also have a seat where these decisions are made? Alignment Assemblies try to answer that question with real gatherings.

In short

In AI research, making AI fit human values is called "alignment". Alignment Assemblies are efforts to decide this alignment not only inside companies, but by bringing ordinary people together (an "assembly") to discuss it, and feeding the results into the decisions of AI companies and governments [1][2].

They were started in 2023 by the Collective Intelligence Project (CIP), a US nonprofit. Taiwan's Ministry of Digital Affairs (moda) became a partner the same year [1][2].

For example

Anthropic and a constitution written by the public (2023). CIP and the AI company Anthropic asked a representative sample of US adults how AI should behave. They used Polis, a tool in which people put forward statements and vote on them. 1,002 people wrote 1,127 statements and cast 38,252 votes. From the statements supported across groups with different views, they built a "public constitution" and used it to train a language model (an AI that generates text). According to the paper, this model was less biased than a model trained on Anthropic's standard constitution on all nine social dimensions measured, with similar performance [3].

A workshop on generative AI in Taiwan (2023). moda gathered opinions on generative AI online (403 people, 8,591 votes and 126 statements as of August 1, 2023). It organized the results with Talk to the City, a tool that shows opinions like a map. Using this as material, a team from National Chengchi University held face-to-face workshops in Taipei and Tainan with 95 people in total [4][5].

Taiwan's citizens' deliberation on using AI to promote information integrity (March 23, 2024). ("Information integrity" means keeping trustworthy information intact.) From "111", the government's official text-message number, 200,000 invitations were sent at random. From 1,760 valid responses, participants were chosen so as not to skew by gender, age or place of residence, and 447 people discussed online in 44 groups. The topic was how large platforms can use AI to protect information integrity against things like AI-made fake videos (deepfakes). moda writes that it linked the results to a government anti-fraud bill covering scam ads and deepfake ads [6].

How it works

Why it started

In the May 2023 piece that launched the project, CIP's Divya Siddarth and Saffron Huang argued as follows. The usual approach leaves technology to technologists, what gets built to investors, and risks to regulators. That, they argued, does not work for a technology like AI that reaches all of society. Deciding through the market tilts toward the interests of a few companies, and policy is slow and blunt. So, the two argued, people's views need to be built into the foundations of AI itself and of the systems that govern AI [1].

The name has meaning too: "alignment" from aligning technology with people's values, and "assemblies" from assembling ordinary people [1]. Audrey Tang of Taiwan was also listed on the launch team [1].

How it runs

In a November 2023 look back, CIP wrote that every assembly starts by deciding four things [7]:

What it aims to change (the intended outcome)
Whom to bring together (the people concerned)
What to ask (the scope of discussion)
How to run it (tools and procedures)

CIP put special weight on tying each assembly to a specific decision it can influence. The tools were mainly digital, such as Polis, and it also tried combining them with face-to-face meetings [7].

Who took part

As of November 2023, CIP had held assemblies with, or worked with, the UK government's AI taskforce, OpenAI, Anthropic, Taiwan's moda and others. For the assemblies with OpenAI and Anthropic, it brought together about 1,000 people each, chosen so that age, gender, income and region matched the US population [7].

The thinking in Taiwan

At a meeting with OpenAI in October 2023, Tang argued that the smallest unit of an assembly should be the community, not the individual: "If you aggregate individual signals, no matter how you measure, you lose the plurality that is imbued in community interactions. And so we argue that the smallest unit of alignment assembly should not be an individual" [8]. In a December 2025 lecture, she described Taiwan's assemblies this way: "from one side looks like a regular citizen assembly. But on the other side, it looks like a poll. So you can actually think of it not as deliberation but rather as a survey that surveys not individuals but groups of ten" [9].

What is debated

It stays advisory. CIP itself admits that most participatory processes so far have been advisory, relying on those involved to make commitments voluntarily, and produce few mechanisms for holding anyone accountable [7].
The difficulties of digital participation. In the assembly with OpenAI, statements that made no sense, hate speech, off-topic statements and ones that could not be acted on had to be reworded or removed according to a policy. CIP also notes that relying on digital tools can leave behind people unfamiliar with them [7].
Whose voices does it represent? The official report on the 2023 Taiwan workshops stated that most participants had come of their own accord or were invited experts, and did not represent society as a whole [5]. The random invitations of 2024 were an improvement, but only 1,760 people responded to 200,000 messages, less than 1% [5].
Did the assembly make the law? The book Broad Listening (in Japanese 『政治・自治体・企業に広がるブロードリスニング』, written by volunteers from DD2030, 2026) writes of the 2024 assembly that "85% of citizens supported the view that 'platforms should bear the primary responsibility for improving information integrity'" [5]. Together with the fact that an anti-fraud law came into force four months later, it describes the deliberation as having pushed the legislation forward [5]. By contrast, the glossary of Tang and colleagues' 6-Pack of Care adds a caution: "The mini-public produced a public record of recommendations and an institutional response path while an Executive Yuan bill was already moving; it neither stood for the whole population nor enacted the statute" [10].
Should AI do the deliberating for us? In her 2025 lecture, Tang mentioned research in which individuals' views are gathered by interview and their AI "digital twins" then deliberate with one another. Leaving deliberation to machines, she said, is "exactly like sending our robots to the gym to lift weights for us" [9]. In her framing, assemblies are a "civic gym" where citizens train themselves [9].

Where it fits in Plurality

In chapter 2-2, "The Life of a Digital Democracy", the book Plurality lists Alignment Assemblies as one of Taiwan's practices. It writes that "Taiwan has pioneered convening, increasingly common around the world, of citizen participation in the regulation and steering of AI foundation models" [11]. The Japanese edition translates the term as "アライメント集会", and the Traditional Chinese edition as "對齊大會" [11][12]. For more, see Taiwan's digital democracy.

Chapter 7-0, "Policy", also lists "alignment assemblies and other augmented deliberations on the direction of technology" among the tools that let the public take part in decisions about technology. The Japanese edition translates the term here as "アライメント会議" [13].

After the book came out, Tang and colleagues' 6-Pack of Care included these assemblies in its glossary as one of its "instruments" [10]. Polis and Talk to the City, the tools used in the assemblies, are also central examples in the book's chapter on Augmented Deliberation. In preparing society's side for AI, the idea is close to reverse alignment.

Sources

CIP, "We should all get to decide what to do about AI" (2023-05-23) — https://www.cip.org/blog/alignment
moda, "對齊大會 Alignment Assemblies" (in Chinese) — https://moda.gov.tw/major-policies/alignment-assemblies/1453
Collective Constitutional AI (arXiv 2406.07814) — https://arxiv.org/abs/2406.07814 ［CC BY 4.0］
moda, "AI 未來民主化審議式工作坊" (2023, in Chinese) — https://moda.gov.tw/major-policies/alignment-assemblies/2023-ideathon/1459
Broad Listening, 11-1 "The Evolution of Deliberative Democracy in Taiwan" (in Japanese; English translation in the same repository) — https://github.com/digitaldemocracy2030/broad-listening-book/blob/f69c60c838a91b91668abf109d7093b240610139/en/11_01_taiwan.md ［CC BY-NC 4.0］
moda, "運用 AI 促進資訊完整性" citizens' deliberative assembly (2024, in Chinese) — https://moda.gov.tw/major-policies/alignment-assemblies/2024-deliberative-assembly/1521
CIP, "Alignment Assemblies: Nine Months In" (2023-11-16) — https://www.cip.org/blog/alignment-assemblies-nine-months-in
Audrey Tang, meeting with OpenAI (2023-10-10) — https://archive.tw/2023-10-10-meeting-with-openai ［CC0］
Audrey Tang, Right Livelihood lecture (2025-12-04) — https://archive.tw/2025-12-04-right-livelihood-lecture ［CC0］
6-Pack of Care, Glossary, "Alignment Assembly" — https://civic.ai/glossary/ ［CC0］
Plurality, 2-2 "The Life of a Digital Democracy" — https://www.plurality.net/v/chapters/2-2/eng/ ［CC0］
Plurality, 2-2 "デジタル民主主義の日常" (Japanese edition) — https://www.plurality.net/v/chapters/2-2/jpn/ ［CC0］
Plurality, 7-0 "Policy" — https://www.plurality.net/v/chapters/7-0/eng/ (Japanese edition: https://www.plurality.net/v/chapters/7-0/jpn/) ［CC0］

We welcome error reports and questions ("What is…?"). Submissions become public issues.

Report an error in this article ・ Ask a question