Taiwan's Alignment Assemblies on AI (2023–)

Background

Taiwan has a decade of experience with digital deliberation, from the vTaiwan consultations begun in 2014 to the use of the Polis opinion-mapping tool in regulating ride-hailing and online alcohol sales. In May 2023 its Ministry of Digital Affairs (moda), then led by Audrey Tang, partnered with the Collective Intelligence Project (CIP), a non-governmental organisation that had launched "Alignment Assemblies" in early 2023 to involve the public in decisions about artificial intelligence. Taiwan became the main governmental test site for the idea that citizens, not only AI companies and regulators, should help define what AI systems should and should not do.

How it is governed

The Taiwanese assemblies combined online opinion collection, face-to-face workshops and a deliberative poll.

Democratizing AI Futures (2023). An online consultation asked citizens for their views on generative AI, and the AI Objectives Institute used the open data to produce an interactive report. Deliberative workshops followed in Taipei (26 August 2023) and Tainan (2 September 2023). Participants identified concerns about the redistribution of labour freed by automation, the difficulty of identifying AI-generated information, and the effects of AI assistance on creativity and human connection, and recommended regular public hearings, provenance and traceability rules, and AI review committees within companies.

Utilizing AI to Enhance Information Integrity (2024). On 23 March 2024 moda held an online citizens' deliberative assembly using the Deliberative Polling model, organised by National Yang Ming Chiao Tung University with Stanford's Deliberative Democracy Lab. Invitations went out as 200,000 random text messages from the government's "111" SMS number; 1,760 valid responses were received and about 450 participants were chosen by stratified random sampling on residence, gender and age, with digital-sector workers over-sampled in line with their share of the workforce. On the day, 447 citizens took part in 44 small groups (12 of digital practitioners, 32 of general citizens) on the Stanford Online Deliberation Platform, which automatically manages speaking time and turn order and produces transcripts. Participants answered surveys before and after deliberation.

Economy and public goods

The assemblies were funded by government and civil-society partners and produced public reports in English and Chinese. Their output is a public good of information: structured evidence about informed public preferences on AI, available to legislators, companies and other governments.

Conflict and disputes

The 2024 assembly revealed real disagreement rather than consensus. Some participants favoured self-regulation by large platforms based on user feedback; others wanted strict state standards, heavy fines for AI abuse that undermines information integrity, and suspension of offending systems until they were reviewed. Participants leaned towards giving each user a unique anonymous digital identity and requiring platforms' generative AI to cite diverse sources, positions that raise their own privacy concerns. moda reported that the recommendations aligned with the Executive Yuan's draft anti-fraud legislation requiring online advertising platforms to keep fraudulent content out of advertisements and to disclose ads using deepfake images of people. The bill was enacted as the Fraud Crime Hazard Prevention Act, promulgated on 31 July 2024; articles 30 and 31 bar fraudulent content in online advertisements and require platforms to label ads that use deepfake or AI-generated images of people.

Critics of such exercises note that results are advisory, that the convening ministry chooses the questions and the framing, and that it is hard to trace how much a given assembly changed a law that may have been drafted anyway.

Outcomes

The assemblies showed that a government can recruit a representative online sample cheaply by text message and run structured deliberation on technical AI questions at national scale in a single afternoon. They fed into anti-fraud and platform-accountability policy and into the Executive Yuan's drafting of an AI basic law, and they became a reference model for CIP's work with AI companies, including the Collective Constitutional AI experiment with Anthropic. Their binding force remains low.

Lessons for institution design

• Text-message sortition from an official number is a low-cost way to recruit a representative sample, provided trust in the sender is high.
• Small-group online deliberation with automated moderation can scale while keeping equal speaking time.
• Link the question to a live decision. Assemblies tied to a pending bill have clearer influence than open-ended consultations.
• Report disagreement honestly. The value of deliberation lies partly in mapping where informed citizens still differ.
• Advisory status limits accountability. Without a duty to respond, governments can cite assemblies selectively.

Further reading

• moda: Alignment Assemblies
• moda: 'Utilizing AI to Enhance Information Integrity' Citizens' Deliberative Assembly
• Collective Intelligence Project: Alignment Assemblies
• CIP: Alignment Assemblies, nine months in (2023)
• Fraud Crime Hazard Prevention Act (2024), Laws & Regulations Database of Taiwan