Skip to content
Citizen Infrastructure

Taiwan's Alignment Assemblies: deepfake scam ads (2024)

In 2024, deepfake “investment advice” ads — fake videos of celebrities like Nvidia’s Jensen Huang — were flooding Taiwanese social media and defrauding people. Rather than censor, Taiwan ran an Alignment Assembly: it texted 200,000 random citizens, gathered a statistically representative ~450 of them into AI-facilitated online rooms, and reached cross-party agreement on a package of measures that became law within months. By 2025, fraudulent ads had largely vanished from Taiwan’s feeds.

At a glance

Section titled “At a glance”
Process typeAI-facilitated deliberation / alignment assembly (random recruitment + representative mini-public)
When2024 (“Utilizing AI to Enhance Information Integrity”)
WhereTaiwan, online
Run byTaiwan’s Ministry of Digital Affairs (moda), with the Collective Intelligence Project and Stanford’s Deliberative Democracy Lab
Recruitment200,000 SMS sent from the trusted government “111” number; ~thousands volunteered; ~450 chosen to mirror the population by region, age and gender
Format45 rooms of 10 — about 30 rooms of lay citizens, 15 of practitioners — facilitated by AI, not humans
The questionWhat should be done about deepfake fraud and information integrity online?
Result85%+ agreement across party lines on a package of anti-fraud measures; fast-tracked into law

What happened

Section titled “What happened”

The trigger fit Taiwan’s rule of thumb for when to convene the public: a problem that is urgent and not the sole job of one government department. People who got the 111 text were asked one plain question — how do you feel about information integrity online, and what should we do? From the thousands who volunteered, ~450 were selected as a representative microcosm and seated in 45 rooms of ten, like small Zoom calls.

Crucially, the room itself facilitates — there is no human moderator and no AI avatar pretending to be one. As people speak, a live transcript appears; a gentle “poke” nudges anyone who’s been too quiet; and AI surfaces the uncommon ground between rooms. This deliberately up-regulates the quiet and down-regulates the chatty, the opposite of social media, where the loudest dunk wins. This is AI as assistive intelligence, not as a replacement for the human conversation.

Inputs → outputs

Section titled “Inputs → outputs”

In: citizens’ own feelings and ideas about deepfake fraud, deliberated across 45 rooms, with experts interpreting between sessions.

Out: concrete proposals that won broad agreement — for example, that an ad not signed off by the named person should be presumed a scam; that platforms which fail to verify a signature should be jointly liable for losses; and that platforms ignoring the rules could have their connections throttled. More than 85% of participants agreed on the package regardless of party, and the rest could live with it.

Impact

Section titled “Impact”

Because the result was visibly the public’s own conclusion, big-tech firms could not lobby against it — “there is no pro-fraud party” — and the draft fast-tracked through a parliament where no party held a majority, a rarity in Taiwan. The reported outcome: throughout 2025, fraudulent and deepfake ads had effectively disappeared from Facebook and YouTube in Taiwan. It also builds on a decade of habit — over a hundred prior consultations since the vTaiwan / Uber process — and a backstop where 5,000 counter-signatures on the national participation platform can force another round.

How it went

Section titled “How it went”

Research with Stanford’s Deliberative Democracy Lab found the effect is durable: exposure to such a citizens’ assembly still shaped how people weighed issues — over party loyalty — a year later, a kind of pre-bunking inoculation. The honest limits: it leans on a decade of public trust that newer adopters won’t have, and on a government willing to actually enact the result. The same model has since been exported — to Engaged California and Bowling Green, Kentucky.

Sources

Section titled “Sources”