Taiwan is routinely described as the world’s lab for automated fact-checking. The documented record says something else: on Cofacts, replies come from volunteer editors, while in the 2024 election generative AI appeared mainly on the attackers’ side, mass-producing spam and varying the output of inauthentic accounts. The asymmetry is worth understanding before anyone copies the model.
A user in Taiwan gets an alarming message on LINE — the messaging app that dominates the island, somewhere between WhatsApp and a private social network — forwarded by a relative. She forwards it on to a chatbot. The message lands in a public database. Sooner or later someone reads it: not a model, a person. A volunteer editor on the Cofacts platform looks for sources, writes a rebuttal, publishes it. If several editors disagree, the user doesn’t get a single verdict. She gets all the answers, and sorts it out herself.
That, according to the available reporting, is the actual architecture of Taiwan’s most-cited civic tool against false information. Worth saying up front, because the story circulating abroad is a different one: Taiwan as the laboratory of automated fact-checking, language models dismantling Chinese influence operations in real time. Where artificial intelligence verifiably shows up is on the other side of the table. During the 2024 election cycle, generative AI — systems trained on vast amounts of text, able to produce plausible variants of the same message at near-zero cost — was used for large-scale spam campaigns and to vary the posts and posting times of inauthentic accounts, according to analysis by Doublethink Lab, the Taiwanese research centre founded in 2019 to study Chinese influence operations.
What Cofacts actually does
Cofacts came out of g0v, Taiwan’s civic tech community — a long-running movement of developers and activists that builds “shadow” versions of public services. The LINE bot is credited to Johnson Liang, and the idea grew from a mundane, valuable observation: MyGoPen, another fact-checking account on LINE, was being buried under messages that were overwhelmingly duplicates. The same chain letter, a thousand times over. The problem wasn’t producing a thousand checks. It was recognising they were the same thing and serving an answer that already existed.
That is a piece of labour engineering before it is a piece of technology. A shared database, reusable replies, collaborative verification. The fact that dissenting editors produce multiple answers rather than one authoritative synthesis is a political design choice, not a bug: it tells the user that the truth about a contested message can itself be contested, and leaves the judgement with her. A language model, by construction, does not do this. It produces one fluent answer, in the register of a matter already settled.
Whether and to what extent Cofacts today uses machine learning components — semantic search to match a new message against an already-verified one, deduplication, clustering — is a reasonable question, and it would be odd if a project born precisely out of the duplicates problem had never considered it. But it isn’t something you can assert without opening the repository and the documentation, and this writer hasn’t. Worth flagging, because the slide from “uses a classifier to group similar messages” to “uses AI to debunk hoaxes” is exactly what makes this debate unreadable.
The other half: Doublethink Lab is not a debunker
The second recurring confusion concerns Doublethink Lab. Founded in Taiwan in 2019, led by co-founder and CEO Wu Min Hsuan (known as “TTCat”, previously at the Open Culture Foundation), with Puma Shen as former chairperson, it is a research and analysis centre. It does not refute individual false stories: it reconstructs account networks, posting patterns, amplification chains. The 2021 report Deafening Whispers, on Chinese information operations around Taiwan’s 2020 election, is work of that kind, and on 19 January 2024 the lab published an analysis arguing that China would continue to manipulate Taiwan’s information environment with more diversified tactics, aimed at amplifying internal conflict.
That report contains the single most useful methodological point in the whole Taiwanese file, and it is about AI:
The presence of AI-generated material does not establish who created it. Attribution requires analysing account networks, posting behaviour and links to state media.
Put differently: a synthetic-text detector, assuming it works at all, answers a question nobody needed answered. Knowing a post was generated by a model tells you nothing about whether a state apparatus, a contracted agency or a bored teenager is behind it. The point travels well beyond Taiwan, and European debate about election deepfakes keeps skipping it.
2024 did not go the way it gets told
The two largest campaigns observed in 2024 were ancient in content: an alleged “illegitimate son” of candidate Lai Ching-te, and the “secret history” of outgoing president Tsai Ing-wen. No technological marvel — tabloid material. The novelty was in the industrial production of variants and in distribution: compared with 2022, direct participation by PRC state media and Weibo influencers was notably low, and spread ran mostly through unofficial accounts. Traces of state involvement were deliberately thin — which makes attribution harder and, not coincidentally, Beijing denies any interference.
Lai Ching-te won with roughly 40% of the vote, and independent assessments of the case — among them a Thomson Foundation report by Chen-ling Hung of National Taiwan University and one from the FNF Global Innovation Hub by Austin Wang — agree that the disinformation campaign did not flip the outcome. None of them, however, quantifies how much of that result should be credited to civic fact-checking and how much to other factors: entrenched pre-existing polarisation, high public awareness after years of exposure.
A Leiden Asia Centre report pushes the argument further: AI makes disinformation cheaper and more scalable, but manipulation in Taiwan draws its force mainly from pre-existing polarisation and platform incentives. It amplifies divisions rather than creating them. If that’s right, an automated verification system attacks the most visible symptom and leaves the structure untouched.
The limit no metric captures
Then there is a problem that should make anyone suspicious of coverage percentages: content circulating in closed LINE groups remains largely invisible to officials, researchers and fact-checkers. A model, however capable, cannot classify messages it never sees. The Cofacts bot is not a crawler. It is a voluntary funnel: it depends on someone deciding to forward. Which means the sample is self-selected by definition — what arrives is what somebody already doubted.
Meanwhile pressure is building on other fronts. Taiwan’s National Security Bureau reports that in 2025 the island’s critical infrastructure absorbed an average of 2.63 million Chinese cyberattacks a day — up 6% on 2024 and 113% on 2023. Ahead of local elections set for 28 November 2026, in August 2026 the governing DPP warned that Beijing could deploy deepfakes, fabricated recordings and voter microtargeting; the Central Election Commission is setting up a task force with an online platform where candidates can report suspected deepfakes, with confirmed cases referred to the platforms.
Here is the part an outside reader probably doesn’t expect. What Taiwan is building for 2026 is not a detector. It is a reporting procedure with narrow standing — candidates — and a channel to whoever can take content down. The hard part isn’t recognising the fake. It’s deciding who has the right to say it’s fake and who has the power to remove it. Automation has touched neither, and the piece of Taiwanese infrastructure actually worth copying, if any, is that one.