On 19 October 2026, Taiwan's ruling Democratic Progressive Party held a forum in Taipei on countering Chinese disinformation ahead of the 28 November local elections. According to the Taipei Times, speakers urged the government to build AI-related databases and screening systems for AI-generated content and to strengthen media-literacy teaching. No new platform law was named. The conversation has moved from legal duties on platforms toward technical filters and public education.
The case for acting is real
The threat is not imaginary, and a serious regulator would not wave it away. Reuters, reporting on Taiwan's National Security Bureau, says the bureau counted more than 2.5 million pieces of disputed or false online content targeting Taiwan in the first nine months of 2026, about 48% more than the roughly 1.69 million a year earlier. The bureau says the content spreads mainly on Facebook, Threads, YouTube and online forums, and that China is using AI-generated video and fake accounts. Beijing has previously denied interfering in Taiwanese elections. A government facing that volume has a legitimate interest in speed, and a defender of stricter platform duties would say that voluntary measures will always lag a state-backed campaign.
What the existing law actually covers
Taiwan does have an election-period deepfake rule. According to Tech Policy Press (1 October 2026), Article 51-3 of the Public Officials Election and Recall Act, added by a May 2023 amendment, lets a targeted candidate request a police forensic examination of manipulated audio or video. If police verify it as a deepfake, the candidate can demand removal, and the commentary says platforms must act within two days or face fines from the Central Election Commission. I could not retrieve the statute text itself, so these details rest on that analysis.
The same analysis identifies the limits. The mechanism applies only from the official election notice, which it describes as typically 40 to 120 days before polling. It protects only candidates, prospective candidates and parties to a recall vote. It would not have reached the 2023 Ko Wen-je audio, which surfaced outside a campaign window. It also would not cover a non-candidate such as the president. Police findings are, in the author's words, notifications of fact that candidates cannot appeal in court, and serving a takedown request on an offshore platform within a campaign is often impractical. Most striking, the author reports that as of May 2026 the Central Election Commission had received zero requests under the mechanism.
The Taipei Times reported in September that the Ministry of Justice plans faster forensic tracing, prompt removal of deepfakes, and prosecutions, with penalties of up to seven years for spreaders and penalties for platforms that ignore removal orders. That is a tougher enforcement posture, but it still operates on the same narrow, candidate-initiated trigger.
Why filtering is the wrong centre of gravity
A mechanism nobody uses is a design problem, not a sign that disinformation is absent. Moving the debate to AI filtering databases risks repeating the mistake in a worse form. A government-sponsored screening system for AI-generated content must decide what counts as synthetic, what counts as deceptive, and what counts as satire. Detection tools produce false positives, and when the screen is tied to political speech during an election, the cost of an error lands on lawful speakers. Tech Policy Press already flags that Article 51-3 sets no objective standard for deception and leaves the call to police. An AI filter would only automate that ambiguity.
Taiwan also has a recent example of how platform obligations can be done with narrower aims. The Ministry of Digital Affairs fined Meta NT$50,000 for each of two Facebook ads (NT$100,000 in total) for failing to disclose the publisher and sponsor, under Article 31 of the Fraud Crime Hazard Prevention Act. That rule regulates a verifiable fact, who paid, rather than whether content is true. The approach needs no truth adjudicator and is hard to turn against dissent.
A proportionate agenda
The evidence supports four modest steps rather than a new censorship layer:
- Fix the window problem. If deepfake takedown is justified, the harm does not stop at the official election notice. Extend eligibility for forensic review to the year-round period, while keeping the claim narrow: provably fabricated depictions of real people.
- Add an appeal path. Candidates and uploaders should both have access to judicial review of forensic findings, not only a police notification.
- Require a local legal contact. Large foreign platforms should name a domestic representative who can receive requests, which answers the offshore-service problem without content mandates.
- Publish transparency data. Count requests, removals and reversals, so lawmakers can see whether a rule works before they extend it.
Media literacy deserves support, but it is slow, and it should not be sold as a substitute for a working legal channel for demonstrable fabrications. Equally, a database of approved or flagged content is not a literacy programme.
Conclusion
The forum shows where Taiwan's debate is heading, not a flaw in the diagnosis. The country faces a growing, documented flow of hostile content, and its one dedicated deepfake tool is narrow and, by one account, unused. The pro-speech, pro-innovation response is to repair that tool with a wider but still tightly defined scope, due process and transparency, instead of building AI filters that decide political truth for everyone.