Taiwan's Election Needs a Deepfake Law That Works

As voters in both the US and Taiwan head to the polls this November, Taiwan’s experience regulating deepfakes carries lessons for lawmakers in the US and beyond.

As a principal target of information manipulation by China, Taiwan stands on the front line of cognitive warfare, the sustained influence operations aimed at shaping public opinion and influencing election results. In May 2023 — anticipating the weaponization of generative AI ahead of the January 2024 presidential election — Taiwan’s legislature amended its election laws to establish a dedicated mechanism to regulate political deepfakes. Under Article 51-3 of the Public Officials Election and Recall Act, candidates targeted by digitally manipulated audio or video may request a police forensic examination. If the content is officially verified as a deepfake, a candidate can demand that online platforms remove it. Platforms must act within two days, and noncompliance is subject to fines imposed by the Central Election Commission (CEC).

Free speech has been protected in Taiwan since democratizing in the 1990s, under doctrines borrowed largely from US constitutional law, so the deepfake election law was designed with several safeguards in place to balance freedom of expression. Unfortunately, these protections have also made the law largely ineffectual. In three years — in contrast to separate laws to curb sexual and scam deepfakes — not one election-related deepfake has ever been reported through the official mechanisms established by the act.

As the next election approaches, the risk of political deepfakes undermining democratic institutions is rapidly escalating. Enacting robust legislative reforms is an urgent necessity before synthetic manipulation irrevocably damages election integrity.

A voice that was not the president’s

“My name is benzopyrene. I am a Group 1 carcinogen, definitively proven to cause cancer in humans.” The voice is nearly indistinguishable from that of Taiwan’s President Lai Ching-te. It narrates a short video released on July 24 by Wei Chun-yu, a former director of the opposition party Kuomintang (KMT) and a creator of satirical political videos. The video was made to rally support for a KMT protest the next day over the government’s handling of a scandal over cooking oil contaminated with the carcinogen.

Within two days, according to Wei, police officers appeared at his home asking whether he had made the video and whether he would take it down. The National Police Agency has said the visit was only to verify facts and that officers never asked for a removal. An official analysis by the Criminal Investigation Bureau determined that the video’s narration bore the hallmarks of AI-generated audio. The Interior Ministry said the video might constitute forgery, defamation, and violations of personal-data law.

Today, the video remains online, but it has split Taiwan’s politics along party lines. The Presidential Office argued that parody is protected, but that the unauthorized use of a person’s voice is not — while the KMT accused the government of selective enforcement and attempted censorship. The controversy has reignited a question that US lawmakers and courts are also struggling to answer: how can a democracy regulate AI-generated political speech without trampling freedom of expression?

What US courts have said

Several US states have enacted deepfake election laws over the past few years, but not all have survived. In California, for example, a federal judge preliminarily blocked AB 2839 in Oct. 2024, within weeks of its enactment and permanently enjoined it in Aug. 2025 on First Amendment grounds. The law would have allowed candidates, election officials, and even any voter who received a materially deceptive deepfake to sue whoever distributed or reposted it, and to seek a court order to halt its distribution as well as seek damages. That same month, the same judge held that its companion law, AB 2655, which would have required large platforms to remove or label materially deceptive election content, was preempted by Section 230 of the Communications Decency Act, which protects online platforms from liability for third-party content and permanently enjoined it as well. California’s appeal is pending before the Ninth Circuit.

Taiwan’s law, by contrast, has stood for more than three years without a single constitutional challenge, and without a single case to test its merits.

Why the law was never used

On paper, the Taiwanese deepfake law’s direct statutory intervention appears robust. The CEC must be copied in on every removal request by law.

Earlier this year, after I urged the CEC’s open data advisory group to publicly disclose the number of requests they received, they confirmed that they had received zero requests as of May 2026. Taiwanese press coverage reaffirmed this last week.

AI-generated political content is prolific in Taiwan, but three legislative and operational deficiencies explain why the mechanism has gone unused.

First, the Act is too narrow, in both whom it protects and when it applies, even if the purpose is to stay within the bounds of constitutionally protected speech. It protects only candidates, prospective candidates, and the parties to a recall vote — and only from the date of the official election notice, which is typically 40 to 120 days before election day.

Meanwhile, foreign cognitive warfare is continuous and does not follow campaign calendars. High-profile deepfake incidents in recent years occurred outside this statutory window, such as an incriminating fabricated audio recording of Taiwan People’s Party (TPP) presidential candidate Ko Wen-je that circulated in August 2023. This leaves many deepfaked candidates without recourse under the Act.

Second, the framework is undermined by both definitional ambiguity and technological limitations. Article 51-3 defines a deepfake as content produced by digital composites or other technological methods, which “convincingly performs speech and actions that are not those of the actual person.” It provides no objective standard for determining when that threshold is met. Instead, it hands the authority to determine what counts as deception or legitimate political satire to the police rather than the judiciary.

In another deepfake scandal from 2023 involving a video that purportedly showed then-Vice Premier Cheng Wen-tsan entering a hotel room with a woman, the police said they could not conclusively document the inauthenticity of the video (a requirement of the deepfake election law). Detection software can flag whether a clip is likely to be synthetic, but is limited in assessing whom it depicts. If a creator denies that a video impersonates a particular candidate, it is difficult for authorities to prove otherwise. Moreover, because police findings are classified as “notifications of fact” rather than as formal administrative decisions, candidates have no right to appeal them in court.

Third, the framework suffers from a jurisdictional gap with respect to offshore platforms. The Act places the burden on the candidate to serve a formal takedown request directly to the relevant platform, but most of the major platforms claim to reside outside of Taiwan’s jurisdiction. TikTok’s corporate entities are based in Singapore, while Meta’s local subsidiary (Taiwan Facebook Limited) is only registered for market research and advertising services — as is common for global platforms. For example, in earlier defamation and consumer rights cases concerning Meta, Taiwanese courts have repeatedly held that legal requests must be sent to the parent company in the US. For a political candidate to serve legal documents to a foreign entity within the compressed time frame of a campaign is practically impossible.

By contrast, Taiwan’s sexual-imagery and anti-fraud laws, both of which are routinely used, place the burden of notifying platforms to take down content on the government rather than on claimants. However, deepfakes make up only a small portion of those cases, and nudity or fraud is easier to determine than political satire.

Ahead of Taiwan’s local elections on November 28, the Ministry of Digital Affairs (MODA) has launched a site called “Election Deepfake Report,” to help relay police-verified requests to Meta, Google, LINE, and TikTok through official channels. While the site may streamline the process for candidates, it still depends on the platforms’ voluntary cooperation. MODA itself stresses that the underlying legal procedures are unchanged, meaning that even if the CEC would fine platforms for noncompliance on takedown notices, formal requests by the candidates to overseas headquarters is still required.

Opposition legislators now allege that MODA’s reporting site turns front line officers into “political censors”, even though the KMT supported the law when it was first enacted. The government insists it is merely digitizing an existing process. Both sides have a point — the real problem is that the process was flawed from the start.

Taiwan’s lessons for global policymakers

The KMT benzopyrene video shows what these gaps look like in real time. The president is not a candidate in November’s local elections, and the video appeared outside of any statutory window, so the deepfake law offered no route at all: neither for a forensic analysis, nor for a request for a takedown from YouTube. What remained were general criminal law and a visit by the police. That is the worst possible outcome. The deepfake statute cannot touch a deepfaked voice of the head of state, while police action is triggered by a politically partisan video. This is precisely the type of informal, unreviewable enforcement that may chill expression without giving actual victims of deception a remedy.

Taiwan’s experience carries three lessons for lawmakers who are drafting, or defending, deepfake election laws everywhere. A removal mechanism needs judicial oversight at the decision point to give it legitimacy, to allow for appeals, and to keep the executive branch from turning it into a tool for suppressing speech. A narrow election-period time window is mismatched to a threat that runs year-round, especially when it originates from abroad. And any takedown regime is only as strong as the state’s ability to reach the platforms: smaller democracies without significant market leverage should consider requiring foreign platforms to designate a domestic legal representative, as the EU’s Digital Services Act does — and as Taiwan’s own 2024 anti-fraud law already requires of large online advertising platforms, with gradual but notable results.

The ideal free speech protection we can strive for remains the one that US Justice Louis Brandeis set out nearly a century ago in 1927: “If there be time to expose through discussion the falsehood and fallacies, to avert the evil by the processes of education, the remedy to be applied is more speech, not enforced silence.” The court that struck down California’s law relied on exactly that logic. But the condition in Brandeis’s sentence is time, and a cloned voice released the day before a rally, or the week before a vote, is designed to leave none.

A deepfake election law worthy of a democracy should neither rely entirely on ‘more speech’ nor chill free expression. Instead, it must offer a remedy that is fast and comprehensive enough to remain effective beyond elections, yet narrow enough to spare satire, all while remaining subject to judicial oversight.