On September 20, 2026, Taiwan's Ministry of Justice said it would crack down on election deepfakes ahead of the year-end local elections. According to the Taipei Times, people who spread deepfake audio or video targeting candidates face up to seven years in prison, and platform operators that ignore removal orders also face penalties. Authorities plan more forensic tracing, faster takedowns, and coordination with news outlets, digital platforms and messaging apps.
The announcement puts two different models of platform duty side by side. One is already in force and is narrow. The other is pending in the legislature and is broad. Taiwan should keep them apart.
The case for the crackdown
The strongest argument for the government is that election deepfakes are a harm that arrives faster than any correction. The Taipei Times reports that candidates and party leaders from both the ruling and opposition camps have been targeted, including a fabricated clip of DPP candidate Chen Su-yueh and a 2023 fake audio recording of Taiwan People's Party founder Ko Wen-je. A debunking that lands after election day does not repair a lost vote. If a fake can circulate for weeks, the only remedy that works is removal within days, and only platforms can do that. Taiwan is also a top target for foreign influence operations, which raises the stakes further.
That argument has real force. It is also why the existing rule is reasonable.
Why the existing rule works
The deepfake provisions sit in the Public Officials Election and Recall Act, which the announcement refers to as the Civil Servants Election and Recall Act. Article 51-3 defines a deepfake as a digital composite that convincingly shows speech or actions that are not the real person's. Once a report is verified, broadcasters must stop airing the voice and likeness, and internet services must restrict access and remove it, within two days. Platforms must keep electronic records for six months from the removal request.
Four features make this a defensible design:
- A defined harm. The target is a synthetic imitation of a real person's voice or face, not a category like "false information".
- A verification step. Removal follows a verified report, not a platform's own guess.
- A short window. The duty applies only from the public notice of an election, or the declaration of a recall campaign, until the day before polling.
- A concrete deadline. Two days gives platforms a standard to meet.
A duty of this kind can be audited. A platform knows what it must remove, how quickly, and for how long. That is what proportionate regulation looks like.
The law still carries costs. A seven-year maximum sentence is heavy for speech offences, and satire or parody can look like a deepfake to an overzealous investigator. The Taipei Times editorial page noted this tension in August. When police questioned the creator of an AI video imitating President Lai's voice, the country argued over whether that was legitimate enforcement or "state intimidation". Taiwan's law would be stronger with an explicit exemption for clearly labelled parody and satire, and with published takedown and error statistics after each election.
The contrast: draft Article 4-1
The stalled National Security Act amendment is a different kind of instrument. Per the Taipei Times, the Cabinet's draft would let the Ministry of the Interior order internet access providers to halt domain resolution or restrict access. The triggers are war advocacy, false information harmful to national security, or Chinese Communist Party political propaganda, where circumstances are serious or there is immediate danger. Providers that fail to comply face fines of up to NT$1 million, with repeated fines for continued violations. A separate Taipei Times report says the amendments passed the Executive Yuan in December and await review by the Legislative Yuan. A government official there said the opposition has blocked national security changes 850 times.
Officials have offered reassurances. One minister said private family conversations would not be an offense, and the government says it will not patrol the internet constantly. Those are statements of intent, not statutory limits.
The difference from Article 51-3 is structural:
- The harm is a category of opinion. "False information harmful to national security" and "propaganda" are contestable judgments, unlike a fabricated clip of a named candidate.
- The decider is a ministry. The order comes from the executive, with no verification step or independent check in the reporting text.
- There is no sunset. The election rule expires at the vote. The security power is permanent.
- The tool is blunt. Halting domain resolution blocks an entire domain, not a single piece of content.
The current text of the National Security Act contains no such takedown mechanism, so this would be a new executive power over online speech. A government of any party could use it against critics.
What Taiwan should do
The deepfake crackdown suggests a workable template. A takedown duty is easiest to defend when it has a precise definition, a verification step, a fixed deadline, a limited duration and a way to measure results. Taiwan can use the coming election to show that the model works, and can publish removal counts and reversal rates as evidence.
If legislators revive Article 4-1, they should require prior judicial or independent review for all but genuinely urgent cases. They should also replace vague categories with defined offenses, limit orders to specific content rather than domains, and add a sunset and reporting requirements. Without those changes, the fact that the election rule works is a poor argument for the security rule.
A democracy facing sustained information warfare needs tools. It also needs tools that a future, less friendly government cannot easily turn on its own citizens.