How Deepfakes Spread in Taiwan's 2026 Local Elections: Overseas Accounts, Bot Amplification and AI Labels
Ahead of Taiwan's November 28, 2026 local elections, a legislator said several deepfake videos had targeted a Taipei mayoral candidate, and the Minister of Digital Affairs said many such deepfakes came from overseas accounts and were liked and shared by bot accounts. The ministry also asked major platforms to scan for AI-generated material that is not properly labeled. This page sets out what was reported, how this kind of amplification typically works, what labeling and takedown rules can and cannot do, and how coordinated amplification can be detected in public data.
Definition: coordinated deepfake amplification
Coordinated deepfake amplification is the organized promotion of synthetic or manipulated video or audio that falsely depicts a real person saying or doing something, using multiple accounts acting together, often with automated likes and shares, to increase its visibility.
The definition separates two problems that are often merged. The first is the content: a fabricated clip that shows someone saying or doing something they did not. The second is the distribution: accounts acting together to make that clip more visible. A deepfake with no amplification may reach few people, and an authentic video can also be amplified inauthentically. Detection and response differ for each, so it helps to name them separately.
Taiwan's election law defines the content side. Article 51-3 of the Public Officials Election and Recall Act covers a person's voice or likeness produced with digital synthesis or other technical means in a way that could lead others to mistake it for words or actions the person actually said or did.
What was reported in November 2026
On November 10, 2026, the Taipei Times reported the following from a Central Election Commission (CEC) announcement and a session of the Legislative Yuan's Transportation Committee the previous day. Each item is listed with the person or body it was attributed to; this page adds no findings of its own about the case.
| Topic | What was reported | Attributed to |
|---|---|---|
| Public warning | The CEC released a TV commercial warning about AI-generated content intended to influence the November 28 elections, and urged people to verify election information before sharing it. | Central Election Commission |
| Deepfakes targeting a candidate | Several deepfake videos made false allegations against a Taipei mayoral candidate. | A legislator from the candidate's party |
| Takedown time | The legislator said the campaign informed the Ministry of Digital Affairs (MODA) on October 30 and the videos were not taken down until November 5 because of a criminal investigation; although the ministry could get deepfake videos removed in 73 minutes, the falsified material had circulated for seven days. A November 5 ETtoday report quoting the campaign gives a more detailed account: a Threads post reported on October 30 was removed the next day, and a separate Facebook Reels video filed through MODA's reporting site on November 5 was removed 73 minutes after submission. | The same legislator; the campaign, as quoted by ETtoday |
| Overseas accounts and bots | Many such deepfake videos came from overseas accounts and were liked and shared by bot accounts. | Minister of Digital Affairs Lin Yi-ching |
| AI labeling | Meta requires users to label AI-generated content. MODA asked Meta, Google and LINE to regularly scan their platforms for AI-generated material that is not properly labeled, and said users who fail to label AI content would face legal consequences. | Ministry of Digital Affairs, as reported |
Deepfakes are not a one-party problem in Taiwan. A Tech Policy Press analysis (October 1, 2026) describes earlier incidents involving politicians from different parties, including an audio recording that circulated in August 2023 and a video released in July 2026 that used a voice resembling a national leader's. The methods below apply regardless of who is targeted.
How a deepfake can move through the feed: five possible stages
The report describes overseas accounts and bot engagement but does not establish the full distribution sequence of the November 2026 videos. The five stages below are an illustrative model for planning monitoring. They are possible patterns with clues analysts can check, not necessary features of every deepfake and not proof of coordination.
| Stage | What may happen | Clues analysts can check in public data |
|---|---|---|
| 1. Fabrication | A short synthetic or manipulated clip is produced, often paired with a caption that makes one specific allegation. | Whether reliable source footage can be found; whether the caption makes a claim the footage itself does not show. |
| 2. Seeding | A small number of accounts post it first. Some may be new, operated from outside the country, or have little history in local politics. | The posting histories of the earliest accounts visible in the collected data. |
| 3. Amplification | Automated or coordinated accounts like, share and comment within short windows, which can make the clip look popular and help it surface in ranked feeds. | Engagement arriving in bursts; many accounts reacting within minutes with similar or identical text. |
| 4. Migration | The clip is re-uploaded, screen-recorded, trimmed or forwarded to other platforms and messaging groups. | New uploads of the same clip on other platforms, which may no longer carry the original's label or metadata. |
| 5. Persistence | Copies may continue circulating while a report is reviewed or after the original is removed. | Discussion that cites copies or screenshots of a removed post. |
Stage 4 is why the time between first post and removal matters: copies on additional platforms may require separate review and removal requests.
Why labels and takedowns leave a gap
Labels depend on disclosure or detectable signals. In February 2024, Meta announced that it would require people to use its disclosure tool when posting organic content with photorealistic video or realistic-sounding audio that was digitally created or altered, and that it may apply penalties if they fail to do so. At the time, Meta said AI tools that generate audio and video had not yet started including labeling signals at the same scale as image generators, so it could not yet detect those signals and relied on uploader disclosure. Its current help page explains how detected signals and uploader disclosures feed its AI content labels. Disclosure can be omitted, and detectable markers are not always present, so labeling cannot be assumed to catch every synthetic clip. MODA's request that platforms scan for unlabeled AI material targets that gap.
The legal takedown route requires a forensic examination first. Under Article 51-3, from the election announcement through the day before voting, candidates and prospective candidates may ask the police to examine a deepfake depicting themselves. If it is confirmed, they send the platform or broadcaster a written request with the forensic report and notify the election commission; the provider then has two days to stop broadcasting, or to restrict access to and remove the content, and must preserve the related records. MODA's election AI deepfake reporting site relays such reports to platforms; the official guidance names Meta (Facebook, Instagram, Threads), Google (YouTube), LINE and TikTok. Reports must be submitted online by the candidate or prospective candidate themselves, verified with their own Citizen Digital Certificate, and include a police forensic report. Separately, Article 104 makes it a criminal offense to spread falsehoods through a deepfake of a candidate's voice or likeness with intent to make a candidate win or lose, where this is capable of harming the public or others.
Because each step has to be completed in turn, removal can take time. Accounts of the reported case differ in detail: the legislator said the falsified material had circulated for seven days, while the campaign described individual posts removed the day after reporting or, through the reporting site, 73 minutes after filing. While a clip remains online, monitoring public data can surface leads about suspected amplification and show an event timeline within the collected data. It does not establish who operates the accounts, where they are based, or the full extent of circulation.
How InfoMiner detects coordinated amplification
InfoMiner, LargitData's social listening and risk intelligence platform, collects publicly posted content with web crawlers across more than 500,000 channels. Platforms that can be covered include PTT, Dcard, Facebook, YouTube, Threads, TikTok, LINE communities, X and Instagram. Actual coverage depends on the plan, gaps can occur when platforms change or block access, and private messages and non-public groups are out of scope.
On top of that collection, InfoMiner builds event timelines and runs real-time intelligence assessment, returning suspicious accounts, account clusters and leads for analysts to verify. Its video and deepfake detection adds suspicion signals for the clip itself. A suspected deepfake push is assessed across the signal layers below. No single layer is decisive; a lead is stronger when several layers point the same way.
| Signal layer | What is assessed | What it cannot prove on its own |
|---|---|---|
| Copies of the same item | Posts that appear to reuse the same clip or repeat the same allegation, reviewed together so one clip can be followed as one story; analysts confirm whether the items are actually related. | Different versions may have been made independently by different people. |
| Event timeline | The earliest copy observable in the collected data, and how the clip spread across platforms over time. | The earliest observed copy is not necessarily the true origin. |
| Timing and synchronization | Bursts in which many accounts post, share or comment on the same item within a narrow window, often with near-identical wording. | Real supporters can also respond together, for example to a call to share. |
| Account attributes | What public profiles show about the accounts involved, such as account age, posting history and whether they have engaged with local topics before. | A new or unusual account is not proof of automation, of where it is run, or of who runs it. |
| Interaction network | Who shares and replies to whom, to find tight clusters that mainly amplify each other. | Close-knit communities of real users can form similar clusters. |
| Video and deepfake signals | Suspicion signals from InfoMiner's video and deepfake detection for the clip itself. | A suspicion signal is not a forensic finding, and even a forensic examination can be inconclusive. |
The output is a set of leads, not a verdict. InfoMiner looks for behavioral signs of coordination; it does not identify who operates an account and does not match identities across platforms. Whether content is false is decided by fact-checkers on evidence, and whether it is unlawful is decided by the competent authorities. Criticism and ordinary political speech do not, by themselves, indicate fabrication or coordinated manipulation, and are not monitoring targets.
Reading the signals carefully
- "Overseas" describes an account, not a sponsor. An account operated from abroad, or bots liking a post, does not by itself establish who commissioned the campaign. Attribution needs evidence beyond public behavior.
- Bot engagement does not show who arranged it. Automated likes on a post do not reveal whether the post's subject, their opponents or an unrelated party paid for them.
- The doubt runs both ways. A fabricated video can be taken as real, and a real video can be dismissed as fake. Both claims need to be checked against original material.
- Engagement is not proof of exposure or persuasion. Likes and shares alone do not establish how many people saw or believed a clip.
Related reading
Frequently asked questions
Sources
- Shelley Shan, "Election body airs deepfake warning," Taipei Times, November 10, 2026.
- Ministry of Justice, Public Officials Election and Recall Act, Articles 51-3 and 104 (Chinese text), Laws & Regulations Database of the Republic of China (Taiwan).
- ETtoday, report quoting the campaign office on deepfake takedowns (in Chinese), November 5, 2026.
- Ministry of Digital Affairs, Election AI Deepfake Reporting Site (in Chinese); Taipei City Police Department, guidance on the reporting site (in Chinese).
- Eddy Yen-Ting Lin, "Taiwan's Election Needs a Deepfake Law That Works," Tech Policy Press, October 1, 2026.
- Meta, "Labeling AI-Generated Images on Facebook, Instagram and Threads," February 6, 2024; and "How to identify AI content on Meta products."
Monitoring public information during an election?
LargitData can walk you through InfoMiner's capabilities, limits and customization options to see whether it fits your needs.
Contact us