China's AI election influence in Taiwan: A report
The debate around AI-generated content and its influence on public opinion is gaining importance. Concrete use cases of AI in information operations are particularly evident in the context of elections, such as in Taiwan. The line between authentic and synthetic content is increasingly blurred, requiring a critical examination of the current reporting status.
Source: Own Representation
Introduction
The question of whether a clip or post is real or "synthetic" is becoming increasingly relevant. This hesitation is part of the problem. In the context of Taiwan, evidence is mounting that AI is not just being used for memes, but as a building block in information operations targeting election environments. This happens sometimes crudely, sometimes precisely. Microsoft describes this.
Taiwan served as a testing ground for influence actors with its presidential and parliamentary elections on January 13, 2024. This was observed. . The pressure on Taiwan is justified by Beijing's claim to the country, which creates a motive to control narratives, undermine trust, and exploit domestic fault lines. A report from American Progress illustrates this. Taiwan's security agencies report a high pace and volume of disinformation disseminated through major platforms and video formats. AP News documented this.
Reporting Status 2024
An example of AI in the election context comes from Microsoft: on election day in Taiwan, the campaign, identified by Microsoft as Storm-1376, allegedly spread AI-generated audio clips. These put words into Terry Gou's mouth about an election recommendation that he did not make. Microsoft Security Insider reported on this. Microsoft classifies Storm-1376 as part of the pro-Chinese Spamouflage/Dragonbridge activity. The use of generated memes, video formats, and "news"-like assets was described. Microsoft Blogs provided further details.
Google reaches a similar conclusion: DRAGONBRIDGE (Spamouflage Dragon) is the "prolific" IO actor that Google's Threat Analysis Group (TAG) is tracking. It is characterized by high content volume, but often low actual reach. Google TAG Blog confirmed this. Fact-checkers on the ground also documented concrete deepfake and "cheapfake" cases in the run-up to the election. Manipulated videos were prepared in typical Taiwanese narrative styles to appear credible. Das Taiwan FactCheck Center showed examples. A later analysis by the Taiwan FactCheck Center described how AI videos circulated during the election campaign and why "seeing" is no longer sufficient proof. Further information on this.
Source: welt.de
China's geopolitical interests in Taiwan are a central driver for influence attempts, including in the digital space.
GoLaxy Documents (2025)
On December 28, 2025, the Taipei Times reported, citing the Yomiuri Shimbun, about internal documents of a Chinese AI company named GoLaxy. Documents allegedly suggest that AI-powered propaganda and "public opinion" tools were used in Hong Kong and Taiwan. Taiwan's local elections in 2026 and the presidential election in 2028 are seen as target areas. The article describes functions that go beyond traditional bot farms: network monitoring, identification of opinion leaders, psychological and linguistic profiling, and even "fictional characters" with appropriate dialect features who intervene in discussions.
Verifiability is important: The Vanderbilt Institute of National Security announced in 2025 that it would make nearly 400 pages of primary documents on GoLaxy publicly accessible. This is intended to serve as an archive containing evidence of AI-driven propaganda and information operations. The accompanying project "The GoLaxy Documents" describes the goal and content of the archive in a similarly clear manner: data collection, "precision profiles," and scaled AI propaganda. Further information on this.
AI Tools and Functionality
When one hears "election influence," one often thinks of a single viral deepfake. Current reports show more of a bundle that reinforces itself: AI lowers production costs for variants, accelerates message testing, and makes micro-adjustments easier. A report by CSET Georgetown. Microsoft Blogs provided examples.
Google TAG emphasizes the cross-platform pattern in DRAGONBRIDGE/Spamouflage: high output, many identities, low actual resonance – but with a learning curve and adaptation as soon as platforms interfere or delete. Google TAG Blog reported. OpenAI has been documenting since 2024 that influence actors are also misusing AI systems as a "content engine": for drafting posts, comments, articles, persona texts, and campaign logic – often in combination with human control and distribution outside the AI platform itself. OpenAI Threat Intelligence Reports provided insights.
The recurring pattern rarely begins with a deepfake, but with target group knowledge. In the GoLaxy descriptions, precisely this "front" in the process is the focus: collecting data, finding influence nodes, deriving language and value profiles, and creating "suitable" figures or voices that feel like local users from them. Die Taipei Times described this. Then comes the content in variants. A clip is not created once, but twenty times – once angry, once ironic, once "concerned." Generative AI scales here, as small changes are cheap and A/B tests don't have to feel like a campaign. CSET Georgetown analyzed this.
Only then follows distribution: coordinated networks, fake personas, comment swarms, reuploads, platform hopping. Meta describes such operations as "coordinated inauthentic behavior" (CIB) and documents that accounts often also use likely AI-generated profile pictures to create mass and "local color." Meta Adversarial Threat Report demonstrated this. The key is the feedback loop: measuring reactions, "covering up" contradictions, pushing new claims. Taiwanese fact-checkers describe this dynamic from the opposite perspective: it is not the individual fake that is dangerous, but the continuous shift in what people consider plausible. Die Thomson Foundation reported.
Source: deutschlandfunk.de
The digitalization of electoral processes harbors new attack surfaces for manipulation by AI.
Detection and Defense
After the 2024 election, Taiwan experienced how quickly "election fraud" rumors can gain traction – and how much of it is spread through short videos, screenshots, and messenger forwards. AP News reconstructed this. The counter-response often came not from a single authority, but as a "whole of society" reaction: election commission, ministries, fact-checkers, creators, community groups worked on speed and transparency. Civil society tools play a concrete role in this. Cofacts is known in Taiwan as a collaborative fact-checking project that also works via bots and workflows in closed messengers. OCF Taiwan described Cofacts.
In parallel, reporting from the Taiwan FactCheck Center provides examples of what deepfakes look like in practice – including typical warning signs that can be recognized not only "technically" but also narratively: implausible source chains, missing original footage, speech breaks, reuploads without context. Das Taiwan FactCheck Center provided insights. Technical provenance standards can support this process if they are widely used. The C2PA specification ("Content Credentials") aims to digitally bind origin and editing steps so that editorial teams, platforms, and users can classify fakes more quickly. Die C2PA-Spezifikation is relevant here.
Source: cnbc.com
Chinese AI technologies such as 'deepseek' could be used for targeted disinformation campaigns.
Outlook
Taiwan's Central Election Commission has set the election day for the local elections in 2026 for November 28, 2026. OCAC Taiwan confirmed this. The Taipei Times quoted on December 28, 2025, explicitly the expectation that "public opinion warfare" could be intensified in the run-up to the upcoming local elections – also as preliminary logic for 2028. Institutionally, this fits into the calendar: Taiwan's president and vice president are directly elected and have regular four-year terms. Taiwan.gov.tw provided information on this.
Beyond Taiwan, the mechanism remains the same: Microsoft already warned in 2024 that state-sponsored actors could use AI to disrupt election contexts in other countries as well – explicitly referring to "learning" from Taiwan. Microsoft Blogs issued this warning.
The documented development is less about "the one deepfake that tips everything over" and more about the normalization of an environment where identities, voices, and video evidence become cheaply imitable. Microsoft's documented audio case on Taiwan's election day in 2024 shows the tactical side. Microsoft Security Insider reported. The GoLaxy documentation, as described by Vanderbilt and picked up by Taiwanese media, points to a strategic next stage: profiling, personalized role characters, and dialog-capable "personas" as infrastructure – not as individual fakes. Vanderbilt University provided details. Taiwan's counter-model appears down-to-earth: quick clarifications, strong fact-checking networks, public learning processes – and as little mystique around technology as possible. AP News highlighted this.