China’s Disinformation Operations in Taiwan Are Changing with AI

An undeleted line from an AI prompt was found in a post on a Chinese content farm. The prompt instructed writers to target a Taiwanese audience, rewrite the article in Traditional Chinese while preserving historical accuracy, and limit it to 500 words. The line was deleted within two minutes of publication but remained in the post’s editing history, where it was identified by the research team led by Dr. Austin H. Wang, a political scientist at the University of Nevada, Las Vegas, and RAND. The incident has since been logged in the OECD's AI Incidents and Hazards Monitor and reported by multiple Taiwanese news outlets.

This evidence points specifically to an effort to produce content for a Taiwanese audience. It provides stronger evidence of Taiwan-specific targeting than of an operation aimed at Traditional Chinese-language readers in general.

This kind of coordination is already documented. Meta published a report this March, documenting its disruption of a network originating in China that targeted audiences in Taiwan. The cluster accounts used Taiwan-based proxy IPs to appear to be of Taiwanese origin, engaging in the promotion of pro-Beijing narratives and criticizing the Taiwanese ruling party across several Facebook pages. The network disguised itself as legitimate advertisers by paying the $15,000 ad fee in Hong Kong Dollars, Chinese Yuan, and Taiwan New Dollars.

While this particular network appears to have relied on manual coordination and proxy IPs, other PRC-linked operations have incorporated language models in their campaigns. Powered by generative AI, these campaigns used large language models for content creation and strategic planning, as documented in OpenAI's February and June reports on covert ChatGPT-assisted campaigns targeting Japanese Prime Minister Sanae Takaichi and US policy debates. Together, these reports show that AI-assisted operations are cross-border efforts.

Apart from coordinating account activities, AI is also being used to create misleading video content. The Australian Strategic Policy Institute found an information campaign during the Taiwan election in 2024, in which the CCP's largest network of inauthentic social media accounts, known as Spamouflage or Dragonbridge, used AI-generated news anchors to spread and amplify content from what's called "The Secret History of (outgoing president) Tsai Ing-wen.” The incident was also documented in a Google report. The analysis shows the campaign "had limited reach, with practically no engagement from organic users." William Yang, a Northeast Asia analyst at the Belgium-based International Crisis Group, said the election interference efforts from China will certainly be on the rise, with the assistance of advanced AI technology. This raises a question: if these campaigns rarely convince anyone, why do they persist?

Part of the answer lies in evidence suggesting China’s goal is to distract the public and change the subject. Even when people are aware that content can be fabricated, as fake material becomes easier to produce, they may grow increasingly unwilling to believe anything at all. This is sometimes called the “liar’s dividend”— the ability to dismiss genuine evidence as fake simply because fake evidence is now easy to produce. Research has documented how generative AI can make this problem worse.

Currently, most reports released by these AI companies or social media platforms focus on identifying which actors did what and blocking them, without examining collaborative behavior patterns to analyze how the operations work and what intents are involved. AI companies and social media platforms should not only censor harmful speech, but also examine how these accounts operate together.

For example, coordinated actions that use narratives targeting specific country policies within a short period reveal transnational interference intent. Meta has mentioned in its report that their focus on behavior rather than content has appeared effective, a conclusion OpenAI's own threat reports have echoed. However, neither has this approach become an established standard among these companies, nor has it brought enough transparency.

By far, the European Union has the Digital Services Act (DSA) to combat disinformation by legally requiring Very Large Online Platforms (VLOPs) to assess and mitigate risks to public discourse. In the US, the proposed Platform Accountability and Transparency Act (PATA) would require large social media platforms to open their data to independent researchers and the public.

Taiwan has not been inactive. In 2022, the National Communications Commission drafted a bill to regulate platforms, but it never reached the legislature amid concerns over freedom of speech. Ahead of Taiwan’s 2026 year-end elections, the District Prosecutors’ Office has established a special task force to combat deepfake content, while the Ministry of Digital Affairs plans to launch a reporting network in August.

However, the focus of these mechanisms remains on identifying the fake content itself, and has not yet addressed how that content is distributed, amplified, and coordinated. Taiwan has taken part in a GCTF (Global Cooperation and Training Framework) workshop with the US and other partners this July, addressing the misuse of AI, deepfakes, and disinformation. The next step should be to bring the distribution and coordination models outlined here into that existing dialogue.

These efforts matter, but none of them will be enough unless they can match the scale of the threat. The threat lies not in the fabricated content itself, but in the scale, speed, and coordination of the operation behind it.