[Big read] Satire, deepfakes and virtual voters: AI enters Taiwan’s election race

(Edited and refined by Vianne Chia, with the assistance of AI translation.)

The creative team behind satirical fictional character Principal Lyle, which went viral during last year’s mass recall campaign in Taiwan, has released a new political satire series, titled Cat Bro’s Troubles. One episode takes aim at Puma Shen, the Democratic Progressive Party (DPP)’s Taipei mayoral candidate who has generated a political “Puma wave” (洋流) ahead of Taiwan’s local elections this year.

In the roughly one-minute AI-generated skit, the main character, Cat Bro, lounges on a sofa and coughs twice before saying, “My throat needs some medicine.”

A slightly overweight man with curly hair then walks into the living room holding a syringe. “I’ve got drugs too. I really can’t get through a day without them,” he says.

The character Dr Sore Tumour is a caricature of Shen. The Chinese term for “sore tumour” sounds like “Puma wave”, a reference to Shen’s political momentum. The character uses the English word “drug”, which can refer to both medicine and narcotics — a jab at Shen’s controversial past remarks on drug policy.

Satire at a tenth of the cost

The series was produced by Fengxiang Co. Its director, Wei Chun-yu, a former head of the Kuomintang (KMT)’s new media department, told Lianhe Zaobao (LHZB) that the team had produced the series on its own initiative to show support for incumbent Taipei mayor Chiang Wan-an, who is seeking re-election.

The team’s earlier production, Principal Lyle, featured a fictional school principal that alluded to Taiwan President Lai Ching-te. “Lyle” sounds like the English word “liar”. The video satirised the DPP government and the mass recall campaign, racking up more than 400,000 views within four days of its release. While some scenes in Principal Lyle still had to be drawn by hand, Cat Bro’s Troubles was made entirely with artificial intelligence (AI). Wei said producing an animation of comparable quality without AI would cost ten times as much.

As the cost of production falls and the technology becomes more accessible, AI is also finding its way into political parties’ battles for public opinion. In August, the KMT introduced an AI-generated spokesperson, Zheng Xiaowen, who comments in short videos on issues including cash handouts, defence spending and food safety.

The trend extends beyond Taiwan. In the run-up to India’s 2024 general election, the ruling party used AI to dub Prime Minister Narendra Modi’s speeches into eight regional languages and circulated them on social media to reach more voters.

Taiwan’s first AI election?

Chuang Po-chung, a professor in the Department of Journalism at Chinese Culture University in Taiwan, considers 2026 the “first year of AI election campaigning”. Generative AI had already appeared during Taiwan’s 2024 presidential election, he said, but the technology was less mature then. It is now being used across a wider range of campaign activities.

He noted that political communication has evolved from newspapers and television to social media and AI. “Although technology changes in countless ways, the goal remains the same: to win votes,” he said.

Data released on 15 October by the Wesleyan Media Project in the US showed that at least 220 political advertisements in this year’s US midterm elections had been generated partly using AI, with more than US$137 million spent on them. By comparison, spending on such advertisements during the 2024 election cycle was less than US$50,000.

AI allows campaign teams to produce multiple versions of their materials faster and more cheaply. It can also help them monitor online sentiment and analyse voter responses. But Chuang said professional judgement remains essential when interpreting complex public sentiment and developing campaign strategies.

The making of Principal Lyle offers a glimpse into how its creators assessed the political climate. Wei Chun-yu recalled that, ahead of the first round of recall votes last year, the team released Flipping the Mahjong Table, a live-action video that mocked the recall campaign through a story about a sore loser who, as the title says, flips a mahjong table.

After the first round of voting, Wei worried that supporters might have vented their frustration and grown complacent before the second round a month later, depressing turnout. The team therefore wrapped the nuclear power referendum into a lighter, animated format. The result was Principal Lyle.

Wei believes Principal Lyle had some impact before the second round of recall votes. He said it disrupted the opposing camp’s rhythm, helped rally the pan-Blue (KMT-led) and White (Taiwan People’s Party) camps, and encouraged their supporters to vote.

Unproven impact of AI campaigning

Yet the issue of whether AI-generated political content can persuade voters, let alone change their voting decisions, remains an open question.

Adrian Rauchfleisch, a distinguished professor at National Taiwan University’s Graduate Institute of Journalism, said AI is advancing rapidly, but for now it mainly serves to support and reinforce existing forms of political communication. New technologies take time to become widely adopted, and their real impact may not become apparent for years.

Can AI genuinely persuade voters? Rauchfleisch said several experiments have found that AI chatbots can be even more persuasive than humans. In a study in Britain that he co-led, simply telling participants that they were talking to an AI did not significantly weaken its persuasive effect. But when participants were told that the conversation was intended to change their views, the effect was halved.

He cautioned that experimental findings may not translate directly into real-world settings. Voters may be unwilling to engage with campaigns they disagree with.

“The main bottleneck is, in a real-world setting or in a campaign setting, whether people really start interacting with such chatbots,” he said.

Research has also found that people still tend to prefer talking to humans. Campaign teams that use chatbots to try to change voters’ political positions could therefore provoke a backlash.

Rauchfleisch also noted that election outcomes are shaped by many factors, including incumbency advantages, media exposure and voter mobilisation. Even if AI lowers the cost of campaigning, it does not necessarily follow that campaigns will reach more voters, let alone that AI’s effect on results can be proven.

Deepfakes threaten to erode trust in democratic societies

More concerning than AI’s ability to sway elections is the growing sophistication of AI-generated deepfakes. As people find it harder to distinguish real information from fabricated content, they may start to doubt what is real, eroding the foundations of trust in democratic societies.

With Taiwan’s local elections approaching at the end of the year, a string of AI deepfake controversies has emerged. 

In June, Puma Shen was targeted by a manipulated image showing him alongside a scantily clad woman and what appeared to be drugs. The IP address of the account suspected of creating and posting the image was traced to mainland China, but the identity of the actual creator remains under investigation.

In September, DPP Changhua county magistrate candidate Chen Su-yueh discovered a deepfake video circulating online that made her appear to say, “I could just lie back and still win.” In October, DPP Yunlin county magistrate candidate Liu Chien-kuo was targeted by an AI-generated video insinuating a romantic scandal.

The risks extend beyond fabricated audio and video. Bot accounts can also spread messages on a large scale, creating the illusion of widespread public support.

Lai You-hao, deputy director of the Democratic Governance Program at the Research Institute for Democracy, Society and Emerging Technology, a policy research think tank set up under Taiwan’s National Science and Technology Council, said bot accounts in the past often posted identical or similar comments, making them relatively easy to identify. Today, AI allows large numbers of accounts to mimic human interactions and promote the same arguments in different ways, creating the impression that a political view enjoys broad support.

AI has made the behaviour of these accounts more complex, he said, so it is now “harder to determine whether they come from the same coordinated manipulation network”.

Taiwan’s National Security Bureau recently reported that intelligence agencies had collected and documented more than 2.5 million controversial or false items of information linked to mainland China’s cognitive warfare against Taiwan in the first nine months of this year — about 48% more than a year earlier. The authorities said mainland China had used AI-generated audio and video, fake accounts and coordinated commenting to manipulate public opinion. They also said it had collected candidates’ policy positions and polling data in an attempt to influence the election. Beijing has previously denied interfering in Taiwan’s elections.

Lai believes such operations may not directly change how people vote. But as manipulated content floods the online space, it could become increasingly difficult for the public to find reliable information. In the long run, he worries that people may lose the willingness to base their judgements on trustworthy information — undermining a fundamental basis of democratic self-determination.

How much should we trust AI-generated results?

Every day at noon, 17,555 Taipei “voters” are asked who they would support if the election took place tomorrow — all without a single phone call or street interview.

These are virtual citizens created using AI by the Taiwan Public Opinion and Political Behavior Research Centre (TPOC), a big data analytics firm. Each has an age, occupation, place of residence and political leaning. They receive news and social media information at scheduled intervals and simulate how real voters’ attitudes might change.

TPOC’s data as of 22 October showed that incumbent Taipei mayor Chiang Wan-an had a narrow lead in the closely watched mayoral race, with 40.9% support against 39% for DPP candidate Puma Shen. A traditional poll released by the Xinmin Cross-Strait Research Association on 20 October, however, showed Shen well ahead, with 50.3% support compared with Chiang’s 33.3%.

TPOC founder Bubble Chen said traditional polling has become increasingly difficult in recent years, as rising scam calls make people wary of answering and fewer households keep landlines.

At the end of April this year, her team began using ChatGPT as the basis for its simulations, combining it with past election results, demographic data and online public opinion to create virtual voters with different personalities and model shifts in public sentiment.

For example, the model could simulate how voters with children might respond to news about child abuse and how that could shape their views on child protection policies.

Chen stressed that she did not consider these simulations to be opinion polls. “Calling it a poll just makes it easier for people to understand.”

AI models may have political leanings

Chen has found that AI itself may have political leanings. In one experiment, her team asked three large language models — ChatGPT, Gemini and Claude — to simulate voting without any corrective parameters. All three produced results favouring the pan-Blue camp.

The team still needs to calibrate the models against traditional polling data from different organisations, she said. “You can’t just leave everything entirely to AI. You still need external calibration.”

The US-based Pew Research Center has likewise shown that AI cannot yet replace rigorous traditional polling. The centre used data from real respondents to create AI “digital twins” and simulated their answers to nearly 300 survey questions. The results differed from those of real respondents by an average of 12.4 percentage points.

For example, the AI simulations showed that 46% of American adults approved of US President Donald Trump’s performance, while the actual approval rating among respondents was just 34%.

Taiwanese polling expert Tai Li-an said the scientific foundations of traditional polling rest on representative sampling, questionnaire design and survey methods. He questioned whether virtual voters whose backgrounds and values are set by researchers could produce reliable results. Generating tens of thousands of virtual respondents, he argued, was not the same as interviewing tens of thousands of real voters.

Human thinking, he said, does not work like a mechanical response system. Citing the shooting of Taiwan presidential candidate Chen Shui-bian on 19 March 2004, the day before the presidential election, he noted that unexpected events could rapidly affect voters’ emotions — something AI might not adequately capture.

Referring to the “Puma wave” surrounding Shen’s candidacy, Tai asked: “There’s a ‘Puma wave’ outside. How is it flowing inside the model? It has no emotional currents.”

He also questioned how much weight traditional polling data carried when used to calibrate AI models, and how much independent value the simulations offered as a reference.

Chen acknowledged that AI-based public opinion simulation was still at an exploratory stage. Taiwan’s local election results on 28 November, she said, will provide an opportunity to test its effectiveness. “If it succeeds, we gain something. If it fails, we learn from the experience.”

Satire or deepfake?

In July this year, a scandal erupted in Taiwan over cooking oil contaminated with the carcinogen benzo[a]pyrene. A political satire video titled Oil, You Can’t Afford It was narrated in the “voice” of the carcinogen, criticising the government’s handling of the scandal. But the narration bore a striking resemblance to President Lai Ching-te’s voice, drawing producer Wei Chun-yu into a controversy over AI deepfakes.

According to Wei, prosecutors have opened a preliminary inquiry (他字案) into the issue, to gather evidence and establish whether any offence might have been committed. The controversy has raised broader questions, such as where the line should be drawn between political satire and AI deepfakes, and how the law can deter malicious fabrication without compromising freedom of speech.

Wei believes that the issue should not be judged solely by how closely a voice or image resembles the person concerned. He argued that it is also necessary to consider whether the creator intended to mislead viewers and whether the work involved malicious fabrication.

He stressed that Oil, You Can’t Afford It merely continued the character-based approach used in Principal Lyle. According to Wei, the team had no intention of making viewers believe that the narration was Lai’s actual voice. He also questioned why similar voice impersonations had not drawn complaints in the past, but were now being challenged as potential deepfakes.

Experts: current laws are inadequate

Eddy Lin, a non-resident fellow at Cornell University’s Centre on Democracy who participated in discussions on amendments to Taiwan’s Public Officials Election and Recall Act in 2023, said the controversy exposed ambiguities in existing legislation. In particular, he said that there is no clear standard or threshold for determining whether content is “likely to mislead others into believing it is true”.

The relevant provisions also mainly cover specific groups, including candidates, officials facing recall and the initiators of recall cases. Ordinary members of the public may find themselves without the same protections.

Lin cited an incident during last year’s KMT chairmanship election, when former Taipei mayor and party chairmanship candidate Hau Lung-bin and Taipei City councillor Liu Tsai-wei were targeted by a fabricated kissing video. Because an intra-party election does not fall under the Election and Recall Act, those involved could not seek redress under it.

He argued that lawmakers should consider expanding the categories of people protected, while setting appropriate thresholds for distinguishing between public figures and ordinary citizens. After all, information manipulation in Taiwan is not confined to election periods, and the law should not focus solely on electoral activities.

Lin also suggested requiring election-related AI content to carry clear labels, making it easier to determine whether regulations had been violated. “Either it’s labelled, or it isn’t,” he said. Such a provision would avoid further disputes over whether the content was likely to mislead people into believing it was real.

Lai You-hao noted that Taiwan had spent nearly 40 years transitioning from authoritarian rule to democracy, leaving society particularly wary of government intervention in speech and content.

He suggested that Taiwan could look to the European Union’s Digital Services Act, which he said shifts the focus of regulation from the content of speech to manipulative behaviour. The law requires very large online platforms and search engines to assess and mitigate systemic risks, including the effects of fake accounts, bots and coordinated manipulation on public discourse and electoral processes.

Despite the legal uncertainty, Wei Chun-yu has no plans to stop his work. His team now labels every production as AI-generated, he said.

He also seems unfazed by the prospect of further legal issues. “After all, if I’m going to be reported once, I might as well be reported twice. Bring it on.”

This article was first published in Lianhe Zaobao as “从助选到介选 AI如何改变选战？”.

Related: When Taiwan’s stars must prove their political loyalty | Taiwan’s local elections: Can DPP disrupt KMT’s Taipei stronghold?