The opposition Kuomintang (KMT) candidate holds a 10.7-percentage-point lead in the race for mayor of Taichung, Taiwan's second-largest city by population, according to an artificial intelligence voter simulation released on October 21, ahead of the 2026 Taiwan local elections on November 28.
The study by TPOC, a Taiwanese issues research center, found that after undecided respondents were pressed on which way they lean, 46.2% of its AI-generated voters backed KMT lawmaker Chiang Chi-chen, also known as Johnny Chiang, while 35.5% supported Ho Hsin-chun, a legislator from the ruling Democratic Progressive Party (DPP). Chiang also led by 13.6 points when the virtual voters were asked which candidate was better suited to be mayor.
The results are not a conventional opinion poll. TPOC built 18,971 AI "virtual voters" modeled on the demographics of Taichung's 29 districts and had all of them answer the survey, rather than interviewing real residents.
Ho Leads Only Among Voters In Their 40s
Of the six age groups in the simulation, Ho came out ahead in just one. Among virtual voters aged 40 to 49, she drew 44.5% to Chiang's 40.9% and led him by 6.8 points on favorability.
Chiang's strongest group was voters aged 50 to 59, where he led by 26.3 points. He was also ahead by 15.4 points among those aged 20 to 29, though 25.2% of that group remained undecided.
By gender, Chiang led by 14.4 points among men and 7.3 points among women. Women were more likely to be undecided, at 18% compared with 10.9% for men.
Chiang Dominates Xitun And Nantun Districts
Chiang's widest margins came in Xitun and Nantun, two of Taichung's fast-growing central districts, where he led by 32.1 and 30.8 points respectively. He was also 22.9 points ahead in Fengyuan.
Ho led by 13.9 points in the Central District, 13.6 points in Dali, and 11.3 points in Taiping, and was ahead in the coastal districts of Qingshui and Wuqi. The East District, South District, and Dajia were all within 2 points.
Among virtual voters set up as supporters of the Taiwan People's Party (TPP), the smaller opposition party, Chiang took 41.1% and Ho 38.3%, while 12.7% backed neither. Ho edged Chiang among politically neutral virtual voters, 27.7% to 26.7%, but 42.4% of that group was still undecided. TPOC said Ho's room to grow lies mainly in extending her lead among voters in their 40s and winning over undecided women and centrists.
Traditional Polls Show A Wider Range
The AI simulation falls between the results of recent telephone polls. A survey commissioned by the KMT caucus in the Taichung City Council and conducted by the TVBS Poll Center from September 7 to 11 put Chiang at 45.1% and Ho at 32.5%, a 12.6-point gap, with a margin of error of plus or minus 3 percentage points.
A poll released on October 12 by a pro-DPP civic group, the Taichung chapter of the Friends of Lai Ching-te Association, found a much tighter race. Conducted by Evidence-Based Polling from October 5 to 6, it showed Chiang at 34.3% and Ho at 32.4%, a gap of 1.9 points, with 32.2% of respondents undecided. That poll surveyed 1,071 adults and had a margin of error of 2.99 percentage points.
Taichung is currently governed by KMT Mayor Lu Shiow-yen, who is serving her second and final term, making the open seat one of the most closely watched contests of the 2026 Taiwan local elections.
How The AI Voters Were Built
TPOC said its system calibrated age, gender, and education level to the population of each of Taichung's 29 districts, then added prior probabilities of party identification based on past party-list voting patterns. The project was led by TPOC founder Chen Pi-shu and funded by Influence Data Co., with the model run on October 19, 2026.
The study used a technique called "silicon sampling" and ran on OpenAI's GPT-5.4 mini model, March 17, 2026 version. TPOC said it applied post-processing to correct the model's partisan bias, and that the percentages were repeatedly weighted using weights ranging from 0.14 to 2.85. Because the exercise simulates an entire virtual population rather than drawing a probability sample, the concept of sampling error used in traditional polls does not apply.
TPOC acknowledged clear limits. The virtual voters' views come from the model's estimates of the population and public sentiment, which may amplify arguments that get heavy media attention and understate the views of undecided or politically disengaged voters. The center said the results reflect the leanings of a virtual population under its model settings, are meant as a reference for assessing the race, and do not represent how real voters will cast their ballots.