SIAgentNews.com
← All news

Asked about Taiwan's mayoral candidates, ChatGPT and Gemini paint different pictures

A REALI study of 12 Taiwanese politicians found the two chatbots describe the same person differently and lean on different sources.

· First reported by technews.tw

7,472source citations REALI tracked across 779 valid answers

The short version

  • REALI asked each chatbot 15 times about each of 12 candidates in Taiwan's six biggest cities.
  • ChatGPT matched Chiang Wan-an to Doraemon and Shen Po-yang to Luffy from One Piece.
  • ChatGPT used 'record in office' for Chiang in 12 of 15 answers; Gemini never did.
  • News outlets were about 50% of ChatGPT's cited sources and 75% of Gemini's.
  • REALI says the results are not a poll and do not show voter opinion.

Why it matters

More people may ask chatbots about candidates before they vote. This study shows the answer can change depending on which chatbot you ask and which sources it draws on. Treat a chatbot's description of a politician as one view, not a neutral summary.

Cartoon characters for candidates

As technews.tw reported, REALI, an AI communications company, studied how ChatGPT and Gemini describe the main mayoral candidates in Taiwan's six biggest cities. The study came ahead of the 2026 local elections. The headline's 'anime styles' refers to cartoon characters the AI compared each candidate to, not to drawn pictures.

Taipei Mayor Chiang Wan-an (蔣萬安) was likened to Doraemon, a character with a tool for every problem. Shen Po-yang (沈伯洋) became Luffy from One Piece, a leader who gathers a crew and takes on challenges. In New Taipei, Li Szu-chuan (李四川) matched Bob the Builder and Su Chiao-hui (蘇巧慧) matched Elastigirl from The Incredibles.

How the test worked

REALI used its own tool to put the same request to both chatbots 15 times for each of 12 politicians. Each time, the chatbot gave five keywords describing the person. That made 30 answers per person, and REALI counted which keywords came up most. ChatGPT then picked the cartoon character that fit those keywords best.

For Chiang, the top keywords were 'incumbent advantage' (26 times) and 'steady' (24). For Shen, they were a nickname-style wave label (22), 'national security / civil defense' (21) and 'challenger' (16).

Where the two chatbots disagree

ChatGPT described Chiang with a 'record in office' label in 12 of 15 answers, and Gemini never used it. A 'participatory governance' label for Shen showed up 11 times in ChatGPT and 0 in Gemini. Some labels held steady, though: Li's 'engineer type' tag appeared in 27 of 30 answers across both models.

The sources differed too. News outlets made up about 50% of ChatGPT's citations and 75% of Gemini's. ChatGPT cited Central News Agency most (712 times), while Gemini cited United Daily News most (496). Dcard, a Taiwanese forum, was cited 108 times by ChatGPT and 3 by Gemini. YouTube went the other way, with 66 citations from Gemini and 6 from ChatGPT.

What the study does not show

REALI says this is an analysis of AI answers to a specific set of questions. It is not an election poll. The results do not show what voters think of any candidate or how much support they have.

What to watch

  • Whether REALI repeats the study closer to election day to see if the labels change.
  • Whether the chatbot makers say anything about how they pick sources for political questions.

Try it

Ask two chatbots the same question about a local politician, then compare the labels they use and the sources they cite.

Sources and links

First reported by technews.tw. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news