[The Public = Reporter Kim Jong-yeon] As controversy over alleged design bias in a Daejeon mayoral poll commissioned by TJB Daejeon Broadcasting from JoWon C&I and Research & Research on April 18–19 has grown, The Public submitted the original questionnaire to three leading artificial intelligence models for independent analysis. Each model reached the same conclusion: the questionnaire contains structural bias.
The models we used were OpenAI’s GPT (version 5.4), Google’s Gemini (Pro), and Anthropic’s Claude (Opus 4.6). For each system we opened a fresh session, attached the same image of the original questionnaire, and asked only that it identify problems in the poll’s items. We gave no additional prompts intended to produce a particular outcome.
The poll in question was conducted by TJB at JoWon C&I’s request using CATI (computer-assisted telephone interviewing) with 100% virtual mobile numbers. The sample comprised 800 Daejeon residents aged 18 and older, selected by proportional random sampling across sex, age group, and region. Interviewers made 6,983 call attempts and obtained 800 completed interviews, yielding an 11.5% response rate. The reported margin of error is ±3.5 percentage points at a 95% confidence level. The published results showed Democratic candidate Heo Tae-jeong at 46.3% and People Power Party incumbent Lee Jang-woo at 22.9%, a 23.4 percentage-point gap.
Core finding all three models flagged: the priming effect
All three AI models pointed first and most forcefully to a priming effect driven by the order of questions.
In the questionnaire, the main sequence runs: Q1 party support → Q1-1 reclassifying respondents who say they have no party → Q2 evaluation of President Lee Jae-myung’s governance → Q3 frame of supporting versus checking the administration → Q4 suitability for the next Daejeon mayor.
GPT characterized that sequence not as neutral measurement but as a frame-inducing strategic design. By putting respondents into a party→administration frame before asking about candidates, the questionnaire measures reactions after respondents are already placed inside a partisan context rather than capturing an unaffiliated assessment of the candidates.
Gemini made a similar point: asking about party support, presidential approval, and the support-vs.-check frame before assessing candidate strength effectively primes respondents with national partisan logic. To reduce distortion and better capture local voter sentiment, Gemini recommended asking about mayoral suitability first, followed by party support and national governance questions.
Claude likewise noted that activating party identity and presidential evaluation before asking candidate suitability encourages respondents to judge candidates as extensions of a party or president rather than on their individual merits. Standard survey practice, Claude said, is to ask about candidate suitability and vote intention first, with party support and national evaluations afterwards. The current ordering, Claude concluded, risks overemphasizing party dynamics.
Problem of forcing nonpartisans into a category
The three models also criticized Q1-1, which asks respondents who selected “no supported party” in Q1 to indicate a party they nevertheless feel some affinity for.
GPT warned that pressuring politically disengaged or genuinely nonpartisan respondents to pick a party produces artificial party leanings; the poll thereby misrepresents true party support by forcing nonpartisans into categories. Gemini described the approach as a fatal error that effectively collapses active nonpartisans and refusal voters into ambiguous responses by coercing them into choosing “don’t know” or “other,” preventing accurate measurement of the nonpartisan vote.
Order of party names and the primacy effect
Claude raised an additional technical issue the others did not emphasize: the questionnaire presents parties in the order of National Assembly seats, which means the Democratic Party is always read first. Claude noted that the Central Election Polling Deliberation Commission recommends rotating order by respondent or using alphabetical order. Listing parties by seat count can produce a primacy effect that boosts the first-listed party—here, the Democratic Party—by roughly 2–4 percentage points. Claude also flagged a logical inconsistency: the Justice Party appears on the list despite holding zero seats, so using seat count as the ordering rule is inconsistent.
Leading wording on administrative integration item
Claude also critiqued Q10, which refers to “Daejeon-centered administrative integration.” He noted the formal policy name is Chungcheong regional administrative integration or Daejeon–Sejong–Chungnam administrative integration; the phrase “Daejeon-centered” frames the proposal positively for Daejeon respondents by implying local leadership, producing a pro-integration bias. Gemini flagged the education superintendent item as similarly problematic: because education superintendent elections are legally nonpartisan, asking about them immediately after heavily politicizing respondents—and then asking whether they prefer a progressive, centrist, or conservative superintendent—nudges respondents toward ideological selection, a textbook framing effect.
At the end of its analysis, Claude estimated the combined impact of these design choices. Priming via the presidential-evaluation question can create a pro-Democratic atmosphere; listing parties by seat count produces a primacy boost for the Democratic Party; and the asymmetric frame in Q3 can suppress ruling-party support. With these effects acting together, Claude judged there is a plausible scenario in which the conservative candidate’s support is understated by roughly 3–7 percentage points. GPT echoed this assessment, noting that the candidate-suitability measure produced by this questionnaire will likely reflect not a neutral “who is suitable” judgment but rather “who fits the party or administration frame the respondent has just adopted.”
