Whose answers do the numbers represent?
A percentage labeled support can come from a person on a telephone or from a generated character. Taichung coverage now includes differing human surveys and AI simulations. Can these share one trend line? Senge argues that the data-generating process and target of estimation must be classified first. Averaging before establishing comparability can create an appearance of stability.
TPOC discloses 18,971 virtual agents, a November 10 execution date, and that its work is not a sample of real voters. Agent count describes simulation scale; representation of society still requires external validation.[6]
Human surveys also need to be separated
TVBS’s November 3 report gives Ho Hsin-chun 41.1% and Johnny Chiang 30.4% in the Hsinmin survey, while separately listing projected vote ranges. Reported interviews ran October 28–30, with 1,077 completed landline and mobile interviews. Support and projected vote share are distinct quantities.[1]
NOWnews’s November 9 report on a KMT council caucus-commissioned Shih Hsin survey gives Chiang 39.2% and Ho 30.0%. Reported fieldwork ran November 5–7, with 1,068 interviews using telephone-directory numbers with randomized final digits. Timing and contact methods differ. This raises questions but cannot establish a dramatic switch by the same people.[2]
Compare three kinds of numbers
| Evidence | Production | Permissible use | Cannot establish directly |
|---|---|---|---|
| Human support | Interviewers record answers at a particular time [1][2] | Stated preference under specified conditions | Identical preferences among nonrespondents or on election day |
| Projected vote share | Responses combined with turnout and other parameters [5] | An estimate conditional on assumptions | Raw support or a guaranteed result |
| AI simulation | Virtual agents generated under model settings [6] | Output conditional on the settings | A count of real people interviewed |
This classifies uses rather than ranking institutions in advance. Human surveys may have coverage and nonresponse problems; models may have useful applications. Quality depends on sources, assumptions and validation, rather than a technology label.
What remains uncertain when counts grow?
Consider a conceptual example: if generated characters share an incomplete specification, adding characters may stabilize their internal average without supplying missing human experience. This is not an error estimate for a particular institution. It distinguishes stable output from accuracy about the world.
TPOC warns about amplification of visible narratives and says conventional sampling margins do not apply. Senge interprets this as a change in uncertainty sources, not their disappearance. Subsequent checks should cover model versions, input and execution dates, adjustments, and validation on data not used to tune the model.[6]
Human surveys require the same scrutiny. Reported margins of roughly three percentage points do not cover every effect of nonresponse, question order or contact mode. Without original data, this article computes no additional significance test for candidate differences and attributes no entire gap to institutional bias.[1][2]
Agreement and explanations still to test
CNA reports both campaigns’ reactions to Hsinmin; Rti follows later campaigning and another survey. TVBS and Newtalk organize the same Hsinmin material differently. Five articles permit comparison of editorial choices, not five independent measurements. Newtalk describes turnout and historical-result inputs to vote projections; without original calculation tables, we do not reproduce that estimate.[1][2][3][4][5]
The Taiwan Society for Targeting Misinformation discusses nonresponse and the spiral of silence as possible mechanisms requiring research. Differences in party identification might reflect composition, questions or real attitudes. An assumed correct partisan distribution cannot by itself invalidate all surveys departing from it.[7]
CNA quotes Ho describing enthusiasm associated with Ocean Current events. That is a candidate’s assessment. Estimating a vote effect needs evidence on exposure, prior preferences and later behavior, while addressing the greater inclination of existing supporters to attend. These reports cannot establish that causal effect alone.[3]
Conclusion: demand validation before a verdict
Senge recommends three public fields: who supplied the answers, which transformations produced the result, and which external data validated it. Reporting should retain fieldwork or execution dates, commissioners and limitations, while separating modeled output from human responses. This is more informative than a single ranking of percentages.
Candidates and readers can ask which groups are difficult to reach, which policy needs were not asked about, and whether conclusions survive changed assumptions. This article forecasts no winner and treats no simulation count as proof of predictive strength. Visible evidentiary boundaries make subsequent evaluation possible.