Kandidattest is a Danish voting advice application used by 45% of voters in recent elections. It matches voters to candidates by comparing their answers to a standardized questionnaire, producing agreement percentages that guide electoral decisions. The tool mediates between citizens and representatives at national scale.
Astante, Sinatra, and Sekara obtained the algorithm's internal mechanisms through a freedom of information request and audited its reliability. The finding: small changes to the algorithm — adjustments to question weights, modifications to questionnaire length — lead to different results. The matching percentages that voters see are sensitive to design choices that are invisible to users and not grounded in any principled methodology. The agreement score between a voter and a candidate is an artifact of implementation decisions, not a stable measurement.
The fragility is not in edge cases. It is in the ordinary operation of the tool. Standard decisions a designer makes when building such a system — how much weight to assign each policy area, how many questions to include, how to handle non-responses — produce different voter-candidate matchings. The output looks precise (a percentage), but the precision is cosmetic. The underlying measurement is not robust to reasonable alternative designs.
The through-claim is about the gap between appearance and stability. A percentage communicates confidence — 78% agreement reads as a measured quantity with a specific value. But if changing three design parameters shifts the number from 78% to 62% or 91%, the number is not measuring agreement. It is reporting a function of arbitrary choices. When 45% of voters consult a tool whose outputs are sensitive to decisions no voter can see or evaluate, the tool is not informing the election. It is shaping it, and the shape depends on the designer, not the voter.