It Takes So Little to Change So Much: Investigating the Robustness of a Danish Voting Advice Algorithm
A popular Danish election tool's recommendations change drastically with minor tweaks, raising trust concerns.
A new research paper from Giovanni Astante, Roberta Sinatra, and Vedran Sekara reveals significant vulnerabilities in a widely used Danish election tool. Through a freedom of information request, the team gained access to the inner workings of the 'Kandidattest', a Voting Advice Application (VAA) implemented by a major Danish news outlet for general, municipal, and European elections. The study's core finding is that the algorithm's output is highly sensitive to minor adjustments, meaning the agreement percentages shown to voters—which 45% reported following in the last election—cannot be considered stable or fully trustworthy.
The researchers conducted an algorithmic audit, using simulated responses to test the tool's brittleness. They discovered that small changes to the algorithm's weighting of questions or the total number of questions in the questionnaire could produce different matching results for the same user. This lack of robustness suggests the tool's recommendations are not as objective as they appear, potentially influencing voter behavior and, by extension, election outcomes in Denmark's multi-party system. The paper calls for greater transparency and rigorous evaluation of VAAs, which play an outsized role in modern democratic processes but have largely escaped technical scrutiny.
- The 'Kandidattest' VAA, used by a major Danish media outlet, was found to be algorithmically brittle.
- Small tweaks to question weights or count led to different candidate matches in simulated audits.
- 45% of surveyed Danish voters reported following the tool's recommendations in the last general election.
Why It Matters
Highlights critical need for algorithmic transparency in tools that directly influence democratic decision-making for millions.