Do AI chatbots systematically bias voters toward extreme parties?

The Dutch Data Protection Authority (Autoriteit Persoonsgegevens) warned voters against relying on AI chatbots for voting advice, after its own research found the tools systematically biased toward two parties. The authority tested four well-known AI chatbots against official Dutch voting aids such as Kieskompas and StemWijzer and found the chatbots "frequently recommended only two political parties" — the far-right Party for Freedom (PVV) and the left-wing GroenLinks–PvdA alliance — "when asked which party best matched the users' views." PVV was the top recommendation in more than 30% of tested cases, GroenLinks–PvdA in nearly 25%, while centrist parties such as the Christian Democratic Appeal and the Christian Union were suggested in only 1.3% and 0.3% of cases. Vice-chair Monique Verdier said "chatbots seem like clever tools, but as a voting aid, they consistently fail."

Researchers called this a "vacuum cleaner effect": left-leaning users were drawn toward GroenLinks–PvdA and right-leaning users toward PVV, producing what they termed "distortion and polarization" of the political landscape. The mechanism the watchdog identifies is a mismatch between appearance and function — a tool framed as personalized matching advice instead collapses many different inputs into two fixed attractors. Verdier spelled out the resulting risk to voters directly: "if, regardless of what you enter, you end up with two parties, you might think, 'I have to vote for that party,' even though that party doesn't align at all with your preferences." The authority's remedy is procedural, not technical: it urged developers to have chatbots refuse voting-advice requests and redirect users to official election-guidance sites.

This cuts against Can sycophantic AI advice still push people away from polarized views?, where advice from a measurably sycophantic model moved participants away from their initial leanings on average across 30 decision environments. The discrepancy looks like a difference in task structure rather than a contradiction: that paper's model offered open-ended considerations, while the Dutch chatbots performed a fixed multi-party classification that collapsed toward two attractors regardless of input. It extends the same structural point as How do feed ranking weights shape what content gets produced? to a different mechanism — a tool presented as neutral matching is a political lever on which options voters even see — and it resembles Can we detect when language models flip their stance to please users? in kind, both describing chatbot output converging toward a narrow set of attractors, though here the convergence appears across many users rather than within one user's stated preference. It also extends Can a simple warning reduce how much LLMs persuade people? from one-on-one persuasion to recommendation-matching as a second channel of democratic risk.

The excerpt does not name the four chatbots tested, the prompts or methodology used to elicit recommendations, or whether the 30%/25% figures reflect many prompt variations or a single test run. It does not report statistical methodology or confidence intervals, and it does not establish that the clustering changes actual votes — only that the pattern exists and that the regulator infers the behavioral risk from it. The defensible reading is narrow: a national regulator found a specific matching failure in four unnamed chatbots and recommended refusal rather than a technical fix, which does not establish that AI voting tools are biased in all deployments or beyond this Dutch sample.