Why do chatbots asked who to vote for keep steering everyone toward the same couple of parties?

This explores why structured voting-advice tools like Kieskompas give more trustworthy recommendations than general-purpose chatbots. A caveat first: the collection has no study of Kieskompas itself. What it does have is a close look at how chatbots fail at the same job, and that shows clearly what a structured tool does differently. The short answer is that Kieskompas works from a fixed, checked map of where each party stands. A chatbot works from whatever the internet says about politics, and that is a different kind of input.

The key evidence comes from the Dutch Data Protection Authority. It tested four chatbots as voting advisors and found they recommended just two parties, PVV and GroenLinks–PvdA, in more than half of all cases. That happened even when the test user's answers clearly matched a different party Do chatbots steer Dutch voters toward the same parties?. Centrist parties came up less than 2% of the time. The regulator called this a 'vacuum cleaner effect' that pulls voters toward the two poles of the debate Do AI chatbots systematically bias voters toward extreme parties?. The cause it identified matters here: chatbots rely on unstructured internet text rather than structured data on party positions. The loudest, most-discussed parties take up the most space online, so they also take up the most space in the chatbot's answers. A tool like Kieskompas gives every party a position on every statement, so a quiet centrist party gets the same weight as a loud populist one.

Getting facts right does not fix this. By early 2026, ChatGPT and Google's AI were making essentially no verifiable factual errors on US voter questions. Yet they sent people to official election sites less than half the time and leaned on editable sources like Wikipedia Do AI chatbots give voters accurate election information?. A broader study of more than 3,000 AI answers about news found serious sourcing or accuracy problems in a large share of them Do AI assistants reliably answer questions about news?. A chatbot can state every fact correctly and still give a skewed overall picture. Where a structured tool keeps its advantage is in being traceable: you can see which statements produced your match. Research on AI evaluation points the same way. AI judges that gather explicit evidence before deciding are about 100 times more stable than ones that rely on their own impressions Can agents evaluate AI outputs more reliably than language models?. Grounding a judgment in checkable evidence is what makes it consistent.

The part you might not have expected is about trust, not accuracy. People tend to trust ChatGPT because talking to it feels responsive and natural, and that trust has little to do with whether it is correct Does conversational style actually make AI more trustworthy?. So the chatbot's weakness is hidden by the very thing that makes it appealing. A questionnaire feels mechanical but is reliable. A chat feels personal but is biased toward the same two parties. This matters at scale: an estimated 1.8 million Dutch voters, mostly younger ones, say they might ask AI for voting advice How many Dutch voters might seek AI voting advice?. Use is lower in the US, at about 15% of voters, and also concentrated among younger people Will voters actually use AI chatbots for election information?.

Chatbots are not useless for political topics. In one study, short conversations with an AI representing the other political side corrected people's mistaken beliefs about that side and made them feel warmer toward it. The effect came from accurate information, not persuasion tactics Can AI chatbots reduce partisan misperceptions and warm cross-party feelings?. The pattern across these studies is that chatbots do well when they rest on a defined, accurate body of information. They drift toward whatever gets the most attention online when they don't. Kieskompas is reliable mainly because of its structured data on party positions, not because it uses an older format.