Do chatbots steer Dutch voters toward the same parties?

Ahead of the Netherlands' 2025 election, the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) tested several AI chatbots and compared their output to two purpose-built voting-advice tools, Kieskompas and StemWijzer. Per its October 2025 report, chatbots "ongeacht de antwoorden van de gebruiker" — regardless of the user's answers — "verrassend vaak" (surprisingly often) returned the same two parties as the best match: PVV or GroenLinks-PvdA. One of those two topped the result in more than 56% of cases across the chatbots tested, and in more than 80% of cases for one chatbot specifically. Parties including D66, VVD, SP, and PvdD "nauwelijks voor[kwamen]" (barely appeared), and BBB, CDA, SGP, and DENK were "vrijwel onzichtbaar" (virtually invisible) — even, the report notes, when a user's stated positions matched them.

The AP locates the cause in how chatbots are built rather than in how they are used: unlike Kieskompas or StemWijzer, general-purpose chatbots are "niet ontworpen voor politieke doeleinden" (not designed for political purposes), and their answers are generated from "ongestructureerde gegevens van internet" (unstructured internet data) including social media and dated news sources — inputs the report treats as the likely source of a skew toward particular parties that holds regardless of the specific question asked. AP vice-chair Monique Verdier is quoted summarizing the finding bluntly: "Chatbots lijken slimme hulpmiddelen, maar als stemwijzer falen ze consequent" (chatbots look like smart tools, but as a voting guide they fail consistently), warning that voters "kunnen ongemerkt worden gestuurd" toward parties that do not match their actual views.

The mechanism the AP describes — a chatbot that sounds authoritative while its content tracks something other than the user's input — is the same one named more generally in How do chatbots enable distributed delusion differently than passive tools?: a conversational register that invites trust the content has not earned. It is the behavioral counterpart to How many Dutch voters might seek AI voting advice?, which measures how many Dutch voters say they would consult such a chatbot in the first place; this note measures what those voters would actually be told if they did.

The excerpt does not give the number of chatbots tested, the number of trials per chatbot, or the exact prompts used, so whether the PVV/GroenLinks-PvdA skew is stable across phrasings of the same political question, or specific to the chatbots AP happened to test, is not established here. What the report does support, at the strength given, is the general point it states explicitly — "Een chatbot klinkt misschien zelfverzekerd, maar dat betekent niet dat het antwoord klopt" (a chatbot may sound confident, but that does not mean the answer is correct) — applied to one of the highest-stakes judgments a chatbot can be asked to support.