AI chatbots have been shown to be “unreliable and clearly biased” when providing election recommendations, the Dutch data protection authority (AP) has reported, issuing warnings about a possible risk to democracy just over a week before national elections.
Study Shows Repeated Partiality in AI Suggestions
The four chatbots tested by the authority “often end up with the same two parties”, irrespective of the user’s question or command, the regulator noted in a report published ahead of the late October vote.
For over 50% of instances, the AI recommended either the far-right Freedom party (PVV) of the politician Geert Wilders or the progressive GroenLinks-PvdA led by the ex-EU commissioner Frans Timmermans.
Some parties, such as the moderate CDA, “are rarely ever cited”, even when the voter’s query exactly matches the stances of these political entities, the study indicated.
Official Warns Against Relying on AI Assistants for Voting Guidance
The deputy head of the AP, the official Monique Verdier, stated that while AI assistants appear to be smart resources, “as a voting aid, they consistently fail”. Voters were being pushed toward a political group that might not correspond to their political views, she added.
“This immediately affects a fundamental pillar of democracy: the integrity of free and fair elections,” said Verdier. “We thus caution against using AI chatbots for voting advice, as their operation is unclear and difficult to verify.”
Election Context and Political Landscape
The Dutch head to the polls on October 29 in an electoral contest that will be keenly observed around Europe for the performance of the PVV.
The right-leaning group of Geert Wilders is leading the polls but the lead over the left-wing and centre-right parties seems to be closing, with numerous electors still undecided.
All major parties have ruled out an alliance with the PVV, meaning the runner-up party is most likely to provide the next prime minister.
Unintentional Bias Attributed to Chatbot Functioning
The AP stressed that the AI systems were not intentionally prejudiced, and that their “identified shortcomings” are a consequence of the manner in which AI assistants function.
