Privacy regulator has major concerns about ‘colored voting advice’ from AI bots: ‘This affects the integrity of fair elections’

Privacy regulator Dutch Data Protection Authority finds it worrying that more and more voters are using chatbots powered by artificial intelligence (AI) for voting advice. These opinions are “unreliable and clearly biased,” AP concludes Tuesday after our own investigation. For more and more voters, the AI ​​chatbots complement long-standing online voting aids for advice about the House of Representatives elections in eight days.

In its own research, in which voting advice was requested a total of 21,000 times, the regulator compared four well-known chatbots — those of ChatGPT, Gemini, Mistral and Grok — with well-known voting aids such as Kieskompas and Stemwijzer. Regardless of the user’s question or assignment, the AI ​​bots “remarkably often” recommend the same two parties: in 56 percent they end up with GroenLinks-PvdA or the PVV.

According to AP, one (unnamed) chatbot even reaches those two parties in 80 percent of the cases. It is not immediately clear why GroenLinks-PvdA and the PVV are relatively often given as voting advice. According to AP, this may be “partly due to the extent to which and the way in which these parties appear in the training data”.

Denk and CDA ‘almost never’ come up

Unlike GroenLinks-PvdA and the PVV, parties such as D66, SP, VVD and Party for the Animals appear much less often as the best-connected party. According to the research, BBB, CDA, SGP or Denk “almost never” emerge as the most important voting advice, even if the user exactly reflects their positions in conversations with the bots.

AP vice-chairman Monique Verdier states on the AP website that the AI ​​chatbots “systematically miss the mark”, with potentially major social consequences. Wrong voting advice “hits the cornerstone of democracy,” says Verdier, namely: “the integrity of free and fair elections.” “We therefore warn against using AI chatbots for voting advice, because their operation is not clear and verifiable. And we call on chatbot providers to prevent their systems from being used for voting advice.”

The AI ​​bots are trained with information retrieved from the internet. However, that data is not verifiable and may be outdated. For example, a simple question about a specific topic from one of the election programs with chatbots sometimes leads to an answer from a program from previous elections, which may have since been revised.

Give a gift





Part





Email the editor