Researchers noted that AI chatbots were dynamic, making the experiment hard to replicate. In a series of a dozen tests with similar questions carried out by POLITICO on Tuesday, the chatbots either declined to respond entirely or else had updated responses with links directing users to the EU institutions’ websites.

Meyer-Resende said the experiment was, however, large enough to be representative. It also provided new evidence about the risks of so-called AI hallucinations — which often occur because of insufficient training data, biases and false assumptions —ahead of the European election, which takes place from June 6-9.