AI-based chatbots like ChatGPT and Copilot often falsify news and have trouble separating fact from opinion. A new survey conducted by 22 public broadcasters, including Deutsche Welle (DW), found that four of the most popular “artificial intelligence assistants” distort the news 45% of the time – regardless of language and region.

Journalists working for organizations such as the BBC (UK) and NPR (US) evaluated responses from ChatGPT, Microsoft Copilot, Google Gemini and Perplexity AI. Criteria included accuracy, citation, context, journalistic wording and the distinction between fact and opinion.

Major content errors

The study showed that almost half of the responses presented a serious problem, while 31% had deficiencies in the sources and 20% had significant errors.

DW noticed that 53% of responses to its own queries had problems, of which 29% were accuracy issues. Among the mistakes were examples such as wrongly mentioning Olaf Scholz as chancellor of Germany, when Friedrich Merz had already taken office, and mentioning Jens Stoltenberg as NATO Secretary General, even though Mark Rutte had already taken the post.

This is also something that particularly worries the authors of the study. Their research clearly shows “that these omissions and errors are not isolated incidents,” said Jean-Philippe de Tender, deputy director-general of the European Broadcasting Union (EBU), which coordinated the study. “They are systematic omissions and have neither specific borders nor language. We also believe that this puts public trust at risk. When people don’t know who to trust, they end up trusting nobody and that can affect participatory democracy.”

Broadcasters involved in the study are now calling on governments to take action and initiatives. The EBU said its members intend to put pressure on European and national authorities to implement laws on information integrity, digital services and media pluralism. In addition, they call for independent auditing of artificial intelligence systems due to the rapid development of new applications.

Edited by: Chrysa Vachtsevanu