European Broadcasting Union: AI Systems Often Misrepresent News Content

A study by the European Broadcasting Union (EBU) on the news integrity of AI chatbots concludes with the alarming finding that artificial intelligence continues to be an unreliable source for news consumption. The large-scale, cross-market investigation, in which 22 public service media organizations from 18 countries and leading AI assistants such as ChatGPT, Copilot, Perplexity, and Gemini in 14 languages were evaluated, shows: Errors in handling news do not occur in isolation with these systems. They are systemic and extend across all languages and platforms.

Overall, according to the study published on Wednesday, 45 percent of all over 3000 answers to the 30 core questions about current events contained at least one significant problem that could mislead the reader. When answers with certain deficiencies are included, even 81 percent of the results show an inaccurate representation.

By far the biggest issue area is source tracking, which researchers classified as significantly flawed in 31 percent of all answers. These deficiencies manifest in generative AI systems citing claims that are not supported by the provided source, or omitting source references altogether. Google Gemini, in particular, stands out negatively in this area: 72 percent of its outputs showed significant deficiencies in sources. This far exceeds the error quotient of the other assistants, which is less than 25 percent.

Outdated Information, Fabricated Links

The accuracy of facts and the provision of sufficient context also remain critical weaknesses, leading to significant deficiencies in 20 and 14 percent of cases, respectively. Scientists frequently identified fundamental factual errors, such as the mention of outdated political officeholders, for example, in the case of former Chancellor Olaf Scholz (SPD). Or – as a recurring problem with ChatGPT, Copilot, and Gemini – the listing of Pope Francis, who died in April, as the current head of the church in May. 2025. Perplexity claimed in its answer to "Is Viktor Orbán a dictator?" that Tagesschau describes the Hungarian Prime Minister's rule as authoritarian and illiberal, which was not in the cited source. Gemini, Copilot, and ChatGPT fabricated links that appeared credible but did not exist at all.

The study was led by the BBC; ARD and ZDF participated from Germany. While the results indicate a general, slight improvement in AI quality since a previous BBC study, with the proportion of answers with any kind of significant deficiencies falling from 51 percent to 37 percent. However, according to the authors, the awareness of the problem is exacerbated by users' unfounded confidence; for example, more than a third of adults in Great Britain trust AI summaries. This is particularly delicate, as the refusal rate for AI assistants to answer questions has dropped to just 0.5 percent. This proves the tendency to provide a low-quality answer rather than no answer whatsoever.

Furthermore, errors in AI-generated news threaten the reputation of trustworthy media brands, as 42 percent of surveyed adults place less trust in the original news medium if the AI answer contained errors. According to the analysis, this represents a significant risk, especially since AI assistants often distort, misattribute, or add editorial judgments to content from public service media.

Call for Legislators and Regulators

Considering the serious deficiencies, the EBU demands: AI developers must urgently reduce error rates and establish transparent reporting on their performance by language and market. Publishers and broadcasters need more control over the use of their content and a clear, agreed citation method with prominent links to the original sources.

Thirdly, policymakers and regulatory authorities must hold AI providers accountable for the quality of their products and ensure that public service media content is presented visibly and prominently. Users should also be better informed about the limitations of chatbots. Manufacturers such as OpenAI, Microsoft, or Google must urgently develop standards for safety, accuracy, and transparency. The research team has published a "toolkit" for this purpose.

Just last week, the German state media authorities warned based on a report by Hamburg computer science professor Dirk Lewandowski: "AI-based search answers create new content and displace established information sources." This has "far-reaching consequences for the visibility of journalistic offerings, the refinancing of media, and the diversity of online information." Traffic losses for publishers and broadcasters threaten the refinancing of content production, "which is essential for a diverse information landscape."

(vbr)