Largest study of its kind shows AI assistants misrepresent news content 45% of the time – regardless of language or territory
gemini
| Source: Mastodon | Original article
A joint investigation by the European Broadcasting Union and the BBC has revealed that AI‑driven assistants distort news content in nearly half of their replies. The study, the most extensive of its kind, sampled responses from the major commercial platforms—ChatGPT, Claude, Gemini, and several regional bots—across ten languages and three continents. Overall, 45 % of the answers contained at least one factual error, a mis‑quotation, or a misleading omission. Gemini performed the worst, with errors in 58 % of its outputs, while the others hovered between 38 % and 48 %.
The findings matter because consumers are turning to voice‑activated and chat‑based assistants for quick news updates, often trusting the brevity of AI summaries over traditional outlets. Misrepresentations can amplify misinformation, skew public opinion, and undermine confidence in both media and the technology that delivers it. The cross‑language consistency of the problem suggests that the issue is rooted in the underlying model architectures and training pipelines rather than isolated data‑set quirks.
Regulators are already watching. The European Commission’s AI Act, slated for final adoption later this year, calls for “high‑risk” AI systems to meet strict transparency and accuracy standards. The study’s results are likely to accelerate scrutiny of AI news‑generation features and could prompt mandatory third‑party fact‑checking or real‑time provenance tagging.
What to watch next: major vendors have pledged to tighten internal validation, and the EBU plans a follow‑up audit focusing on mitigation techniques. Meanwhile, consumer‑rights groups are preparing a petition for mandatory disclosure when an answer is AI‑generated. The next few months will test whether industry fixes can close the 45 % gap before the technology becomes an even more entrenched source of daily news.
Sources
Back to AIPULSEN