Are AI Chatbots Spreading Foreign Propaganda?

The fear is easy to state. Feed a chatbot a steady diet of the internet, including the parts seeded by state propaganda operations, and it will parrot whatever falsehood is loudest. So when NPR and the media-rating group NewsGuard set out to test how the popular assistants handle foreign disinformation, the expectation in many quarters was that they would fail.

They mostly did not. In a study published on 30 August, the researchers built 30 questions around false narratives pushed by China, Iran and Russia between December 2025 and July 2026, then posed them to assistants including OpenAI's ChatGPT and Google's Gemini alongside the major search engines. The chatbots debunked the false claims more often than search did, according to NPR.

What the test found

Take one example the researchers used. After Russian forces shelled a historic Ukrainian monastery in June, Kremlin-aligned accounts claimed Ukraine had damaged the UNESCO site itself. Asked about the incident, the assistants generally flagged the claim as a known falsehood rather than repeating it.

The weak point was not the chatbots. It was the AI summaries that now sit above traditional search results. Those blurbs performed worse than the search links beneath them, and worse than the standalone chatbots, per the same reporting. That is an uncomfortable result for anyone who assumes the tidy answer at the top of the page is the most trustworthy thing on it.

Why this is not the whole story

None of this means the assistants are immune. Earlier research has shown that models can be nudged into repeating disinformation when a topic is obscure and the propaganda is dense, and the NewsGuard team has documented cases where chatbots absorbed Russian-planted claims. The results depend heavily on how a question is worded, and on whether the model reaches for a live search or answers from memory.

Still, the honest takeaway is a modest one. On a careful test of known falsehoods, today's chatbots were better at spotting foreign propaganda than the search tools most people still rely on. That is not licence to trust them blindly. It is a reason to retire the assumption that they simply launder whatever the internet tells them. If anything, the softer target is the summary box, not the chatbot. For a related look at what these systems get wrong on their own, see our piece on why models make things up.