Chatbots resisted state propaganda better than AI search summaries

An experiment conducted by NPR with NewsGuard found that leading AI chatbots mostly pushed back against false narratives promoted by Russia, China and Iran, while the AI-generated summaries that now sit atop search results performed less well. The researchers built 30 questions from narratives that first surfaced between December 2025 and July 2026 and put them to chatbots including OpenAI's ChatGPT and Google's Gemini, alongside the largest search engines.

One example cited: after Russian shelling damaged a historic Ukrainian monastery in June, Kremlin-aligned outlets and accounts claimed that Ukraine had damaged the UNESCO site instead.

The result runs against the intuition that grounding an answer in live search should make it more reliable, and the reason it does is worth understanding. A chatbot answering from its trained parameters, with safety tuning applied to known influence operations, draws on a corpus in which the debunking is usually better represented than the claim. A search summary is constrained to synthesise the pages the ranker returned. If a coordinated campaign has succeeded in getting its outlets to rank for a narrow query — which is the entire point of the campaign — the summary does not evaluate those pages so much as compress them, handing the reader a laundered version in the search engine's own voice.

Retrieval, in other words, becomes a liability when the retrievable corpus is the thing under attack. That inverts the usual assumption that grounding reduces fabrication, and it does so precisely where accuracy matters most: narrow, recent, contested events, where authoritative coverage is thin and adversaries have had time to fill the gap.

The practical implication is that the two systems need different defences. A model can be evaluated against known false narratives before release. A retrieval pipeline has to be defended continuously, because its inputs change daily and its adversaries are optimising against it — and a fluent summary is a poor place to discover that.