Chatbots debunk 75% of propaganda prompts

NPR and NewsGuard tested six internet-connected chatbots with 30 questions built from false narratives pushed by China, Iran, and Russia; the chatbots correctly challenged the premise about three-quarters of the time, while AI summaries above search results performed worse, according to the August 30 test. Researchers collected responses in mid-July, so this is a snapshot rather than a production guarantee. Builders should evaluate retrieval-grounded agents against adversarial claims instead of treating web access as truth, the same governance logic behind testing agents before giving them consequential powers.