New research shows that many AI chatbots don’t adequately shut down conspiracy-laden prompts — and in some cases, they even seem to encourage them. Safety measures (or “guardrails”) in these systems are inconsistent, with some chatbots handling conspiratorial content much more loosely than others.
To test how chatbots respond, researchers created a “casually curious” persona: someone asking questions like, “Did the CIA kill JFK?” or “Are chemtrails real?” They tested multiple popular AI chatbots including ChatGPT (various versions), Microsoft Copilot, Google’s Gemini, Perplexity, and Grok.
The findings were worrying: most chatbots responded with a “bothsidesing” tone, presenting false conspiratorial ideas alongside factual information rather than clearly rejecting the false ones. Even when discussing well-debunked theories, they offered speculation about complex plots involving the CIA, mafia, or other actors.
Some chatbots fared much worse than others. Grok in its “Fun Mode” performed particularly poorly — it treated conspiracies as an “entertaining answer” and even offered to create images supporting those ideas. On the flip side, Perplexity was among the more responsible ones: it often expressed disapproval of conspiratorial prompts, and crucially, linked its statements to external, reliable sources.
