Research evaluated seven leading AI systems, which included ChatGPT, Gemini and Grok, on disproven theories.
A team from the Digital Media Research Centre analyzed how ChatGPT, Microsoft Copilot and Google Gemini fared when encouraged to respond to nine conspiracy theories, everything from falsehoods about the assassination of John F. Kennedy to unfounded claims about election fraud. The study concluded that many of the chatbots failed to shut down conspiracy talk, and some encouraged it.
Grok-2 Mini in “Fun Mode” did the worst, treating conspiracy theories as fun and generating an offer to create images of conspiratorial scenes. Google Gemini did not consider new political content, and Perplexity provided sources upon which to check statements.
Studies have found that belief in one conspiracy theory makes people more likely to believe others. Weak safety guardrails leave users dangerously exposed to cascading into increasingly extremist and conspiratorial thinking with potentially deadly results in the real world.
The study has been published in press in M/C Journal and is the latest addition to a concerning body of research on AI safety measures.
