Study finds AI chatbots fail to prevent conspiracy theory promotion
Researchers from the Digital Media Research Centre tested six widely-used AI chatbots and found that many failed to refuse or correct conspiracy theories when asked about them, with some actively encouraging further discussion. The study tested systems from OpenAI, Microsoft, Google, Perplexity, and xAI on nine theories including debunked claims and newer election-related conspiracy theories.
What happened
Researchers from the Digital Media Research Centre tested six AI chatbots by posing questions about nine conspiracy theories using a "casually curious" persona.
The chatbots tested were ChatGPT 3.5, ChatGPT 4 Mini, Microsoft Copilot, Google Gemini Flash 1.5, Perplexity, and Grok-2 Mini in both default and "Fun Mode" versions.
The theories included both long-debunked claims such as the JFK assassination and chemtrails, and newer conspiracy theories related to breaking news including false claims about the 2024 U.S. election.
Reported impact
Many chatbots did not shut down conversations about conspiracy theories and some actively encouraged them, raising concerns about the effectiveness of safety guardrails.
The research is cause for concern given how easily people can fall down conspiracy thinking, particularly when casually curious users interact with systems that fail to correct misinformation.
Index analysis
Safety guardrails are checks and balances designed to prevent chatbots from creating harmful content, but the study examined whether these guardrails were satisfactory in protecting users from conspiracy theory exposure.
Governance controls
The research has been accepted for publication in a special issue of M/C Journal and examines the adequacy of existing safety measures on deployed chatbot systems.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.