Abstract
Large language models (LLMs) can persuade people to revise their beliefs, but it is unclear which beliefs are most "debunkable," and why. Across three pooled studies, participants (N = 3,341 online U.S. adults) discussed a self-reported conspiracy or other epistemically suspect belief with GPT-4, instructed to debunk it using facts and evidence. An LLM classified each belief by content domain (i.e., conspiratorial, paranormal/supernatural, pseudoscience, political, religious, other) and evidential status, from strong counterevidence to supported by evidence. Confidence decreased in every domain, but belief updating tracked evidential status rather than domain: reductions were substantial wherever the belief was open to counterevidence — strong, weak, or uninvestigated — and small for non-falsifiable and supported beliefs. Domain no longer predicted change once evidential status was accounted for, and the counterevidence the LLM deployed partially mediated the gradient. AI debunking appears most effective for falsifiable beliefs resting on weak evidence.