The Challenge
For decades, conspiracy belief has been treated as a kind of psychological bedrock, something people arrive at to satisfy deeper needs for certainty, control, or identity, and something facts alone can't dislodge. MIT's Thomas Costello, together with Gordon Pennycook (Cornell) and David Rand (MIT Sloan), set out to test that assumption directly: could a single, personalized conversation with an AI model meaningfully, and durably, change someone's mind about a conspiracy theory they already held?
Testing this required a sample most panels simply can't produce. The researchers didn't want people who would check a box for a few extra cents; they needed thousands of Americans who held real, self-articulated conspiracy beliefs, described in their own words, along with the specific evidence they thought supported them. On top of that, the study design was operationally demanding: each participant needed to complete a real-time, three-round back-and-forth dialogue with GPT-4 Turbo, tailored to their individual conspiracy and stated reasoning, sustained across an average of 8.4 minutes of interaction. And, with data clean enough to survive review at one of the world's most rigorous scientific journals, followed by 10-day and 2-month follow-up waves to test whether the effect held.
The Solution
The research team fielded both core studies through CloudResearch's Connect participant pool, drawing quota-matched American samples (balanced on age, gender, race, and ethnicity) totaling more than 2,000 respondents across two experiments. Connect's screening tools let the researchers pre-qualify for writing quality and coherence, critical for a design that lived or died on participants' ability to articulate genuine, detailed conspiratorial beliefs rather than low-effort survey noise. That same attentiveness carried through the demanding, multi-round AI conversation itself, which called for real engagement, not click-through completion. Connect also supported the study's longitudinal component, enabling the team to re-contact the same participants 10 days and 2 months later to test whether belief change held, recontact studies that retained roughly 84% and 68% of the original treatment sample, respectively.
Results
The findings, published in Science in 2024, were striking. Across two studies, participants who engaged in a personalized AI conversation reduced their belief in their own chosen conspiracy theory by roughly 20% on average relative to control: an effect that held steady, virtually undiminished, at both 10-day and 2-month follow-up. The reduction occurred across a wide range of conspiracies, from JFK and the moon landing to COVID-19 and 2020 election fraud, and even among participants whose beliefs were most deeply entrenched and central to their identity. The debunking effect also spilled over: participants became measurably less likely to endorse unrelated conspiracy theories after the conversation, and more likely to say they'd disengage from people or social media accounts promoting conspiracies.
The study was covered by The New York Times and became one of the most widely discussed findings at the intersection of AI and behavioral science.
What's Next
The study reframed a long-standing assumption in psychology, that conspiracy belief is functionally immune to evidence, and demonstrated that even research questions requiring large, hard-to-reach, genuinely-committed populations can be answered rigorously online and at scale. Researchers designing studies that depend on real, verified participants holding specific beliefs, behaviors, or experiences can build similar studies using Connect's screening and recruitment tools.
About MIT
The Massachusetts Institute of Technology is one of the world's leading research universities, based in Cambridge, Massachusetts. Researchers at MIT Sloan and across the Institute work at the frontier of behavioral science and human–AI interaction, producing research that shapes how the field understands technology's influence on human belief and behavior.