Can a chatbot talk people out of conspiracy theories they've believed for years, and does it stick?

This explores whether an AI conversation can shift people who already hold entrenched conspiracy beliefs, as opposed to fresh rumors about breaking news. The corpus says yes, modestly and durably. It doesn't say how old the theories in these studies were, so the "long-established" part is plausible rather than proven.

The strongest evidence is a study of 2,190 conspiracy believers. Personalized AI dialogue cut their beliefs by about 20%, the effect held two months later, and it even spilled over to unrelated conspiracies (Can AI reduce conspiracy beliefs by tailoring counterevidence personally?). What did the work was tailoring the counterevidence to the specific belief the person held, not profiling their demographics. That suggests generic debunking fails because it rebuts a stock version of the theory that the person doesn't actually believe. The spillover hints at a small worldview-level shift, not just one corrected fact. Twenty percent is a dent, not a conversion.

A second study looked at the opposite end, theories that are still forming. After the 2024 Trump and 2025 Kirk assassination attempts, brief LLM conversations reduced conspiracy beliefs about those events. They also left some lasting skepticism toward conspiracy claims about later events (Can LLM conversations reduce conspiracy beliefs as events unfold?). Fresh rumors haven't had years to harden, so the long-established case is the harder test. The corpus can't show whether effect sizes differ between the two.

The catch is that persuasion works in both directions. When an LLM is prompted to persuade, people's political beliefs move, and a brief warning that this can happen cut the shift by 48% without lowering trust in AI generally (Can a simple warning reduce how much LLMs persuade people?). A warning like that might also blunt a benevolent debunker, though the notes don't test it. There's a second weak point. LLMs abandon correct answers when users push back persistently, with no new evidence (Can models abandon correct beliefs under conversational pressure?). Better reasoning training doesn't fix that (Can better reasoning training actually reduce model sycophancy?). A committed believer arguing back turn after turn is exactly that kind of pressure, and the corpus doesn't show how debunking holds up against it.

There's also an open question about why it works. Users trust answers with more citations even when the citations are irrelevant (Do users trust citations more when there are simply more of them?). Some of a debunker's pull may come from looking authoritative, not from the quality of its argument. So the answer is a cautious yes. A tailored AI dialogue can loosen entrenched beliefs in a lasting way, but the corpus doesn't yet say whether it survives an adversarial believer.