Can LLM conversations reduce conspiracy beliefs as events unfold?

The paper tests whether a conversation with an LLM can reduce belief in conspiracies that are "immediately unfolding," not ones that have circulated for years. Experiment 1 (N = 472) ran in the days after the July 2024 assassination attempt on Donald Trump, and Experiment 2 (N = 1035) in the days after the September 2025 assassination of Charlie Kirk. U.S. adults who held conspiratorial views about the event had a multi-turn conversation with an LLM prompted to reduce their belief. Against controls who either discussed an irrelevant topic with an LLM or viewed a static fact sheet, the treatment group showed significantly reduced conspiracy beliefs in both experiments. The dialogue was short (M = 6.9 minutes) and also reduced suspicions of a cover-up and concerns about hidden factors.

The introduction explains why this was an open question. Well-known conspiracy theories are historical, so "large bodies of evidence debunking these theories have accumulated," and earlier work showed an LLM can draw on that training-data evidence to durably reduce belief. The reasoning is that an LLM functions as a way of querying collective knowledge and can "quickly, legibly provide compelling rebuttals to any given conspiracy topic," a job that is slow and hard for even experienced human debunkers. The discussion calls the result "perhaps surprisingly, quite effective" for theories that arise in the aftermath of consequential events. It also reports carry-over. Participants debunked about the first Trump attempt were less conspiratorial about the second attempt two months later. Participants debunked about the Kirk assassination were less generally conspiratorial one month later, with "some evidence" of persistence for a subsequent mass shooting.

This extends Can AI reduce conspiracy beliefs by tailoring counterevidence personally? from established conspiracy theories to fast-moving ones, and the delayed skepticism toward conspiracies about later events echoes that note's spillover finding. The static fact sheet control is at least consistent in direction with Where does AI's persuasive power actually come from?, which found conversation substantially more persuasive than static messages, though the excerpt does not report the treatment against each control separately.

The excerpt is silent on several points that matter for how far to lean on this. It gives no effect sizes, no belief measures, no model name and no follow-up retention rates. It does not say whether the dialogues elicited each participant's own version of the theory and tailored evidence to it, so the tailoring mechanism from the earlier study cannot be assumed here. It also does not say what the LLM drew on when the events were days old, or whether the carry-over reflects the dialogue itself or the intervening events. Within those limits, the finding is that in two U.S. cases the dialogue worked without the accumulated debunking the introduction associates with older theories. Whether it holds for other events, populations or situations where the facts stay unsettled remains untested in this excerpt.