Debunking Antisemitic Conspiracy Theories using AI

Key findings

General conspiracy debunking via LLM. Rand's team asked 1,000 American participants to state a conspiracy theory they believed (rated 0–100) and provided evidence. A three-round dialogue with GPT-4 Turbo (prompted to debunk via evidence) reduced belief by 20 percentage points—a full standard deviation. Effect persisted at 10 days and two months. The model's success did not depend on the target conspiracy (works across 9/11, moon landing, etc.) and was equally effective on people with 100% baseline belief. Mechanism: evidence, not rhetoric. When the model was told to persuade without providing facts, the effect disappeared.

Antisemitic conspiracy theories are debunkable. Historical precedent might suggest antisemitic conspiracy beliefs (blood libels, Protocols of the Elders of Zion, Rothschild/Soros global domination, COVID-vaccine Jewish profiteering, Holocaust denial, great replacement) would be especially resistant—embedded in religious/ethnic/ideological identity and reinforced over centuries. Rand tested this on 2,400 Americans (Democrat/Republican/Independent split). Baseline: ~50% believed at least one of six major antisemitic conspiracies. Soros/Rothschild control (36%) and Jewish government/media control (33%) were most common; Holocaust exaggeration ~10%. Critically, belief in antisemitic conspiracy theories correlated strongly with low favorability toward Jews (r~–0.4), while belief in true conspiracies or aliens did not—showing these beliefs are uniquely tied to prejudice.

Claude 3.5 Sonnet debunking dialog reduced antisemitic conspiracy belief by 15 points (on 0–100 scale, ~16% decrease, >0.5 standard deviations). Minimal treatment (labeling as anti-Semitic without evidence) produced only a 5-point decrease. One-month follow-up showed ~50% effect durability. Crucially, among participants initially unfavorable toward Jews, debunking conversations increased favorability by 7 points (25% increase, ~0.5 standard deviations)—strong evidence for a causal pathway from conspiracy belief to prejudice, not just correlation.

Evidence matters more than source labels. Model framing (human vs. AI, partisan vs. neutral) did not predict efficacy—only the evidence and arguments did. This suggests the mechanism is genuine persuasion by facts, not social proof from trusted sources.

Broader context and applications

Rand frames this work as part of a larger research agenda on using LLM-assisted dialogue for correcting misperceptions. Beyond conspiracy theories, his team has shown similar effects on climate change belief, vaccine hesitancy, child vaccination, environment care, racial wealth gap understanding, and overdose prevention policy. The common pattern: AI tutoring works because models can diagnose what someone is confused about and provide tailored correction, resembling effective one-on-one teaching.

Caveats: not everyone is reachable. In the original conspiracy study, ~25% of believers were talked out of the conspiracy; the rest either stayed fixed or showed no change. In the antisemitic conspiracy study, similarly, some portion remained unmoved despite evidence.

Research implications

1. Conspiracy belief is more evidence-responsive than popular psychology suggests. The "once down the rabbit hole, it's too late" narrative is incorrect if you have the right evidence.
2. Antisemitic conspiracy beliefs are not uniquely protected by identity. They respond to evidence like other conspiracies—though they are embedded in prejudice, the link is causal (conspiracy belief → lower favorability) and reversible (debunk belief → higher favorability).
3. LLMs as tutors, not lecturers. The success of AI dialogue depends on its ability to identify and address specific misconceptions, not on rhetorical persuasion or source credibility.

Notes

This talk captures Rand's research trajectory: from showing that people will believe conspiracy theories despite overwhelming evidence, to showing that the right evidence (personalized, detailed, non-rhetorical) can actually move believers. The antisemitic conspiracy extension is important—it tests whether millennia-old prejudice narratives are truly intractable, and the answer is no. The 25% reachability rate is sobering but not pessimistic: a quarter of conspiracy believers is a substantial population, and the effect size is large.

The LLM tutoring framing is conceptually clean: don't assume a model needs to be a persuader or a trusted source to work. It just needs to diagnose and correct. This contrasts sharply with fact-checking research, which focuses on whether corrections are believed; Rand's work shows corrections are believed when evidence-driven and tailored.

One limitation: all studies are U.S.-based and recruited; generalization to other nations and education levels is open. Another: real-world uptake of AI-assisted conspiracy debunking is not demonstrated—the studies are lab-based. The connection between reduced conspiracy belief and reduced prejudice is measured immediately post-intervention; whether improved favorability toward Jews translates to behavioral change or resistance to future conspiracy narratives is unknown.