A pilot AI system tested on more than 4,000 registered U.S. voters managed to shield their confidence in elections from false rumors — and the protection was still measurable a week later, across party lines. That is the finding from a team led by Caltech researchers who are turning the same technology behind deepfakes and super-spreading social media posts into a tool for good.
Their approach is called "pre-bunking," a simple but powerful idea: warn people about a false claim before they ever see it, so they are less likely to believe it when it arrives. "False claims can spread widely before fact-checkers have time to respond," explains Mitchell Linegar, the paper's lead author and a doctoral graduate of Caltech now at Washington University in St. Louis. "Pre-bunking gives people accurate information before exposure, making them less likely to believe these claims, potentially stopping their spread before it starts."
The challenge has always been scale. Scientists have known for years that pre-bunking works, but each message required a human expert to carefully craft it — too slow to keep pace with rumors that race around the internet in minutes. The Caltech team found a workaround: they combined one reusable, human-made prompt with verified election information, letting a large language model generate pre-bunking articles for new rumors quickly and without further human review. That model, described in the journal Royal Society Open Science, is the first of its kind tested on real voters.
The experiment ran before the 2024 U.S. Election Day using five common election myths. Volunteers read a persuasive article endorsing one of the myths. Some then received an AI-written pre-bunking article about that same false claim; others read an AI-written piece on an unrelated topic. The result was striking: "Purely AI-written pre-bunks were just as effective as pre-bunks receiving human feedback," Linegar says. "Our work showed that a short AI-generated pre-bunk protected voters from false election rumors."
The significance reaches far beyond one election. Betsy Sinclair, a political science professor at Washington University and a Caltech research affiliate, notes the stakes: "The rise of AI has meant that fiction is now easier and cheaper to produce—while verified truthful information is still expensive to produce." As the U.S. heads toward the 2026 midterms, R. Michael Alvarez of Caltech, who co-directs the Caltech/MIT Voting Technology Project, says the tool's speed is its superpower: "This is really important, because AI tools can develop countermeasures quickly, which will help us get ahead of misinformation campaigns."
Designing the template was delicate — a pre-bunk must introduce a false claim clearly enough for people to recognize it, but not so persuasively that it backfires. Now that the method is proven, the team hopes to expand it to help voters anywhere navigate an increasingly confusing media landscape. The researchers collaborated with Sander van der Linden of the University of Cambridge, a leading scholar on fighting misinformation. In a world where fiction has never been cheaper to produce, this work shows the truth can be made fast enough to catch it.