AI ‘Pre-Bunking’ Tools Could Neutralize Election Misinformation, Caltech Study Finds

New research shows large language models can generate effective warning messages at lightning speed, giving voters a psychological vaccine against false claims.

As the 2026 midterm elections approach, a novel weapon has emerged in the fight against digital misinformation. Researchers at the California Institute of Technology (Caltech) have developed an artificial intelligence tool that can automatically generate “pre-bunking” articles—psychological vaccines that expose and defuse false claims before they can spread. Published in the journal Royal Society Open Science, the study demonstrates that these AI-crafted warnings are just as effective as those written by human experts, offering a scalable, rapid-response solution to an information landscape saturated with AI-generated deepfakes and viral rumors. The findings signal a fundamental shift in how democracies might combat the spread of false narratives.

The urgency stems from the sheer velocity of online falsehoods. In recent election cycles, viral social media posts, manipulated images, and deceptive narratives have often reached millions within hours, leaving fact-checkers scrambling to catch up. The rise of generative AI has exacerbated this crisis, drastically lowering the cost and effort required to produce convincing misinformation. “False claims can spread widely before fact-checkers have time to respond,” explains Mitchell Linegar, the study’s lead author and a postdoctoral scholar at Washington University in St. Louis. Traditional reactive fact-checking, while essential, is resource-intensive and chronically slow, often debunking a myth long after it has taken root in the public consciousness. A single viral rumor can dominate the news cycle for days before a credible source debunks it.

The roots of this crisis run deep. The 2020 US presidential election saw an unprecedented wave of false claims regarding mail-in ballots, voting machines, and polling place integrity. These narratives, often amplified by political elites and foreign actors, have persisted and mutated into new forms for the 2024 and 2026 cycles. Research consistently shows that exposure to such misinformation reduces trust in the electoral process, potentially depressing voter turnout and eroding the legitimacy of democratic outcomes. This makes the task of protecting voters before they are exposed to these falsehoods all the more critical.

This is where pre-bunking differs fundamentally. Rooted in psychological inoculation theory, pre-bunking works by exposing individuals to a weakened dose of a false claim before they encounter it in the wild, along with a clear explanation of the manipulation techniques being used. This builds cognitive resistance, allowing the audience to recognize and reject the misinformation when they see it. “Pre-bunking gives people accurate information before exposure, making them less likely to believe these claims, potentially stopping their spread before it starts,” Linegar notes. The technique is analogous to a medical vaccine: a small, weakened sample of the virus triggers the immune system to build defenses. In a media context, this means exposing the logical fallacies and rhetorical tricks used by purveyors of misinformation. However, crafting these interventions manually is a laborious task. Each piece requires a human expert to carefully distill the fallacy, articulate the truth, and frame it persuasively—a process that takes hours if not days, making it impossible to keep up with the constant flow of new rumors.

The Caltech team sought to automate this process using large language models (LLMs). Led by R. Michael Alvarez, the Flintridge Foundation Professor of Political and Computational Social Science at Caltech, the team built a framework that uses a “reusable human expert–crafted prompt.” This prompt provides the AI with guidelines on how to correctly structure a pre-bunking article, while feeding it verified election information. “By combining a reusable human expert–crafted prompt with verified election information, the model could generate pre-bunking articles for new rumors quickly and without the need for further human review,” the researchers report. The key innovation is the separation of human expertise from the repetitive generation process. Human experts essentially “teach” the model the structural rules of an effective pre-bunk once, and then the model can apply those rules indefinitely to new topics.

To test the efficacy of these AI-generated pre-bunks, the team conducted a large-scale randomized controlled trial involving more than 4,000 registered voters prior to the 2024 US election. Participants were randomly assigned to read a persuasive, human-written article endorsing one of five common election myths—ranging from false claims about widespread voter fraud to misconceptions about ballot handling, machine tampering, and election official corruption. Some of these participants then received an AI-generated pre-bunking article specifically addressing that myth, while others received a control article on an unrelated subject. The researchers then measured participants’ beliefs in the myths, their confidence in true election facts, and their overall trust in election integrity. The random assignment ensures that the results are due to the intervention itself and not pre-existing differences between the groups.

The results were striking. The AI-generated pre-bunks were just as effective as human-written interventions at correcting misconceptions. “Purely AI pre-bunks were just as effective as those receiving human feedback,” Linegar stated, highlighting the model’s ability to match human nuance. Even more importantly, the benefits persisted. “Our work showed that a short AI-generated pre-bunk protected voters from false election rumors,” he added. The effects were still measurable a week later, suggesting a durable shift in the participants’ beliefs. Furthermore, the impact of the interventions was consistent across party lines, indicating that this tool can be effective in a highly polarized environment. This cross-partisan efficacy is particularly significant, given that many previous misinformation interventions have shown limited effects with strong partisans.

This cross-partisan resilience is a remarkable finding. In political psychology, the ‘backfire effect’ often causes individuals to double down on their beliefs when presented with contradictory information. However, pre-bunking focuses on the techniques of manipulation rather than the partisan conclusion. By showing a voter how a viral claim uses a specific fallacy—such as scapegoating or emotional language—the intervention avoids triggering partisan defensiveness. The study’s success suggests that this technique-based approach can bypass the psychological barriers that typically hinder fact-checking in polarized environments.

The interdisciplinary nature of the team was key to the project’s success. Alvarez, a leading authority on election technology, provided the domain expertise on electoral processes. Betsy Sinclair, chair of political science at Washington University and a research affiliate at Caltech’s Linde Center for Science, Society, and Policy, contributed her vast knowledge of political behavior and survey methodology. Sander van der Linden, a professor of social psychology at the University of Cambridge and the world’s foremost expert on pre-bunking, guided the psychological foundations. “I think this is one of the ways Caltech has demonstrated tremendous intellectual leadership on one of the most pressing problems of our age,” Sinclair noted, praising the collaboration. The combination of political science, computational methods, and social psychology was essential to designing a robust study that could withstand rigorous scrutiny.

Despite the remarkable success, generating the pre-bunks was not a simple “plug-and-play” process. A significant challenge was programming the AI to explain a false claim without inadvertently spreading it. “A pre-bunking article has to introduce a false claim clearly enough for people to understand it without being persuasive enough that the intervention causes harm,” Linegar explained. “Traditionally, a human expert does this for each new rumor.” The team had to iterate extensively to refine the prompt. “It took a lot of iteration to get the model to distill and weaken the claims without repeating them verbatim,” he said. This delicate balance between education and accidental endorsement was eventually solved, but it underscores the complexity of automating persuasive communication. The researchers had to carefully test multiple prompt variations, analyze the outputs for any residual persuasive effects, and adjust the wording to ensure the AI maintained a neutral, educational tone.

To translate this research into real-world impact, the team has released a public demonstration of the tool. Accessible at electionbot.chat, the application allows anyone to input a new claim and receive a draft pre-bunking article within seconds. “To facilitate real-world use, the team has released a public demonstration of the tool that drafts pre-bunks using trusted factual material,” the study authors stated. This open-access approach democratizes the technology, allowing state and local election officials, journalists, and civic organizations to generate tailored warnings without needing a team of data scientists. The tool is designed to be user-friendly, requiring only that the user submits the problematic claim; the system automatically fetches related factual context and constructs the intervention.

The public demo is intuitive. A user pastes a rumored claim, such as “Voting machines in county X are switching votes,” and the system immediately generates a clear, fact-based response. The response typically explains why the claim is false, citing official sources, and outlines the common rhetorical tricks used to make it seem credible. The interface also allows users to specify the platform or format they need, ensuring the output is suitable for a press release, a social media caption, or a voter information flyer. This ease of use is crucial for state and county election departments, which often lack dedicated communications staff.

The team is already looking beyond long-form articles. “We are exploring shorter formats, such as social media posts and videos, and applications beyond elections,” Linegar said. Given that misinformation primarily spreads through short-form content on platforms like X (formerly Twitter), TikTok, and Facebook, adapting the AI framework to these formats is a critical next step. The goal is to create a suite of easily shareable graphics, videos, and short texts that can be deployed rapidly across the same channels where the rumors are spreading. “The goal remains the same: help accurate information move as quickly as the misinformation it competes with,” Linegar added. This adaptability could make the tool relevant for public health campaigns, climate change communications, and other domains plagued by misinformation.

As the 2026 midterm elections loom, the timing could not be better. Alvarez emphasized the growing threat of intentional disenfranchisement. “The AI-assisted tools will be particularly helpful in responding to attempts to decrease confidence as we head into the 2026 midterm elections, as deliberate misinformation efforts seeking to disenfranchise voters is a major concern,” he said. The speed of AI generation directly addresses the previous bottleneck. “AI tools can develop countermeasures quickly, which will help us get ahead of misinformation campaigns,” he added. This proactive capability transforms the dynamic from a losing race against fact-checkers to a strategic advantage for truth. Election officials, often overwhelmed during voting periods, can use this tool to quickly address localized rumors that might otherwise suppress turnout.

The human toll of misinformation is often borne by frontline election workers. “Across the country there are myriad hard-working election officials who are answering phone call after phone call, email after email, addressing questions that have arisen because of the rise of election rumors,” Sinclair observed. These officials spend hours each day calming voters fears about baseless claims of hacked machines or fraudulent ballots. “Tools like the one we have developed are going to make their jobs easier, while also making it easier for voters to access accurate information,” she added. By automating the initial drafting of responses, the tool frees up valuable human time for higher-level tasks, such as engaging with communities and ensuring the smooth administration of the vote.

While the results are promising, the researchers are careful to note the limitations. The study was conducted in a controlled online setting, and the long-term effects beyond a week were not measured. Future research will need to investigate how pre-bunks perform when embedded in real social media feeds, where users are exposed to a cacophony of competing messages. Additionally, the study focused on specific election myths; it remains to be seen how the model performs on highly novel, unprecedented falsehoods. However, the flexibility of the LLM framework suggests it can be rapidly updated with new fact-checking data. The team is currently working on integrating real-time fact-check feeds to ensure the pre-bunks remain accurate and current.

However, deploying such powerful tools carries ethical responsibilities. Just as AI can generate pre-bunks, it can also generate more sophisticated misinformation. The researchers emphasize that their tool is built on a foundation of verified facts, sourced from official election administrators and reputable fact-checking organizations. They are also exploring ways to watermark the AI-generated content to distinguish it from human-created material. Furthermore, they urge organizations using the tool to maintain human oversight, ensuring that the final messages are reviewed for accuracy and tone before dissemination. The goal is not to automate persuasion for propaganda purposes, but to arm democratic institutions with a defensive capability.

The study represents a significant step forward in the defense of democratic integrity. While it does not replace the crucial work of human fact-checkers, it supercharges their efforts by automating the labor-intensive process of drafting intervention materials. Though AI has been weaponized to create fabricated realities, this research demonstrates that the same technology can be repurposed to inoculate the public against them. With a scalable, rapidly deployable toolkit now available, the balance of power in the information war may be shifting back toward the truth. The findings serve as a powerful reminder that innovation, guided by rigorous social science, can offer a real defense against the corrosive effects of digital manipulation. As the next election cycle approaches, the availability of an automated pre-bunking system provides a powerful shield for voters, ensuring they are better equipped to separate fact from fiction at the ballot box.