Imagine scrolling through your social media feed in the middle of an election season. You see a post claiming that voting machines are being hacked, or that voter rolls are full of fraudulent names. It looks alarming, it fits a narrative you’ve heard before, and within minutes it’s being shared by hundreds of people. This is the modern reality of political misinformation—a flood of false claims that moves faster than any human fact-checker can track. And with the rise of generative artificial intelligence, the problem has only gotten worse. AI can produce endless variations of convincing content in seconds, making it easier than ever to spread doubt and confusion. But what if the same technology could be turned around and used to fight back? That’s the question that a team of researchers, led by Betsy Sinclair at Washington University in St. Louis, set out to answer. Their new study, published in Royal Society Open Science, suggests that AI can indeed be used to counter election misinformation—not by chasing after falsehoods after they spread, but by “pre-bunking” them before they take hold. With more than 4,200 registered voters taking part in the experiment, the results offer some of the clearest evidence yet that AI-generated interventions can protect people’s belief in the validity of elections, even when they are later exposed to highly persuasive misinformation.
The idea behind pre-bunking is a little like a vaccine. Instead of waiting for a disease to spread and then trying to cure it, you give people a weakened version of the threat so their immune system knows what to look for. In the world of misinformation, that means exposing people to a mild version of a false claim—along with accurate, authoritative information that helps them recognize why it’s wrong—before they encounter the real thing. This isn’t a new concept. Researchers have known for years that pre-bunking can work. The problem has always been scale. To pre-bunk every rumor that emerges during an election, you’d need human experts to write custom refutations for each one, carefully tailoring the language and facts to match the specific false narrative. That takes time, expertise, and money—resources that are always in short supply during a fast-moving campaign cycle. By the time a carefully crafted fact-check is ready, the rumor has already spread across the internet. So the research team decided to test whether AI could do the heavy lifting. They recruited 4,293 registered voters through the polling firm YouGov and randomly assigned them to read about one of five common election-related rumors: voter fraud, faulty voter rolls, election hacking, the so-called “blue shift” (when late-counted ballots skew Democratic), and problems with voting machines. Before reading a persuasive article promoting the false claim, some participants were given a short AI-generated pre-bunking article that addressed the claim head-on, while others received a neutral piece about remote work and urban planning. The misinformation articles themselves were not invented from scratch; they were adapted from real articles published by Breitbart, meaning the test was as realistic and challenging as possible.
The results were striking. Among participants who did not receive the pre-bunking article, confidence in the false rumor jumped by an average of 0.43 points on a 10-point scale after reading the misinformation. In other words, the misinformation did its job—it made people more likely to believe a false claim. But among those who received the AI-generated pre-bunking article beforehand, confidence in the rumor actually decreased slightly. Overall, the pre-bunking intervention reduced immediate belief in the false claim by 0.55 points—a meaningful shift, especially in the high-stakes world of election confidence. One week later, the effect had faded to 0.17 points, but it was still statistically significant. That alone would be reason for optimism, but the study found something even more important. Pre-bunking didn’t just reduce belief in specific rumors; it also protected people’s broader trust in the election system itself. Participants who read misinformation without first receiving a pre-bunking article saw a larger decline in their confidence that votes would be accurately counted nationwide. Those who received the pre-bunking article experienced a smaller decline, with the intervention producing an immediate boost of about 0.19 points in national election confidence. Perhaps most reassuring, the researchers found no evidence of backlash. They had worried that challenging people’s beliefs might make them double down and cling even harder to false claims. That didn’t happen. In fact, the reduction in rumor belief one week after the intervention was roughly 1 point among Republicans and ideological moderates—an even larger effect than among Democrats and liberals, who generally entered the study with lower levels of belief in the rumors to begin with.
How did the researchers get AI to produce such effective interventions? The answer lies in a carefully designed process that repositions human expertise rather than eliminating it. At the start of the study, human experts developed a detailed prompt—essentially a set of instructions for the AI—that specified how to write a pre-bunking article. But they didn’t just write the prompt once and move on. They repeatedly generated AI responses, reviewed them for quality, and refined the prompt based on what they found. This back-and-forth process continued until the AI was able to produce a pre-bunking article that met their standards without needing any additional edits. Once that refined prompt was ready, it was used to generate interventions for all the other election rumors in the study. The AI was also given only authoritative election-security information from the Cybersecurity and Infrastructure Security Agency, or CISA, ensuring that the content it drew on was factual and reliable. This distinction is crucial, according to Sinclair. Human expertise was not discarded; it was moved to the front of the process. Instead of every intervention requiring a human to review every article, the humans invested their time in crafting a prompt that could then generate new interventions on demand. This means that once the framework is in place, misinformation defenses could theoretically operate at machine speed—responding to emerging rumors during an election campaign as quickly as they are generated. That’s a potentially revolutionary shift, because the current fact-checking model is fundamentally reactive. It waits for a claim to spread, then tries to debunk it. Pre-bunking at machine speed could flip that script, allowing election officials, media organizations, and civil society groups to stay ahead of false narratives instead of constantly playing catch-up.
Of course, the study’s authors are quick to caution that AI is not a magic bullet. There are important limitations that need to be acknowledged before this approach can be widely deployed. The study tested only five specific election-integrity rumors, all in the United States. It relied on authoritative information from a single federal agency, CISA, and used misinformation articles adapted from a single partisan news outlet. That means the findings might not translate perfectly to other types of misinformation, other countries, or other sources of false claims. The researchers also measured self-reported beliefs rather than actual behaviors. They asked people how confident they were in the rumors, but they didn’t measure whether people shared misinformation, changed their voting plans, or participated in protests. Belief is an important precursor to action, but it’s not the same thing. And there’s another sobering finding: while the effect on specific rumor belief lasted for at least a week, the boost in confidence in election administration itself did not persist for a full week. That suggests that a single pre-bunking intervention may not be enough to overcome deeper, long-standing skepticism about elections. Broader distrust might require repeated “booster shots”—multiple interventions over time, much like a vaccination schedule that requires follow-up doses to maintain immunity. None of these limitations mean the approach doesn’t work, but they do mean that pre-bunking should be seen as one tool in a larger toolbox, not a complete solution.
Still, the implications of this study are hard to overstate. We are living through a moment when democratic institutions are being tested by an unprecedented flood of false information. From the Brexit referendum to the 2016 U.S. presidential election to elections in India and beyond, misinformation has made it harder for ordinary people to distinguish fact from fiction and to participate in civic life with confidence. The same technology that has made this problem so much worse—generative AI—might now be harnessed to fight back. The research by Sinclair and her colleagues, with coauthors from Caltech and the University of Cambridge, offers a concrete, evidence-based path forward. It shows that with thoughtful human oversight, AI can create interventions that are effective, scalable, and safe. It shows that people from across the political spectrum can be reached without triggering the kind of backlash that many researchers fear. And it shows that the fight against misinformation doesn’t have to be hopeless. Instead of trying to debunk every false claim after it has already taken root, we can inoculate people in advance, strengthening their critical thinking skills and their resilience to manipulation. That doesn’t mean we can stop worrying about misinformation or rely on machines to save us. It means we need to be smart about how we use them. The next election is already on the horizon, and the rumors are already starting to spread. But with continued research, careful implementation, and a recognition that human judgment still matters, there is real reason to believe that we can meet the challenge. The technology driving the misinformation problem is powerful—but now we know it can be just as powerful when used to defend the truth.