AI Fact-Checking Tone Key to Correcting Health Misinformation, WSU Study Finds

PULLMAN, Wash. — As bogus health claims continue to spread through social media, artificial intelligence has emerged as a promising tool for correcting falsehoods—but new research led by Washington State University shows that the effectiveness of AI fact-checking depends heavily on the tone of the correction, and that tone must match the beliefs of the person being corrected. The study, published in the International Journal of Human-Computer Interaction, found that people who view artificial intelligence strictly as a technical, machinelike tool are most persuaded by corrections delivered in a neutral, just-the-facts style. By contrast, people who believe AI can be humanlike—tending to assign human traits to machines—respond far better to corrections framed in an empathetic, understanding voice. Interestingly, researchers found that whether the correction came from a human being or an AI agent had little effect on whether people Updated their beliefs. The emotional tone of the message, they said, mattered more than its apparent source. “In multiple studies, our team has found that corrections could work most times, but the tone of the correction is important,” said Porismita Borah, professor in WSU’s Edward R. Murrow College of Communications and corresponding author of the new publication. “In this study, it did not necessarily matter whether the correction came from a human being or an AI agent. What mattered was the tone and how the tone aligned with people’s beliefs about whether AI agents should be more humanlike or more machinelike.” The findings come at a critical time when health misinformation online has become a major public health concern, from false claims about vaccines to dangerous miracle cures, and the research suggests that simply deploying AI to fact-check may not be enough. The messenger may not matter as much as the emotional register of the message itself.

.

The study adds to a growing body of evidence that correcting misinformation is far more complicated than simply presenting accurate facts. It takes more than data to persuade someone that a claim is false; people may feel challenged, insulted, or threatened when their beliefs are contradicted, and they often react defensively. They may also question the trustworthiness of the person or system delivering the correction. As social media has become a primary source of news and health information for millions of people, scholars have been urgently trying to identify which correction strategies work best. Some previous research has suggested that empathetic corrections are better at reducing misperceptions, while other studies have found neutral, factual corrections to be more effective. Borah’s team sought to resolve these conflicting findings by introducing a new variable: anthropomorphism—the tendency to attribute human characteristics, emotions, and intentions to non-human entities such as AI agents. They reasoned that people’s expectations about how an AI should communicate would shape how receptive they are to its corrections. If a person sees AI as merely an algorithm optimized for accuracy, a warm, emotional correction may feel out of place or even manipulative. If a person sees AI as more humanlike—capable of understanding and feeling—a cold, clinical correction may fail to connect. The study’s results confirmed this intuition, showing that alignment between tone and anthropomorphic belief was key to persuasion.

.

To test their theory, Borah and her colleagues conducted a randomized online experiment involving 857 parents of children who were in the age range recommended to receive the human papillomavirus vaccine. The vaccine, commonly known as HPV vaccine, is recommended for preteens and teens to protect against strains of the virus that can cause several types of cancer, including cervical, anal, and oropharyngeal cancers. The vaccine has been proven safe and effective, but it has also been the target of persistent online misinformation, making it a fitting case for studying how corrections work in the messy, emotionally charged world of health communication. Survey participants were first evaluated for their level of anthropomorphism—that is, how strongly they tended to see AI systems as humanlike versus as purely technical tools. Then they were shown a simulated Facebook comment thread beginning with a false claim: “HPV vaccines increase the risk of neurological problems.” An AI correction account then responded to the claim within the thread. The corrections came in two distinct styles. The neutral, fact-based response used direct, plain language:”That’s not true. Scientific studies have shown no link between HPV vaccines and any of those scary neurological conditions.” The empathetic version used a warmer, more understanding tone:”I hear you, but scientific studies have shown….” The researchers measured how likely participants were to accept the correction and reduce their misperceptions about the HPV vaccine after seeing the exchange. The results were clear: corrections were most effective at reducing misperceptions when the tone matched a respondent’s anthropomorphism beliefs. People with low anthropomorphism—who viewed AI as a machine—were more persuaded by neutral, clinical language. People with high anthropomorphism—who viewed AI as humanlike—were more persuaded by empathetic, understanding language. The source of the correction, human or AI, did not significantly change the outcome; tone was the decisive factor.

.

These findings have practical implications for social media platforms, government health agencies, news organizations, and other institutions engaged in the fight against health misinformation. Instead of using a one-size-fits-all correction bot, they could design AI fact-checking agents that automatically tailor their conversational tone to each user’s level of anthropomorphism. For example, a platform could include a simple onboarding step that asks users about their attitudes toward AI or assesses their anthropomorphic tendencies, then adjusts the AI’s communication style accordingly. If a user expresses that AI can be trusted like a human friend, the correction bot might respond with empathy and acknowledgment; if a user sees AI as a simple search tool, the bot might use direct, no-nonsense language. The study’s authors note that this kind of targeted approach could make AI fact-checking significantly more effective without requiring more complex interventions. Borah’s co-authors were Ziyao Zhang, a PhD student at Washington State University; Xiaohui Cao, a PhD student at the University of Wisconsin-Madison; and Danielle Ka Lai Lee, an assistant professor at Hong Kong Shue Yan University. Their work is among the first to demonstrate that the effectiveness of empathetic versus neutral corrections is conditional on individual differences in how people perceive technology—a finding that could shape the next generation of automated misinformation-fighting tools.

.

The broader lesson from the study is that persuasion is not just about information—it is about relationship and expectation. As Borah noted, the problem of misinformation is critical and is not going away, especially in the digital landscape where falsehoods travel faster than facts. “The effectiveness of corrections depends on a lot of factors—for example, the way you talk to someone when providing accurate information,” she said. “An empathetic tone may often work better than a condescending one. Race, gender, and other factors also matter. We’re ultimately trying to study humans—and humans are remarkably complex.” The research underscores that automated fact-checkers, no matter how accurate, must be sensitive to the psychological state of the person receiving the correction. A correction that feels judgmental or dismissive can backfire, entrenching people deeper into their beliefs. A correction that acknowledges the person’s concern—even if it ultimately rejects the factual claim—can lower defenses and open the door to accepting evidence. As AI becomes increasingly embedded in social media platforms and public health communication, understanding how to make these systems feel trustworthy and appropriate to diverse users will be essential. This study offers a roadmap for doing just that: not by replacing human judgment with algorithms, but by programming algorithms to speak with the right emotional register, tailored to the person at the other end of the screen. It also suggests that future research should continue to explore how other social identities, cultural contexts, and personality traits interact with tone, source, and content of corrections—because the battle against misinformation will require more than bots armed with facts; it will require bots that can speak to humans in the way humans need to hear them. In the meantime, health officials and platform designers can take away a simple but powerful lesson: when AI corrects a false health claim, how it says something can matter just as much as what it says.