Imagine you are scrolling through Facebook and come across a post that makes your stomach drop: “HPV vaccines increase the risk of neurological problems.” You have a child in the recommended age range for the vaccine, and suddenly that alarming headline sticks in your mind. Then you see a reply from an artificial intelligence account dedicated to fact-checking. It says the claim is false. Do you change your mind? According to new research led by Washington State University, your answer may depend less on whether the correction is accurate and more on how the AI speaks to you—and what you believe about AI in the first place. In an age of rampant medical misinformation online, AI can be a powerful tool for correcting falsehoods. But the researchers found that tone is everything. People who view AI strictly as a technical tool are more persuaded by a neutral, just-the-facts correction. People who believe AI can be humanlike respond better to an empathetic, understanding tone. Perhaps most surprisingly, the source of the correction—whether it came from a human being or an AI agent—did not really matter. What mattered was whether the tone of the correction matched the person’s expectations about how AI should communicate. This insight has big implications for social media platforms, health agencies, and anyone trying to fight misinformation in an era when trust is hard to earn and easy to lose.
The study, published in the International Journal of Human-Computer Interaction, was led by Porismita Borah, a professor in Washington State University’s Edward R. Murrow College of Communications, along with colleagues and students from WSU, the University of Wisconsin-Madison, and Hong Kong Shue Yan University. The research team conducted a randomized online experiment with 857 parents whose children were in the age range recommended to receive the HPV vaccine. The human papillomavirus, or HPV, is spread through sexual contact and can cause a variety of serious health problems, including several types of cancer. The HPV vaccine is widely considered safe and effective, yet it has been a frequent target of misinformation. The parents were first evaluated for their level of “anthropomorphism”—the tendency to assign human characteristics to non-human things like machines, animals, or AI. Then they were shown a simulated Facebook comment thread that began with a false claim: “HPV vaccines increase the risk of neurological problems.” In the thread, an AI corrections account responded to the claim. Some participants saw a neutral reply in direct, plain language: “That’s not true. Scientific studies have shown no link between HPV vaccines and any of those scary neurological conditions.” Others saw a warmer, empathetic response: “I hear you, but scientific studies have shown….” The results were clear: the correction was most effective at reducing misperceptions when the tone matched the participant’s anthropomorphism beliefs. People who expected a machine-like AI were more convinced by the neutral reply, while people who saw AI as humanlike were more persuaded by empathy. The findings add a crucial layer to our understanding of persuasion and misinformation correction, showing that one-size-fits-all communication is not enough.
Why does tone carry so much weight? Because corrections are not just about facts—they are about feelings. When someone tells you that something you believe is wrong, it can feel like a challenge to your identity, your intelligence, or your values. People may become defensive, double down on their original belief, or dismiss the source of the correction altogether. A blunt “that’s not true” might work for someone who sees information as data and wants the facts delivered quickly and cleanly. But for someone who expects a conversation with a human-like guide, a cold correction can feel dismissive and disrespectful. Empathy, on the other hand, can soften the blow. It signals that the speaker understands where you are coming from, that they are not just correcting you but also hearing you. Yet empathy is not a universal magic bullet. In fact, prior research on whether empathetic corrections work has produced conflicting conclusions. Borah’s team added an important twist: the effectiveness of empathy depends on the receiver’s expectations. If you believe AI is just a machine, then an AI saying “I hear you” might feel phony or manipulative. But if you believe AI is capable of understanding human emotions, the same phrase feels natural and reassuring. In other words, communication is not just about the message and the messenger. It is about the relationship between the message, the messenger, and the person receiving it. That is what makes misinformation correction so complicated—and why scientists are working to understand not just what works, but for whom it works.
This study comes at a critical moment. Social media has fueled the rapid spread of dubious health claims, and misinformation about vaccines has become a public health challenge. The HPV vaccine is a perfect example: it is safe, effective, and can prevent cancers, yet misinformation about its risks has led many parents to delay or refuse it. The false claim used in the study—“HPV vaccines increase the risk of neurological problems”—is exactly the kind of scary, misleading statement that circulates widely online. What makes this finding particularly relevant is that the correction came from an AI agent, not a human fact-checker. Many people are skeptical of AI’s ability to handle nuanced social situations, and some worry that corrections from machines will be ignored or rejected. But Borah’s research suggests otherwise. In multiple studies, her team has found that corrections can work most of the time, regardless of whether they come from a human or an AI. What matters more is how the correction is delivered. “In this study, it did not necessarily matter whether the correction came from a human being or an AI agent,” Borah said. “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.” For those who see AI strictly as a tool, a neutral, matter-of-fact correction is persuasive because it matches their mental model. For those who see AI as more humanlike, a little warmth and empathy go a long way.
These findings have practical applications that could help shape the future of online fact-checking. Social media platforms, government health agencies, news organizations, and even individual creators could design AI fact-checking systems that adapt their tone to each user. The idea is simple: during an onboarding process, a platform could ask users a few quick questions to gauge how much they anthropomorphize AI and other non-human entities. Based on their responses, the AI correction agent could adjust its conversational style—staying neutral and data-driven for one person, while using a warmer and more understanding tone for another. This is not about tricking people or manipulating them; it is about meeting people where they are and communicating in a way they can actually hear. It is the digital equivalent of knowing when to use a firm handshake and when to offer a gentle pat on the back. Of course, the authors acknowledge that this is not the whole solution. Race, gender, culture, and other factors also shape how people respond to corrections. Misinformation is a multifaceted problem, and no single strategy will solve it. But tailoring tone to someone’s expectations is a promising starting point. Instead of broadcasting the same correction to everyone, AI could be taught to listen—figuratively speaking—and respond in a way that respects the user’s worldview. In a noisy digital world, being right is not enough. You also have to be seen as someone (or something) worth listening to.
Ultimately, this research reminds us that fighting misinformation is not just a technical challenge—it is a deeply human one. People are not simple information-processing machines. We bring our emotions, our identities, our biases, and our assumptions into every interaction, including online conversations with AI. The same fact that convinces one person might offend another, simply because of how it is phrased. The same source that gains trust in one context might be dismissed in another. That is why Borah and her colleagues are pushing beyond the question of whether corrections work, toward the more nuanced question of how they work in different contexts. As Borah put it, “The problem of misinformation is critical, and it’s not going away. The effectiveness of corrections depends on a lot of factors—for example the way you talk to someone when providing accurate information—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.” In that sense, the study is not just about AI or vaccines or social media. It is about the art of communication itself. It is a reminder that the best way to change someone’s mind is not necessarily to shout louder or pile on more facts, but to speak in a tone that matches who they are and how they see the world. If AI can learn to do that, it may become one of our most valuable allies in the ongoing fight against misinformation—not because it is smart, but because it can be understanding. And in a world where truth is often drowned out by noise, a little understanding can make all the difference.