You’ve probably seen them: absurd videos of animals that don’t exist, fake audio clips of world leaders saying things they never said, and social media posts full of outrage that were written by no human at all. Artificial intelligence has made it staggeringly easy to create convincing falsehoods, and the internet is drowning in them. Fake accounts spew propaganda, content farms churn out clickbait, and in 2024, voters in New Hampshire received robocalls with an AI-synthvoice of President Joe Biden telling them not to vote. It’s enough to make you feel like you can’t trust anything you see online. So it might seem strange, even counterintuitive, that scientists are now exploring ways to use the very same technology to fight back against misinformation. Why would anyone trust AI to solve a problem that AI itself created? But here’s the twist: the same language skills that let AI generate convincing nonsense also make it uniquely good at spotting, analyzing, and even correcting false information. Researchers are finding that machine learning and large language models—the engines behind chatbots like ChatGPT—can help us identify fake news, understand why it spreads, and maybe even reduce our belief in it. In a world where online lies have influenced elections, fueled violence, and deepened political divides, these tools could be invaluable. Across the globe, most people see online misinformation as a major threat to their country, according to Pew Research Center. The challenge is huge, but experts argue we have to fight fire with fire. As Jevin West, a misinformation expert at the University of Washington, puts it, “We should fight fire with fire.”
For a long time, scientists have used machine learning to detect fake news. The basic idea is simple: you train a computer model on a huge pile of claims that human fact-checkers have already labeled true or false. The model looks for patterns in the language that tend to show up in falsehoods—maybe an overuse of capital letters, exclamation points, or highly emotional wording—and then uses those signals to judge new claims. During the Covid-19 pandemic, researchers trained a model to identify tweets full of coronavirus misinformation, and it agreed with human fact-checkers about 90 percent of the time. That sounds impressive, but there’s a catch. These models are usually trained on carefully curated datasets that only cover specific time periods, topics, or platforms. So they don’t work well in the messy, constantly changing real world where new events, new slang, and new conspiracy theories pop up every day. That’s why researchers have turned to large language models, which are trained on massive amounts of public internet text and can understand not just words, but the relationships between words, phrases, and concepts. They’re better at understanding human language in all its nuance, and many newer versions can also analyze images and audio. Ask them whether a statement is true or false, and they generate an answer based on the patterns they’ve absorbed from millions of books, articles, and websites. But there’s a serious problem: LLMs are not lie detectors. They’re more like language imitation machines. When they don’t have enough information or a claim is ambiguous, they have a well-known tendency to confidently make things up—a behavior called hallucination. They can also be out-of-date, because they aren’t always trained on the latest news. So if you ask an AI chatbot about something that happened yesterday, it might give you a fluent, confident answer that is entirely wrong.
Researchers are working hard to fix this weakness. Dorsaf Sallami, an AI researcher in Montreal, has developed a fact-checking browser extension that lets an LLM search the web for current information before it answers a question. That’s a step in the right direction, but it isn’t bulletproof. One preliminary study found that an early 2025 version of Elon Musk’s chatbot Grok only agreed with human fact-checkers about 55 percent of the time when asked to verify claims. To put that in perspective, human fact-checkers agree with each other about 64 percent of the time. So even people who are trained to check facts often disagree. The task of deciding what’s true is genuinely hard, especially when a claim depends on context. For example, the statement “Mark Carney is prime minister” is true in Canada but false almost everywhere else. Researchers are teaching AI to recognize this kind of ambiguity. Instead of forcing the model to instantly answer yes or no, they train it to ask the user for more information or to say there isn’t enough evidence. That’s exactly what the Dubawa fact-checking bot does in Nigeria. People can message it on WhatsApp, and it checks their claims against articles from reputable media sources. If it can’t find supporting evidence, the bot says so plainly, rather than inventing an answer. Human journalists then step in to investigate thoroughly and publish their findings. This kind of cautious, human-in-the-loop approach is becoming the gold standard for AI fact-checking.
Other projects use AI not to decide if a claim is true, but to look at how it’s being told. A European collaboration called AI4Trust developed tools to fight disinformation—fake news that is deliberately created and spread. One tool prompts an LLM to sniff out 42 common features of disinformation, like alluding to a secret group of conspirators or using emotionally manipulative language. When they compared this AI tool with human fact-checkers, they agreed about 70 percent of the time. That’s not perfect, but it’s good enough to help journalists spot suspicious claims that might deserve a closer look. Social media companies are also getting in on the act, at least to some extent. YouTube says it combines advanced detection systems with human reviewers to remove videos that violate its misinformation policies, and it reports taking down thousands of videos in just one quarter. Other major platforms like TikTok, Meta, and X sometimes stay quiet about their use of AI detection, and experts suspect they aren’t doing enough. In recent years, many social media companies pulled back from moderating content, citing free speech concerns. But West thinks the tide may be turning, especially after court cases in California and New Mexico found that Meta and Google were liable for harming young users, which could push companies to take misinformation more seriously again. There’s also the growing question of legal liability for AI-generated falsehoods: a German court recently ruled against Google over an AI overview that produced false information. So there are real legal and financial reasons for tech companies to start caring about cleaning up their platforms.
AI can do more than just flag individual false claims; it can also help us see the bigger picture of how lies spread. West and his colleagues are using LLMs to track whole clusters of social media posts that promote misleading or false narratives, and to observe how those stories emerge, morph, and evolve over time. Take the “stop the steal” conspiracy theory that spread after Joe Biden was elected president in 2020. There were thousands of posts, some with outright false allegations of voter fraud, others with real but misleading information, like a video showing poll watchers initially being denied access to a polling station without explaining that they were later let in. Trying to summarize the overall narrative from all those individual posts is exhausting and nearly impossible for humans alone. But LLMs are fairly good at labeling the big-picture narrative, which helps crisis managers and fact-checkers understand what they’re dealing with. If you can address the large-scale story, not just one tweet at a time, you can debunk misinformation much more efficiently. And there’s even more exciting evidence that AI can actually change people’s minds. In a 2024 study published in Science, researchers had 2,190 Americans who believed in conspiracy theories—like the idea that the Moon landing was a hoax—chat with a version of ChatGPT that was instructed to persuade them otherwise. The chatbot succeeded in reducing people’s belief in these theories by about 20 percent on average. That’s a bigger effect than most other interventions, like psychological therapy designed to address the feelings that drive conspiracy beliefs. There was a bit of a hiccup: Science later told readers about some issues with the study’s data, and the authors submitted corrected data, saying the original results still hold. But the basic finding is encouraging. As one researcher noted, AI chatbots are incredibly patient and good at using reason and evidence to talk someone out of a belief, one conversation at a time.
Even with all this promise, experts are careful to say that AI should never replace human judgment. These tools are biased, just like the data they were trained on, and they can make confident mistakes. If we blindly trust AI to tell us what’s true, we’re setting ourselves up for more misinformation—just in a fancier package. The most realistic vision is a partnership: AI sifts through the endless flood of falsehoods, flags suspicious content, and offers context, while human fact-checkers and journalists investigate, verify, and make the final calls. Social media companies need human oversight too, and so do the researchers and organizations using AI to fight misinformation. Thanh Thi Nguyen, an AI researcher in Australia, likes to compare training an AI model to raising a child. You want the child to be independent, but you still have to watch their behavior, guide them, and correct them when they go astray. The same goes for AI. It can be an amazing ally in the battle against fake news, maybe even essential—but only if we keep humans in the loop, stay skeptical, and always remember that artificial intelligence is not a magic truth machine. It’s a powerful tool that reflects both the best and worst of human communication, and it’s up to us to use it wisely. In a time when the truth feels more fragile than ever, that might be the most important lesson of all.