AI Tools Show Promise in Fighting Misinformation, But Human Oversight Remains Essential

The same technology that floods social media with fake animal videos and AI-synthesized voices impersonating politicians is now being enlisted to fight the very misinformation it helped create. In 2024, thousands of New Hampshire voters received robocalls featuring an artificial voice of Joe Biden urging them to skip the primary election. That same year, AI-generated content—from fabricated images to doctored videos depicting violence in the Middle East—proliferated across platforms designed to harvest clicks and advertising revenue. Yet researchers at universities across North America and Europe are betting that artificial intelligence, properly constrained and supervised, could become a powerful tool for identifying and understanding false claims before they spread.

The logic is counterintuitive but compelling. Machine learning systems have long been trained to spot falsehoods by learning patterns in human-verified claims—the overuse of capital letters, exclamation points, emotionally charged language. When researchers at American University trained such a model to identify COVID-19 misinformation during the pandemic, it agreed with human fact-checkers roughly 90 percent of the time. But these older systems were brittle, trained on narrow datasets from specific time periods and platforms, leaving them useless in the messy real world. Researchers have shifted toward large language models—the same systems powering ChatGPT—which absorb vast amounts of internet text and learn the relationships between words, concepts, and contexts. These models understand human language deeply enough to analyze claims, summarize narratives, and even persuade people to abandon false beliefs.

Yet the same capability that makes them useful also makes them dangerous. Large language models are fundamentally language-imitation machines, not lie detectors. When given ambiguous or outdated information, they confidently generate false answers in a process researchers call hallucination. An early 2025 version of Grok, a chatbot designed to search the web for current information before answering, agreed with human fact-checkers only 55 percent of the time when asked to verify claims—worse than human fact-checkers agreeing with each other, which happens 64 percent of the time. The problem deepens when claims are genuinely ambiguous. The statement "Mark Carney is prime minister" is true only in Canada; a model given insufficient context might flounder or misinterpret contradictory evidence.

Researchers are developing workarounds. Dorsaf Sallami at McGill University trained her model to recognize ambiguity and ask users for clarification rather than guessing. The Dubawa fact-checking bot, launched in 2024 by a Nigerian nonprofit and accessible via WhatsApp, explicitly tells users when there is insufficient evidence for a claim instead of attempting an answer. A European collaboration called AI4Trust developed tools that analyze not what is said but how it is said—scanning for 42 common characteristics of deliberate disinformation, from allusions to secret conspiracies to emotionally manipulative language. When compared with human fact-checkers, that system agreed 70 percent of the time, enough to flag suspicious claims for journalists to investigate further.

Beyond detection, researchers are using large language models to map how misinformation spreads. Jevin West at the University of Washington and his colleagues track clusters of social media posts that propagate false narratives, watching how stories emerge, proliferate, and mutate over time. After Joe Biden's 2020 election victory, the "stop the steal" conspiracy theory spawned thousands of posts mixing outright fabrications with real but misleading information—videos of poll watchers being denied entry, for instance, without context that they were later admitted. Summarizing the larger narrative beneath such noise is challenging, but West finds that language models excel at this labeling task. When crisis managers and fact-checkers lack resources to address each individual claim, an AI system can quickly characterize the big-picture story so it can be evaluated and, if necessary, debunked.

Perhaps most striking, researchers have found that AI can actually change minds. In a 2024 study published in Science, nearly 2,200 Americans who believed conspiracy theories—that the Moon landing was faked, for instance—chatted with a version of ChatGPT instructed to persuade them otherwise. The chatbot reduced belief in the conspiracy theory by an average of 20 percent, a higher success rate than psychological interventions targeting the underlying psychology of conspiracy adherence. Thomas Costello of Carnegie Mellon University, who co-led the study, attributes this to the infinite patience of language models: they can spend whatever time and effort is necessary to conduct fact-based arguments, something human fact-checkers cannot do at scale.

Yet experts are adamant that these tools should never operate without human oversight. "We cannot just rely only on the AI," says Thanh Thi Nguyen of the University of the Sunshine Coast, comparing the challenge to raising a child—you want autonomy, but you must observe behavior and correct course. Large language models inherit the biases of the human-compiled data they were trained on, making blind trust in their outputs a recipe for the same misinformation problems they're meant to solve. Most researchers see AI primarily as a triage system: flagging content that journalists, fact-checkers, and social media platforms can then investigate thoroughly. YouTube reported removing 11,337 videos for misinformation violations in the last quarter of 2025, using a combination of automated detection and human review. Whether other platforms are deploying such tools remains unclear—TikTok, Meta, and X declined to comment—though recent court cases holding Meta and Google liable for harm to young users may prompt companies to intensify their efforts. The technology shows genuine promise, but only when humans remain in control.