Can This New Way to Fight Misinformation Work?
As AI supercharges fake news, researchers have found that telling people what’s false may be less effective than helping them figure it out for themselves
During the SARS outbreak in 2003, an American political scientist coined the term “infodemic” to describe the parallel epidemic of false information that complicated public health responses. The term caught fire in early 2020, when the World Health Organization used it to describe the global mayhem at the start of the COVID-19 pandemic.
Today, SARS and COVID are largely defanged, yet the epidemic of misinformation has gone from strength to strength. Vaccine misinformation, election fraud narratives and climate denial have each demonstrated how false content, once released into the digital ecosystem, metastasizes faster and further than corrections can follow.
It’s not a fair fight. On one side are social media platforms whose business models are based on the spread of emotionally provocative content — which false and outrage-inducing stories tend to be — rather than measured, factually accurate reporting. And that false content is more convincing than ever. Generative AI can now fabricate credible news articles, synthetic images and deepfake video at essentially zero marginal cost.
On the other side are tools that have proven largely inadequate. Warning labels and “disputed content” flags have produced mixed results at best. Research has repeatedly shown that such interventions can backfire, triggering what experts call “psychological reactance”. When people feel their beliefs are being challenged, they often entrench rather than reconsider. As a result, the warning label that is supposed to make someone doubt a false claim can, paradoxically, make them believe it more stubbornly.
Against this discouraging backdrop, a recently published experimental study offers a refreshingly different way of thinking about the problem of misinformation. What if, instead of telling people what to think, you gave them better material to think with?
Challenge rather than tell
Abayomi Baiyere of Smith School of Business, with Jan Bauer, Ioanna Constantiou and Daniel Hardt of Copenhagen Business School, conducted a series of experiments testing the use of what they term “discursive evidence”. This involved providing people with information relevant to evaluating a news claim rather than simply telling them whether the claim was true or false. They believed that by respecting the individual’s role as the decision-maker, while giving them better material to work with, they could avoid triggering a defensive reaction.
The researchers ran a series of experiments involving more than 1,000 participants. The participants were first asked to evaluate 22 news claims, a mix of true and false. In one condition, they had to judge the claim without help. In another, they were given curated evidence (from PolitiFact) relevant to assessing whether the claim was true — information they could reason with themselves.
The experiments paid off. When the provided evidence was clear, relevant and helpful, participants got significantly better at discerning true claims from false ones. Around six to seven percentage points better on average, which is impressive considering most interventions to fight misinformation produce little to no measurable effect. By contrast, irrelevant or weakly-related evidence made people more doubtful.
Critically, the improvement held even for politically-charged claims that ran against the participants’ prior beliefs. So much for the widespread assumption that people are incapable of updating their views in the face of inconvenient evidence.
The participants, however, needed to be in the right mindset. When they were primed to think critically — before the exercise, researchers had them unscramble words related to truth and deception — even lower-quality, automatically-generated evidence became substantially more effective. Neither the high-quality evidence nor the priming alone was enough; it was the combination that produced significant improvements in judgment.
Can it be scaled?
The study was conducted under controlled conditions, which raises the question: Can these findings be applied at the scale and speed that the misinformation problem demands? This is where artificial intelligence enters the picture, not as the source of the problem but as one part of the solution.
The researchers offer a potential scenario. First, a search engine or retrieval system would generate a broad pool of evidence related to a given claim circulating online. Second, a large language model would filter that pool, identifying and surfacing only the genuinely helpful items that relate to the veracity of the claim. The filtered evidence is then presented to the user alongside the content, enabling them to reason rather than simply react.
As for priming the individual for critical thinking, the researchers suggest that a brief critical-thinking prompt before someone engages with news content could be embedded unobtrusively into existing platform interfaces, similar to reCAPTCHA prompts to verify human users.
The scenario may seem implausible, but the underlying capabilities — semantic search, relevance filtering, claim matching against verified databases — exist and are improving rapidly. Google’s Fact Check Explorer does a basic version of claim matching. More sophisticated retrieval-augmented systems are being tested in newsrooms.
AI is being put to work to counter misinformation in other ways as well. Deepfake detection systems use computer vision models to identify synthetic media by spotting artifacts the human eye misses. Network analysis tools flag coordinated bot activity by detecting behavioural signatures invisible to human moderators. Content provenance systems embed cryptographic watermarks in media files, creating a verifiable trail of where content came from and whether it has been altered.
Each of these approaches represents impressive technological progress. But none resolves the fundamental problem. The economic incentives of major platforms remain aligned with the dissemination of misinformation rather than its disruption.
AI has the potential to sharpen the tools available to those defending the information environment, just as it has sharpened the tools of those attacking it. What it can’t do is resolve the underlying conditions that make the fight so asymmetric in the first place. That work belongs to regulators, institutions and, ultimately, to you and me. That’s considerably harder than building a better algorithm.