Not long ago, AI-generated falsehoods felt like a curiosity—a weird picture of the Pope in a puffer jacket, a synthetic voice saying something outrageous, a video that was clearly too strange to be real. That moment has passed. In the United Kingdom, AI content has moved from the margins to the mainstream of misinformation, and it now occupies a growing share of everything fact-checkers have to investigate. A new study, carried out by Full Fact and the University of Westminster and funded through the AISI Challenge Fund, set out to understand this shift. Researchers analysed 112 pieces of online misinformation and disinformation that were either entirely generated by AI or meaningfully altered by it, circulating between January 2025 and March 2026. These items were collectively seen tens of millions of times. But the project was not simply counting fakes. It was trying to answer a deceptively simple question: which of this content actually has the potential to cause harm? That question matters because we are drowning in a sea of synthetic content, but not all of it is equally dangerous. Some AI fakes are silly, some are simply wrong, and some are genuinely dangerous. Treating them all the same way—either panicking about every piece or ignoring every piece—is a recipe for failure. The study is an attempt to bring calm, careful judgment to a problem that usually inspires fear and confusion, and its findings are more nuanced than the usual headlines suggest.
The scale of the shift is striking. During the study period, confirmed or suspected AI content accounted for 24 percent of all fact checks published by Full Fact—a jump from just 8 percent a year earlier. That is not a slow creep; it is a dramatic change in the information environment. And it is likely to keep growing. The technology itself is becoming harder to spot. The extra limbs, the warped hands, the strangely blurry backgrounds that once made AI-generated images comparatively easy to identify are becoming far less common. Audio-only fakes remain especially difficult to verify with certainty, because the human ear is not well equipped to catch subtle synthetic artifacts, and there are often no obvious technical clues left behind. Detection is evolving too, but it is an arms race. One of the more promising developments is invisible watermarking, such as Google’s SynthID, which embeds a signal into content that can later be detected by software. In this study, watermarks provided the strongest evidence of origin, appearing in 36 of the 112 cases. Still, watermarking is not universal, and many pieces of content are created without any such marker. Fact-checkers increasingly have to make judgment calls based on context, source behavior, and the content’s internal consistency rather than relying on obvious visual glitches. This makes the work slower, harder, and more dependent on careful reasoning. The study reflects that reality: it treated AI involvement as a spectrum, distinguishing between content that was clearly AI-generated, content that was probably AI-generated, and content that was altered with AI tools. The authors were honest about uncertainty, and that honesty is itself an important lesson. In a world where people want simple answers, the truth is often that we cannot be sure exactly how a piece of media was made.
But the bigger question is not simply whether something is fake; it is whether the fake can hurt someone. The study applied a harm-risk assessment model developed over four years by the academic Peter Cunliffe-Jones and published by University of Westminster Press in 2025. This model refuses to treat all falsehoods as equally dangerous. Instead, it asks three sequential questions. First, does the claim create a substantively false understanding of the world, or is it only narrowly inaccurate? A manipulated video that makes a politician say something they never said is substantively false, while a video that slightly exaggerates a statistic may be misleading but not fundamentally untrue. Second, do enough people believe the claim to produce a specific consequence? A lie that no one believes is less concerning than a lie that spreads widely and takes root. Third, do the believers have both the capacity and the motivation to act on it? A false claim that angers people is less dangerous if those people have no power to do anything about it; a false claim that reaches a single person with decision-making authority can be very dangerous indeed. Applying this model to the 112 cases produced a complex picture. On one hand, 94 of the 112 entries—83.9 percent—created a substantively false or misleading understanding, feeding a broader erosion of public trust in information. That is a serious problem in itself. On the other hand, only 46 entries, or 41.1 percent, met all three criteria and were judged to have substantive potential to cause specific real-world harm. The remaining 66—58.9 percent—had little or no such potential. Some were only narrowly inaccurate. Some were simply not believed. And in a few cases, even people who believed them had no realistic way to act on that belief. This is not to say that the harmless-seeming content did not matter; it still contributed to a fog of confusion and distrust. But the study is a reminder that not every false claim is an emergency, and fact-checkers and regulators cannot afford to treat them all as if they were.
One of the most interesting findings in the study is the surprisingly weak link between virality and harm. Conventional wisdom suggests that the more views a piece of content gets, the more dangerous it must be. The data challenges that assumption. More than half of the pieces that were viewed over a million times had limited harm potential. They were widely seen, but they did not meet the full criteria for causing specific, concrete damage. Conversely, some content that reached a much smaller audience still managed to create real consequences. The study gives a vivid example: a fake video of an MP defecting to the Reform UK party. It was not among the most viral pieces in the dataset. Yet it ended up being discussed in Prime Minister’s Questions, meaning it had reached the one institution with the power to amplify it into the national political conversation. This is a crucial insight. Harm sometimes depends not on the size of the audience but on whether the claim lands with the right person—someone who can act on it, repeat it, or make decisions based on it. In an era of algorithmic amplification, we tend to assume that danger scales with views. But the reality is more subtle. A falsehood that spreads through millions of feeds may remain relatively inert, while a falsehood that reaches a single influential person can ripple outward in unpredictable ways. This matters for how we allocate resources. If we focus only on the most viral content, we will miss the quiet but consequential fakes that slip under the radar. The study suggests that fact-checkers need to think like investigators, not just like analysts of social media trends, and ask who might be watching as well as how many people are watching.
Where harm potential did exist, it clustered in eight distinct fields. The first was social unrest and vigilante violence—false claims that could lead people to take justice into their own hands. The second was abuse serious enough to affect victims’ health, including harassment and stalking driven by false accusations. The third was financial scams, where AI-generated content was used to trick people out of their money. The fourth was harm to individual and public health, such as dangerous medical advice. The fifth was trust in the police and the justice system, where AI fakes could undermine confidence in institutions that rely on public cooperation. The sixth was climate attitudes affecting policy, meaning false claims that could shift public opinion and, in turn, influence whether governments take meaningful action on climate change. The seventh was susceptibility to conspiracy theories, which can lead people down a path of rejecting expert information and mainstream institutions. And the eighth was shifts in broader political and social attitudes—slow changes in how people see their society, their leaders, and one another. Notably, the majority of these potential consequences were cumulative rather than immediate. They were not direct triggers of near-term action but slow-burn contributions to attitudes over time. This distinction matters enormously for anyone deciding when intervention is urgent and when it is not. A piece of content that directly incites violence today is clearly an emergency. A piece of content that quietly normalises distrust in elections or vaccines may be just as dangerous in the long run, but it requires a different kind of response. The study does not argue that cumulative harm is less important; it argues that we need to be honest about the difference between a spark and a slow burn. Otherwise, we will exhaust ourselves putting out small fires while the embers continue to smolder beneath the surface.
Finally, the study looked at the patterns in the content itself. Politics was a major focus, featuring in 63 percent of the entries. The signature format was imposter content: AI-generated video or audio of prominent figures, often including the then Prime Minister, Keir Starmer, “announcing” policies that did not exist. Two recurring narratives stood out. The first was fake announcements of unpopular charges—heating fines, pension cuts, a clean water levy—that played on cost-of-living anxiety. These were designed to tap into genuine public fears about money and survival. The second was fake restrictions on personal freedoms, such as flight limits, phone surveillance, and food monitoring. These played on fears of government overreach and control. What made these fakes especially worrying was not just their content but their structure. The near-identical templates across different pieces suggested that they were not random one-off efforts but part of organised or copycat operations. The context also pointed to some creators being motivated by financial reward rather than political conviction; in other words, some of these fakes were simply grift, designed to attract clicks, engagement, and ad revenue. This is an important reminder that misinformation is not always about ideology. Sometimes it is just a business. As AI fakes continue to proliferate, the ability to distinguish genuinely dangerous content from noisy-but-inconsequential content is becoming essential. Treating everything as an emergency wastes finite fact-checking and regulatory capacity, while ignoring the genuinely harmful risks repeating the mistakes that preceded past crises. The study’s central message is that fact-checkers, policymakers, and the public need a more disciplined approach. Before we share, before we panic, before we regulate, we should ask a simple question: what’s the harm? That question is not a way of dismissing concerns. It is a way of taking them seriously enough to know which ones deserve immediate action and which ones require long-term strategy. In a world increasingly saturated with synthetic media, that kind of calm, careful judgment is not just useful. It is essential.