Imagine scrolling through your social media feed and seeing a video of the Prime Minister announcing a new tax, or a photograph of armed men attacking a hospital, or a famous financial expert encouraging you to invest in a cryptocurrency scheme. A few years ago, you might have assumed these things were real. Today, there is a very good chance they were made, or heavily altered, by artificial intelligence. This is the unsettling reality explored in a new report by Full Fact and the University of Westminster, which examined 112 pieces of AI-generated or AI-altered misinformation and disinformation seen in the UK between January 2025 and March 2026. Together, those posts, images, and videos were viewed tens of millions of times. The report’s findings are stark: 94 of the 112 entries—83.9 percent—created a substantively false or misleading understanding of the world. These were not trivial mistakes or harmless exaggerations; they were meaningful distortions that make it harder for people to know what is true. Even more concerning, 46 entries—41.1 percent—had the potential to cause or contribute to specific, real-world harm, from helping scammers drain bank accounts to fuelling street riots and damaging people’s health. The report uses a carefully designed harm-risk model, developed over four years and published by University of Westminster Press, to separate genuinely dangerous falsehoods from merely wrong ones. The model asks three essential questions: Does the claim create a substantially false understanding? Is it believed by enough people to matter? And do those who believe it have both the means and the motivation to act on it? This measured approach is vital because the problem is expanding quickly. Full Fact’s first check of AI-generated content was in early 2023; by the study period, confirmed or suspected AI claims made up nearly a quarter of everything the charity fact-checked. Barriers that once required sophisticated video-editing skills have collapsed, and convincing fakes can now be produced by anyone with a smartphone.
So what does AI-generated disinformation actually look like in practice? Politics was by far the most common theme, appearing in 71 of the 112 entries, or 63.4 percent. The most frequent format was video or audio designed to impersonate prominent politicians—most often the Prime Minister, Keir Starmer—making it appear they were announcing policies that had never been proposed. These fake announcements were carefully chosen to provoke anger: fines for using a heating system after 9pm, fees for entering airport toilets, taxes on Christmas decorations, charges for using public services people rely on. The goal was usually to feed existing anxieties about the cost of living or to create a sense that the government was quietly taking away personal freedoms. Other common themes included events in foreign countries, the wars in the Middle East and Ukraine, police and crime, migration, health and medicine, and long-running conspiracy theories. With the exception of conspiracy content, there was a strong connection between the AI fakes and stories already dominating mainstream news—disinformation tends to ride on the back of genuine public concern. The formats varied too. Some videos were entirely synthetic, presenting fake news events or conflict scenes. Others were “lip-sync deepfakes,” in which a real speech is altered and a new audio track is overlaid. Some were simple still images designed to look like press photographs, such as a picture of a politician meeting a disgraced celebrity or a leader being arrested. How did fact-checkers know AI was involved? They looked for glitches—extra fingers, unnatural pauses in speech, strange cadences—that still appear in even the most sophisticated tools. They looked for incongruities, like a politician wearing the wrong clothes compared with verified footage. They searched for negative evidence, such as the absence of any credible reporting of an event that would have left traces. They also used watermarks, which were found in 36 cases, including invisible marks from Google’s SynthID and visible labels from Grok and Sora. Detection is becoming harder as the technology improves, but these methods provide a starting point for protecting the public.
The most important part of the report is its analysis of what these falsehoods can actually do. The harm-risk model identified eight distinct fields of potential damage. First, AI disinformation can contribute to social unrest and vigilante violence. A fake image of masked men carrying axes in a Birmingham hospital, shared online, fed anti-Muslim and anti-migrant narratives and created a sense of threat that could be used to justify violence. Second, it can trigger abuse serious enough to affect a person’s mental and physical health. An AI-generated image of Arsen Ostrovsky, a Jewish man injured in the Bondi Beach mass shooting, with fake blood being applied to his face, led to antisemitic abuse, accusations that he had staged the attack, and threats against him. Third, it can cause direct financial harm. A fake “BBC” article falsely claimed that financial expert Martin Lewis had been arrested for promoting a cryptocurrency platform called Swiftgate Montark, potentially tricking people who trust him into sending money to fraudsters. Fourth, it can damage public health: the report found AI-generated videos of a real public health expert appearing to promote ineffective and potentially harmful “menopause supplements,” which could lead women to waste money or put their health at risk. Fifth, it can reduce public trust in the police and justice system. An AI-generated image of police officers kneeling before a group of Muslims was designed to portray the police as weak or sympathetic to “invaders.” Sixth, it can distort attitudes to climate policy; one fake video claimed the government was banning people from taking more than two flights a year, feeding resistance to action on climate change. Seventh, it can deepen susceptibility to conspiracy theories, such as claims that the state is tracking every phone or shopping basket. Eighth, and most broadly, it can shift political and social attitudes over time, gradually making distrust and hostility feel normal and thereby influencing elections, protests, and public debate.
This does not mean every piece of AI-generated content is dangerous, and the report is careful to avoid panic. Of the 112 entries, 66—just under 59 percent—had no or very limited potential to cause specific, substantive harm. The reasons vary. Seventeen claims were only narrowly inaccurate: an AI image of Venezuelan president Nicolás Maduro being led from a plane was fake, but he had actually been detained by US forces, so the core understanding viewers took away was essentially correct. Nineteen concerned topics where believing the claim would not lead to any consequential action. An AI image of a supercar somehow surviving the Los Angeles wildfires might be viral and striking, but what could anyone actually do with that false belief? Twenty-four claims failed because not enough people believed them to make a difference. In the case of a fake video about a cruise ship dumping waste at sea, audience comments showed widespread scepticism, with most people recognising it as AI-generated “fake news.” Finally, six claims were harmless because the people who believed them had neither the capacity nor the motivation to act: a dramatic video of Hurricane Melissa might fool viewers, but if they lived outside the storm’s path, the false belief had no practical consequence. The report uses this step-by-step reasoning to distinguish between content that is merely false and content that is likely to cause harm. That nuance is essential. Over-reacting to every AI-generated image risks stoking fear and encouraging pointless restrictions; under-reacting allows genuine dangers to go unchallenged. The key is to focus on the conditions under which false information actually leads to damage: false understanding, sufficient belief, and the practical ability to act.
Audience size is important but not decisive. Some of the most viral content in the dataset had no real potential to harm, while content with fewer views could still be risky if the right person saw it. The report found that reach alone does not determine impact. Among the 22 entries viewed at least a million times, more than half had no or limited potential for specific consequences. A fake image of protesters in Turkey gained more than 11 million views on X, but its practical effects were negligible. Conversely, a false claim that reached only a few thousand people could still be dangerous if one of those viewers was in a position to act—for example, a policymaker or a person inclined to violence. Historically, false information published in academic journals or shared with a small number of decision-makers has influenced public policy with disastrous results, as with the myth that HIV does not cause AIDS, which reached the South African president and led to catastrophic health policies. The report also explains that harms can happen in two ways. They can be direct: someone sees a scam and immediately sends money. Or they can be cumulative: a single fake video about the Prime Minister may not change anyone’s mind, but repeated exposure to the same narrative week after week can gradually normalise distrust, anger, and hostility. This cumulative effect is why the report includes content that might individually seem trivial. The scale of potential consequences also varies widely. In one case, the target was a single individual. Eight outcomes would affect individuals in general, such as scam victims or people who take dangerous health advice. Ten would affect particular communities, such as Muslims, Jews, trans people, or police officers. And 39 could affect a large part or the whole of society, by changing political attitudes, weakening trust in institutions, or damaging public health and climate policy.
The report ends with a clear set of recommendations for platforms, governments, and researchers. Social media companies should require disclosure for AI-generated or AI-edited content in sensitive areas such as politics, elections, health, crime, and international conflict. Anyone who repeatedly fails to label synthetic content should be banned from the platform after three strikes. Companies should work together to detect coordinated campaigns and to make their detection tools more transparent. Governments should support content provenance standards that allow cameras and editing tools to record how an image or video was made, and should require AI companies to add automatic watermarks to everything their tools generate. Large online platforms should have a legal duty to assess and mitigate risks to civic discourse, public security, and public health. Perhaps most importantly, independent researchers need safe, privacy-preserving access to platform data—especially during elections and high-risk periods—so that we are not reliant on the platforms themselves to police their own systems. The report’s conclusion is both sobering and measured. The threat from AI-generated disinformation is serious and growing, and the evidence shows it can contribute to harms ranging from financial fraud to social unrest. But it is not true that all AI fakes are equally dangerous. Nearly six in ten entries had little or no capacity to cause specific damage, and the report provides a credible, tested method for telling the difference. As AI tools continue to improve, the question is not whether we will see more deception, but whether we are willing to invest in the tools, policies, and public education needed to stop it from causing real-world damage. This report is a valuable starting point for that work, offering both a warning and a practical path forward.