AI-generated false content is no longer a fringe curiosity in the UK’s information ecosystem. It’s a fast-growing share of everything fact checkers investigate. In this new study, funded through the AISI Challenge Fund, Full Fact and the University of Westminster analysed 112 pieces of online AI-generated or AI-altered misinformation and disinformation that circulated between January 2025 and March 2026, collectively seen tens of millions of times. The project set out to answer a deceptively simple question: which of this content actually has the potential to cause harm?
The scale of the shift
The trajectory is striking. During this study period, confirmed or suspected AI content accounted for 24% of all fact checks published, up from just 8% a year earlier. As generative tools improve, some of the tell-tale signs we used to identify synthetic content are fading. The extra limbs and warped hands that once made fakes comparatively easy to spot are becoming increasingly rare, and audio-only fakes remain especially hard to verify with certainty. Detection is evolving too, with invisible watermarking systems like Google’s SynthID playing a growing role. In this study, watermarks provided the strongest evidence of origin, seen in 36 of the 112 cases.
Not all false content is harmful
Our 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. Rather than treating everything as equally dangerous, the model asks three sequential questions: Does the claim create a substantively false understanding (or is it only narrowly inaccurate)? Do enough people believe it to cause a specific consequence? And do those believers have the capacity and motivation to act on it?
The results show that our information environment is complex. On the one hand, 94 of the 112 entries (83.9%) created a substantively false or misleading understanding feeding a broad erosion of public trust in information. On the other, only 46 entries (41.1%) met all the criteria for substantive potential to cause specific real-world harm. The remaining 66 (58.9%) had little or no such potential. Of these, some were only narrowly inaccurate, some simply weren’t believed, and in a few cases believers had no way to act on them.
Our work did highlight a potentially fairly weak link between virality and harm. More than half of the pieces viewed over a million times had limited harm potential, while a fake video of an MP defecting to Reform UK reached a much smaller number of views yet ended up discussed in Prime Minister’s Questions. Sometimes harm depends not on reach but on whether a claim lands with the one person positioned to act on it.
A common model for harm?
Where harm potential existed, this clustered in eight fields: social unrest and vigilante violence; abuse serious enough to affect victims’ health; financial scams; harm to individual and public health; trust in police and justice; climate attitudes affecting policy; susceptibility to conspiracy theories; and shifts in broader political and social attitudes. Notably, the majority of the potential consequences identified were cumulative. Here we mean slow-burn contributions to attitudes over time rather than direct triggers of near-term action. That distinction matters for anyone deciding when intervention is urgent and when it isn’t.
Patterns in the content itself
Politics was a major focus, featuring in 63% of entries. The signature format was imposter content: often AI-generated video or audio of prominent figures included then Prime Minister Keir Starmer “announcing” policies that didn’t exist. Two recurring narratives stood out: fake announcements of unpopular charges (heating fines, pension cuts, a clean water levy) playing on cost-of-living anxiety, and fake restrictions on personal freedoms (flight limits, phone surveillance, food monitoring). The near-identical templates across these fakes potentially suggest organised or copycat operations, with context pointing to some creators motivated by financial reward rather than political conviction.
With AI fakes proliferating, 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. Our analysis shows that, more than ever, fact checkers need an approach which first asks: what’s the harm?
