AI deepfakes are fooling Americans as fake images improve

A recent survey by identity verification company Veriff found Americans were about as likely to mistake an AI-generated deepfake for a real image as they were to get the answer right by chance. The result shows how quickly the old giveaways, distorted fingers, unnatural eyes and other obvious visual flaws, have faded as image generators improve.

BBC coverage of the same problem showed how hard that can be in practice. In one test, psychologist Dr Clare Sutherland held up two large photographs, one of an Australian academic leading an international research study and the other an AI-generated deepfake, a setup meant to show how closely synthetic images can track the real thing.

That is why experts such as Hany Farid, a Stanford digital forensics specialist, now say he no longer trusts his own eyes when he judges images. The technology for fake videos and photographs is moving too fast for the old telltale signs to keep pace, leaving viewers with a narrower set of checks that actually work.

The sharper framework starts with provenance. Readers need to ask where an image first appeared, who posted it, whether the source can be traced back to an original account or camera file, and whether the metadata still matches the claimed time and place. Corroboration matters just as much: a striking image that appears nowhere else, or one that is posted without context, deserves far more scrutiny than a photo confirmed by multiple independent sources.

That shift is also showing up in education and policy. The University of Sheffield describes generative AI as a rapidly developing field and says its guidance is meant to support responsible and ethical use of those tools, with much of it applying to other AI systems as well. The university’s School of Information, Journalism and Communication says its research addresses media freedom, disinformation, information literacy, responsible use of AI and digital solutions to complex problems in healthcare.

The University of Sheffield and the National Council for the Training of Journalists have also hosted a symposium on integrating AI into journalism training, a sign that verification is now a core newsroom skill rather than a niche technical concern. As deepfakes get cleaner, the work of checking them is moving from spotting flaws to tracing evidence.