“Look for unnatural blinking” used to be the go-to advice for spotting a deepfake video. In 2026, detection companies are marketing their tools specifically around the fact that this trick no longer works, one platform’s entire pitch is built around moving past “look at the blinking” as a detection method, because deepfakes have simply gotten better at it.
Researchers Confirm What This Already Feels Like
Researchers in Aberdeen built a public test specifically to measure whether people can be trained to identify computer-generated facial images, and the honest framing behind that research says a lot on its own, this is now something that requires training to do reliably, not casual observation. Deepfakes fool most people most of the time, according to multiple 2026 analyses, not because people aren’t paying attention, but because the visual tells that used to give them away have mostly been engineered out.
Detection Software Has the Same Problem
Even purpose-built detection tools aren’t a guaranteed fix. As one expert bluntly put it, you’re never going to have a detection tool that’s able to 100% detect whether AI has been used in text, images, or video, in any form. Detection researchers building benchmarks specifically pull real deepfakes from social media rather than lab-generated test samples, because detection tools that perform well on clean lab data often fail against the messier, compressed, re-uploaded video that circulates in the real world.
Where This Actually Matters Most Right Now
Banking and financial institutions have started requiring deepfake detection specifically to combat voice cloning and synthetic identity fraud, not as a future precaution but as a current operational necessity. That alone signals how seriously institutions with real financial exposure are taking this, well past the point of treating it as a hypothetical risk.
What Actually Still Helps
- Source verification over visual inspection — checking who posted something first and whether it appears on the original source’s verified channel matters more than scrutinizing pixels
- Treating a video alone as insufficient proof — for anything with real consequences, a second, independent confirmation channel matters more than trusting your eyes
- Staying current on which platforms label AI content — several major platforms now add AI-generation labels automatically, and checking for that label is more reliable than any manual visual check
The Honest Answer
No, most people can’t reliably tell anymore, and neither can most automated tools, consistently. The shift worth making isn’t training your eye harder, it’s changing what you treat as sufficient proof in the first place.
Technology Editor. Ethan Parker is a Technology Editor at News in Focus, covering consumer technology, software, digital trends, and emerging innovations. His work focuses on making technology easier to understand by turning complex topics into practical and accessible content for everyday readers.
