Think You Can Always Spot a Deepfake?

Almost everyone thinks they can tell. Ask around and you will hear the same confidence: the hands look wrong, the lighting is off, the voice has a tinny edge, real people blink differently. There is a comforting belief that a careful eye can still catch a fake, and that the technology, however impressive, leaves tells for those who know to look. The evidence says that belief is already out of date.

This is the quiet claim worth examining, because it shapes how people treat what they see. If you assume you would spot a fake, you lower your guard against the ones you do not spot. And the gap between felt skill and actual skill here is wide.

What the people who build it say

The most direct evidence came this month from the companies making the models. At a federal judicial panel in Boston, a Meta product director told judges that media produced by current AI models cannot be distinguished from authentic recordings by a human viewer, a point we covered in our report on the race to detect synthetic video. Not hard to distinguish. Cannot be distinguished. OpenAI's policy counsel went further, saying she doubts a reliable scientific method for detecting AI images will ever exist.

These are not doom-mongers talking up a threat. They are the vendors, testifying about their own products to a panel deciding how courts should handle evidence. When the people with the most to lose from a panic are the ones saying the fakes are undetectable, the usual skepticism runs the other way.

Where the old tells went

The familiar giveaways were real once. Early image models mangled hands, struggled with reflections and left text as garbled nonsense. Early voice clones had a flat, robotic cadence. Those flaws are mostly gone, fixed version by version, and the ones that remain are shrinking. Relying on them now is like identifying a counterfeit note by a watermark the forgers learned to copy two editions ago.

Formal detection has its own trouble. Watermarking schemes differ from company to company, there is no universal reader for them, and a watermark only helps if it survives being screenshotted, compressed and re-uploaded. Research tools that claim to flag synthetic media exist, but they trail the generators and produce both false alarms and misses. The same unreliability dogs AI text detectors, as we found when we looked at whether software can tell if a machine wrote something.

The honest version

None of this means every video is suspect or that the information world is beyond saving. It means the test has moved. You cannot verify a clip by squinting at it, so verification has to come from somewhere else: where it was published, who stands behind it, whether a trusted outlet confirms it, whether it matches other records of the same event. That is slower and less satisfying than a gut check, and it is the only method that still works.

The myth is not that deepfakes are dangerous. They are. The myth is that you, personally, would catch one. On current evidence, you probably would not, and the people who build the things are the first to say so.