AI content and deepfakes

The visual tells are gone or going

Extra fingers, warped background text, impossible reflections, plastic skin. These were real artifacts of a particular generation of image models and they are disappearing generation by generation. Advice built on them ages badly, and worse, it teaches a habit that fails silently: you inspect an image, find no artifacts, and conclude it is real.

There is a second failure mode that gets less attention. Once people believe they can spot AI, genuine photographs start getting dismissed as fake. The damage from synthetic media is not only that false things get believed. It is that true things become deniable, and anyone caught on camera acquires a ready-made defence.

Detectors are evidence, not verdicts

AI detection tools output a probability, and they have meaningful error rates in both directions. They flag human work as synthetic and pass synthetic work as human, and their performance drops on content unlike their training data. Treating a detector score as a ruling is a category error.

Use one the way you would use a smoke alarm: as a reason to go and look, never as the finding itself.

Provenance is the check that still works

Ask where the file first appeared and who published it under their own name. A real event photographed by a real organisation has a trail: an outlet that will correct it if wrong, a photographer, a timestamp, other coverage of the same event from other angles.

For a video of a prominent person saying something explosive, the strongest single check is not visual at all. If it happened, it is news, and outlets that dislike each other would all be covering it. Silence across the board while the clip circulates only on social platforms is the loudest available signal.

Content Credentials and similar provenance metadata are genuinely useful when present and intact, because they record what tool made or edited a file. Their absence proves nothing, since ordinary uploading strips metadata from honest and dishonest files alike. One input, not a verdict.

Selective editing beats deepfakes on effectiveness

The most effective manipulated video is usually not manipulated at all. An eight-second cut of genuine, unaltered footage, starting and stopping at chosen points, can reverse a meaning completely and survive any forensic examination, because there is nothing to find.

This is far cheaper than synthesis and far more common. When a short clip is circulating without a link to the full source, that missing link is the finding. Look for the transcript or the full recording before reacting.

What the measurement actually shows

The Veriff Deepfakes Report 2026 surveyed 1,000 US adults in February 2026. Average detection of manipulated video scored 0.07 on a scale from minus 1 to 1, where zero is random guessing. One clip was correctly identified by 52 percent of respondents and another by only 30 percent. Around half believed they could reliably tell the difference.

Read that carefully: performance near chance, confidence near certainty. If you take one thing from this page, take the gap rather than the number.

Scale is the new part

Fabricating a convincing article, a spokesperson, a review or a whole outlet used to take time and money, and cost was a natural brake. It is not any more. Expect more fabricated local news sites, more invented experts and more fake reviews, all fluent and cheap.

The counter does not change and gets more important: fluency was never evidence. A claim with no identifiable, accountable source behind it is unverified no matter how well written it is, and well written is now free.