How AI Detects Deepfakes

When seeing is no longer believing

For most of history, a video of someone speaking was strong proof that they really said it. Not any more. AI can now generate deepfakes — fake videos, images and voices so realistic they can fool the human eye and ear. To fight back, researchers build other AI systems whose only job is to spot the fakes. This lesson explains how that detective work happens.

What a deepfake is

A deepfake is synthetic media made by AI. A model studies many photos or recordings of a real person and learns to reproduce their face, expressions or voice. It can then paste that face onto someone else's body, or make a cloned voice say a brand-new sentence the person never spoke. These tools build on the same technology as how AI image generators work and the wider world of deepfakes and fake media.

The danger is obvious: a convincing fake of a politician, a teacher or a friend could spread lies, scam people or damage reputations.

How a detector is trained

A deepfake detector is itself a machine learning model. It is trained the same way many AI systems are — by showing it thousands of labelled examples:

• A large pile of real videos and images, labelled "real".
• A large pile of fake, AI-generated ones, labelled "fake".

During training, testing and accuracy, the model gradually learns which patterns separate the two groups. Crucially, no human tells it exactly what to look for; it discovers the give-away features on its own by finding what the fakes have in common.

The clues a detector looks for

Even a flawless-looking deepfake usually leaves tiny fingerprints — too subtle for us, but visible to a model trained to hunt them:

• Unnatural blinking and eye movement. Early deepfakes often blinked too little or too regularly, because training photos rarely show closed eyes.
• Lighting and shadow mismatches. The light on a swapped-in face may not match the light in the room behind it.
• Edges and blending artefacts. Around the hairline, jaw or where the face meets the neck, the model that made the fake can leave faint smudges or warping.
• Inconsistent reflections. The glint in the eyes, or reflections on glasses, may not agree with the scene.
• Audio anomalies. Cloned voices can have unusual frequency patterns, robotic smoothness, or lips that do not quite sync with the words.

A detector weighs up dozens of such signals and outputs a probability — for example, "87% likely fake."

A worked example: thinking like a detector

Imagine a clip of a famous athlete announcing they are quitting. A detection model processes it frame by frame and notices three things: the person blinks only twice in thirty seconds, the shadow under the nose points a different way from every other shadow in the room, and the edge of the jaw shimmers slightly whenever they turn. None of these alone is proof, but together they push the model's "fake" score very high. A human reviewer is then alerted to check the source — and discovers the real athlete posted no such video. The detector did not "know" it was fake; it simply recognised a pattern it had learned from thousands of training examples, much like what is a pattern describes.

The endless arms race

Here is the catch. Every time detectors get good at spotting a flaw, the people building deepfakes fix that flaw. If detectors learn to flag poor blinking, the next generation of fakes blinks perfectly. This back-and-forth is a genuine arms race, and it means:

• No detector stays accurate forever; they must be constantly retrained on the newest fakes.
• A detector that never saw a brand-new technique can be fooled by it.
• The strongest defence combines AI detection with old-fashioned checks: Where did this come from? Does a trusted news source confirm it? Is there a verified original?

This is one reason using AI safely and responsibly matters so much.

Try it yourself! 🧪 (safe version)

You can practise the detective skills a model uses — no special software needed.

1. Find two short clips a trusted adult helps you choose: one ordinary real video and one known, clearly labelled deepfake demonstration (many are published for education).
2. Watch each closely and make a checklist: Does the person blink naturally? Do the shadows on the face match the room? Are the edges of the face and hair crisp or smudged? Do the lips match the words exactly?
3. Note which clues feel suspicious, then check your guess against the label.

You have just done by hand what a detector does at scale: hunting for tiny inconsistencies. The big lesson is to always pause and check the source before believing or sharing a striking video.