How to Spot a Deepfake

Look at the boundary, not the face

The face itself is the part the model worked hardest on. The join is where it gets careless.

Hairline and jaw. Look for softness, a faint halo, or a texture change where the swapped region ends. Hair strands crossing the boundary often break or blur.

Ears. Frequently left from the original person, so they may not match the face. Mismatched ear shapes, or an earring on one side only.

Neck and skin tone. A subtle mismatch in colour or grain between face and neck is one of the most reliable tells there is.

Things that should agree and don't

Lighting. The face is lit by whatever the source footage had; the scene by its own lights. Check shadow direction under the nose and chin against shadows elsewhere in the frame.

Catchlights. The bright reflections in the eyes should match the light sources in the room, and match each other.

Sharpness. A swapped face is generated at a fixed resolution and scaled to fit, ending up softer or sharper than its surroundings.

Teeth. Still poorly handled: merged into a block, changing between frames, or unnaturally uniform.

In video, watch time rather than frames

Video gives you evidence a still cannot: consistency over time.

Blinking. Early deepfakes barely blinked. Modern ones do, but the rhythm is often mechanical — too regular, or oddly absent during speech.

Head turns. Profile views are hard. Watch what happens to the face at sharp angles or when a hand passes in front of it.

Lip sync at speed. Fast speech and plosive sounds are where sync breaks down most visibly.

Face flicker. Momentary shifts in proportion or texture between frames, most visible when the lighting changes.

Live video calls are no longer proof

This matters most, because \u201cget them on video\u201d used to be the standard advice for verifying someone. Real-time face swapping now exists and is used in hiring fraud and romance scams.

What still works on a live call: ask them to turn their head fully to the side, to pass a hand in front of their face, or to hold up a finger close to the camera. Real-time systems struggle with occlusion and extreme angles. Sudden movement often makes the mask slip visibly.

A photo of someone holding a sign with your name is also no longer evidence — that is trivially generated now.

If you are the target

Fabricated intimate imagery is the most common abuse of this technology, and being on the receiving end is not a technical problem, it is a crime.

Do not pay. In sextortion cases payment reliably leads to further demands rather than to the end of it.

Preserve everything — messages, usernames, payment requests, the files themselves. This is evidence.

Report to the platform. Every major platform has a removal process for non-consensual synthetic imagery, and they act on it.

Report to the police. In most countries this is a criminal offence, and many jurisdictions now have dedicated helplines for it.

If you need to check a specific file first, our deepfake detector analyses photos and frames, and the catfish checker covers profile photos.

Key points

• A deepfake is a composite — examine the seam, not the face.
• Ears, neck tone, catchlights and sharpness mismatches are the strongest tells.
• In video, look for inconsistency over time rather than a single bad frame.
• Live video calls no longer prove identity — ask for occlusion and profile turns.
• A photo holding a sign with your name proves nothing any more.
• If you are targeted: do not pay, preserve evidence, report to platform and police.

Frequently asked questions

Check an image yourself

Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.

More from the blog