Deepfake detector

Three kinds of deepfake, three kinds of evidence

Fully generated people

A photo of a person who does not exist, or a real person rendered by an image model in a scene they were never in. These leave the strongest pixel traces, because every pixel came from a generator. GPTTrace’s classifier handles them well when the image is reasonably large.

Face swaps and re-animation

A real video or photo with only the face replaced or its expression changed. Most of the frame is genuine camera footage, so whole-image checks see a mix. Look closely at the boundary of the face — jawline, hairline and ears — where the swapped region is blended in, and at moments when something passes in front of the face.

Lip-sync and voice clones

A real video of a person with new words put in their mouth, usually with a cloned voice. The face changes only around the mouth. Detection depends heavily on the audio, which is why we built an AI voice detector alongside this one.

How the video check works

Video files are large, so GPTTrace does not process every frame. It reads the container first — MP4 and MOV metadata, encoder strings and any Content Credentials from tools like Adobe or OpenAI’s Sora. Then it samples eight frames spread through the clip and runs each one through the same neural classifier and forensic checks used for photos. You see a thumbnail strip with a score per frame; a deepfake that only affects part of a clip can show up as a few high frames among low ones.

Compression is the main enemy. Platforms like TikTok, Instagram and WhatsApp re-encode video at low bitrates, which smooths away much of the fine texture forensic tools rely on. A clip downloaded from the original uploader will give a far more reliable result than a screen recording of a repost.

Visual signs worth checking yourself

• Blinking that is too regular, or eyes that don’t track what the person is looking at.
• Teeth that look like a single white block, or change shape from frame to frame.
• Skin that stays perfectly smooth while the rest of the image is grainy.
• Earrings, glasses and hair strands that flicker, merge or vanish when the head turns.
• Lighting on the face that doesn’t match the room — a shadow on the wrong side, missing reflections in glasses.
• Audio and lip movement drifting out of sync, especially on “p”, “b” and “m” sounds.

None of these is proof on its own, and the best fakes show none of them. Treat them as reasons to verify through another channel.