What is a deepfake?

Where the word comes from

"Deepfake" joins "deep learning", the branch of AI behind these tools, with "fake". It took off in 2017, when face-swapped videos were shared online under that name. Since then it has stretched to mean almost any AI-made or AI-altered media of a person.

Most people use it for media that shows a real, identifiable person. A generated landscape is an AI image. A generated video of a politician saying something they never said is a deepfake.

The main kinds

• Face swaps: one person's face placed onto another person's head and body, in a photo or a video.
• AI face edits: a face that was already there, changed. A new expression, an older or younger look, different eyes, a smile that wasn't there.
• Synthetic faces: people who don't exist, generated from scratch. Common on fake profiles because they don't show up in a reverse image search.
• Lip-sync and talking heads: a face animated from a single photo, or a mouth moved to match new audio.
• Reenactment: one person's expressions and head movements driving another person's face.
• Voice clones: synthetic speech in someone's voice. That's audio, which Checko doesn't check.

How they're made, roughly

Early deepfakes needed a lot of footage of the target, a powerful computer and some patience. Today many of these effects come from apps and web tools that need a single photo.

Face-swap models learn to map one face onto another while keeping the pose and expression of the original. Image generators and editors work from text instructions: describe the change and the model redraws that part of the picture. Avatar tools take a photo and an audio clip and produce a talking head. The common thread is that a model fills in pixels it predicts should be there, and that prediction leaves statistical traces even when it looks right to us.

Why they matter

Not every deepfake is harmful. Film studios use the same techniques for de-ageing and dubbing, comedians use them for satire, and some companies make consented avatars of their own staff. The harms come from using someone's likeness without consent or to deceive:

• Non-consensual intimate imagery, which most often targets women.
• Fraud and impersonation: fake profiles, romance scams, and calls or videos that appear to come from a boss or a relative.
• Misinformation: invented events, or real people shown at places they never were.
• Harassment and reputational attacks on ordinary people, not just public figures.

The other side: doubting what's genuine

Once everyone knows deepfakes exist, anyone caught on camera can claim the footage is fake. Researchers call this the liar's dividend. It is one reason careful verification matters in both directions: to catch fakes, and to stand behind genuine material with evidence rather than a hunch.

How to check a photo

Start with the source: who posted it first, and is anyone else showing the same moment? Run a reverse image search, which often finds the original photo a fake was built from. Look closely at the edges of the face, the hairline, the jaw and anything that crosses them. Check for Content Credentials. Then use a detector as one more signal.

Video deepfakes often get shared as a still, or can be checked one frame at a time. Pick a frame where the face is large, sharp and facing the camera, because that's where face analysis has the most to work with.

Checkobot checks photos and still frames with two detectors: whole-image analysis, and a face analysis built for face swaps and AI face edits. The result says AI or No AI detected, which detector fired, and why.

Deepfake detector

What a detector can't tell you

• Whether a person is who they say they are. It judges the pixels in one file, not the person.
• Which tool made the image.
• Anything about audio, live video calls, or someone holding a printed photo up to a camera.
• That an image is untouched. No AI detected means no signs were found at the threshold, and some deepfakes slip under it. Talking-head avatars from new engines are Checko's weakest area.

If you're the person in it

Save the evidence first: links, screenshots, usernames and dates. Then report it to the platform where it appears; most have specific reporting routes for impersonation and for intimate images. Don't engage with whoever posted it. For intimate images, services such as StopNCII.org can help stop them spreading on participating platforms. Laws differ by country, so if you're considering legal action, talk to a lawyer or a local support organisation.

Questions