Deepfake Detector

Photos: JPG, PNG, WebP, up to 25 MB. Video: MP4, MOV, WebM, up to 100 MB and 3 minutes. Audio: MP3, WAV, M4A, up to 25 MB. HEIC (the default iPhone photo format) isn't supported directly — export it to JPG first. Files are read in this tab only; nothing is uploaded.

Before any scoring runs, we look at what the file says about itself: its format, its encoder clues, its compression history, and any provenance metadata attached to it. A file that has been re-saved five times and screenshotted twice carries a different story than one that came straight off a camera — and that story changes how much weight the score's reading deserves.

For faces and video we look at regional error (blend-boundary proxy), noise consistency and frame-to-frame stability across sampled frames. For audio we look at breathing gaps, room tone flatness and prosody stability. Each signal gets its own reading, because a single number hides which part of the file actually looks wrong.

You get one of three verdicts, the individual signals behind it, and a plain-language line telling you what to do next. Copy the report as text, or save it as an image, if you need to pass it on.

When one face is pasted onto another, the two images have to be merged somewhere: the jawline, the hair edge, the neck, the boundary where a cheek meets a collar. That merge leaves a seam — a strip a few pixels wide where sharpness, colour temperature or noise level shifts in a way a real photograph wouldn't produce. Modern swapping tools blur this seam deliberately, so we look for the blur itself: an unnaturally soft band in an otherwise sharp image is a signal, not an absence of one. On this page that shows up as regional error inconsistency in the ELA-style pass. Where it fails: heavy compression smears real detail too, so on a re-uploaded WhatsApp clip the seam and the compression artefact can look identical.

A real face sits inside a real light. The direction of the key light shows up in the highlights on the forehead, in the shadow under the nose, and in the reflection in both eyes — and all of those have to point the same way. Swapped faces often carry their own lighting from the source photo, so the highlights land slightly off. The browser checks we run pick this up indirectly through regional inconsistency. Where it fails: flat, front-on lighting removes most of the cues. A face shot in an evenly lit room gives this check almost nothing to work with.

Real skin has pores, fine hair, uneven pigmentation and specular highlights that break up across the surface. Generative models reproduce the average of those textures but tend to make them too uniform — too smooth in the mid-face, with a repeating micro-pattern where real skin would be irregular. Some generators also leave a faint recurring texture pattern across the whole frame. Where it fails: beauty filters, portrait-mode blur and noise reduction all erase the same textures, which is why "this looks too smooth" is not by itself a verdict.

Every camera sensor produces a characteristic noise pattern, and every editing operation disturbs it. When a region of an image has been generated or replaced, the noise in that region stops matching its surroundings — it's either too clean or patterned differently. We check how consistent the noise is across the frame and whether any region stands out from the rest. Where it fails: if the whole image was generated, there is no "original" noise to compare against, and this check has to lean on other residue instead.

Some cameras and some AI tools now write signed provenance data into the file, recording what created it. When that data is present and intact, it is the most reliable thing on this page — far stronger than any pixel analysis. We report whether C2PA-style markers are present; we do not cryptographically verify signatures in the browser. Where it fails: most platforms strip metadata on upload, and anyone motivated can remove it. A missing provenance record proves nothing either way; its presence can be decisive.

A human speaker breathes, and those breaths land in the recording. A human also speaks inside a room, and that room has a noise floor. Synthesised speech often has neither — the breaths are missing, or inserted at regular intervals, and the noise floor either vanishes during speech or stays perfectly flat across it. On top of that we look at prosody stability: whether zero-crossing behaviour holds unnaturally steady across the clip. Where it fails: a short clip, a phone recording, or a heavy noise-reduction pass can remove exactly the evidence we're looking for.

Almost every deepfake you'll actually encounter has been through a platform: uploaded to a messenger, downloaded, re-uploaded, screenshotted. Each pass strips high-frequency detail and rewrites the noise — the two things most of the checks above depend on. This is the single biggest reason detection results vary. Where possible, work from the earliest version of the file you can get, and treat any result on a twice-forwarded clip as low-confidence by default.

The checks that apply to this file came back clean, and the confidence is high enough for us to say so. It means we didn't find evidence of manipulation — it doesn't mean the file is provably genuine, because a file can be genuine in ways no pixel analysis can confirm.

Two or more independent checks flagged the same thing, and they agree. This is the strongest verdict we issue, and it's still a probability, not a proof.

This means the file doesn't carry enough usable evidence: it's been compressed too hard, it's too short, or the signals conflict — or confidence fell below our fixed publish threshold. You get the individual readings anyway. Our guidance is the same every time: get a better copy of the file, from as close to the original as possible, and run it again.

Because it wouldn't mean anything. Detection accuracy depends far more on the state of the file than on the detector: the same tool can be highly reliable on an original export and close to useless on a screenshot of a forwarded video. Any site quoting one flat percentage for all inputs is quoting a lab benchmark, not your file. Confidence here is shown as a two-decimal figure such as 0.91, never as fake-precision percent.

Upload the profile picture. If it comes back inconclusive, your next move is a video call — not an argument about the photo. A live call is far harder to fake than any still image, and it settles the question faster than any detector.

Record a segment, if they consent, and check it afterwards. Look at the result alongside what you already know: did the connection "glitch" whenever they moved, did the audio ever desync, was the lighting on their face doing something the room didn't explain? A likely-synthetic reading on a recorded segment is worth more than your impression of it — but the live call is still the best test you have, because a swap that holds up in a recording usually doesn't hold up when you ask them to turn their head, pass a hand across their face, or hold up three fingers and wait for the count.

Ten seconds before you pass it on. If it comes back likely synthetic, don't forward it — and if you do decide to say something, say what the detector said rather than what you concluded: "this scored as likely synthetic" is a claim you can stand behind, "you're being fooled" is an argument you'll lose. If it comes back inconclusive, treat it exactly like any other unverified clip: wait for a source you trust, or simply don't forward it. Nothing bad happens if a real video doesn't get forwarded today.

One more thing worth knowing: sharing is what destroys the evidence. Every forward re-encodes the file, and by the third platform it's been through, the signals this tool reads are mostly gone. If you're going to check a clip at all, check the version you received — not the one you saved and re-uploaded.