Fake Image Detector

JPG, PNG or WebP, up to 10 MB. Most phone photos are HEIC by default — if yours is, export it to JPG first, or the upload will fail. Everything runs on this page: no account, no upload.

Before any pixel check runs, we look at the file's own records: camera make, editing software, save history, and any signed provenance markers (C2PA Content Credentials) still attached. A file that says "shot on iPhone, never edited" and a file that says "Midjourney" tell two completely different stories — and both are more conclusive than anything pixel analysis can produce.

Generator-fingerprint heuristics, metadata and provenance, error level analysis, and noise consistency. They run separately and can disagree with each other. When they do, we show you the disagreement instead of averaging it away.

You get one of four verdicts, the four readings behind it, and the parts we couldn't measure. Copy the whole report as text or save it as an image, if you need to show it to someone.

Every image generator leaves traces in how it builds a picture, and a lot of those traces survive in the final file. Some show up as a faint, repeating micro-pattern spread across the whole frame — an artefact of how the model decoded its output. Others sit in the frequency domain, where generated images tend to cluster differently from camera images. When the file still names a tool in its own metadata, this layer reports that name as the most likely source. Where it fails: anything that re-encodes the image damages this layer. A resize, a screenshot, a chat-app upload, or a social platform's compression will usually erase it completely. Pixel-level attribution here is heuristic, not a trained model — newer models also leave fainter, harder-to-attributable traces than the ones from two years ago.

A file can carry its own history. Camera photos often include the make and model, the shooting settings, the date and sometimes the location. Editing software writes its own tags. And a growing number of cameras and AI tools now attach signed provenance data — a tamper-evident record of what created the file. When C2PA Content Credentials or similar markers are present, we report their presence: that is the most reliable evidence on this page by a wide margin.Where it fails: most platforms strip metadata on upload, and it takes seconds to remove by hand. That cuts both ways — a missing record proves nothing at all, which is why we never treat its absence as evidence of anything. Its presence can be decisive; its absence is silence. We check that markers exist; we do not cryptographically verify signatures in the browser.

When an image is saved as JPEG, every part of it is compressed by roughly the same amount. That means every part should show roughly the same level of compression error. If one region sits at a dramatically different error level than its surroundings, something happened there after the original save — an object removed, a face pasted in, a background regenerated. Where it fails: this layer cannot detect an entirely generated image, because there is no "original" baseline to compare against; the whole frame agrees with itself. It's also easily confused by screenshots and by any image that's been saved more than once. And it's a visualisation, not a verdict — the ELA map highlights where to look, not what happened.

Camera sensors produce a characteristic grain, and every save, resize and re-compress rewrites it in a recognisable way. Genuine photographs carry this grain consistently across the frame. Generated or replaced regions usually don't — they're either unnaturally clean, or patterned differently from everything around them.Where it fails: if the entire image was generated, there's no "rest of the frame" to disagree with. Night mode, HDR, portrait blur and any aggressive noise reduction also flatten real grain, which is why "this looks too smooth" is never a verdict on its own.

The second verdict exists for one specific case: a real photograph with an AI-altered region — a person removed from a crowd, a background swapped, a sky regenerated, a frame extended with generative fill. There is no real photograph underneath that region, but there is one everywhere else, which is what Layers 3 and 4 detect. Most tools call this "AI-generated" and stop there. That's misleading — the photo itself is real, and saying otherwise starts an argument you can't win. We give it its own verdict instead. And when the whole frame is synthetic, that verdict takes priority: a generated image can show local inconsistencies too, and reporting those as "edited" would be backwards. That is why the rules are evaluated top to bottom and the first match wins.

They disagree constantly. Metadata says camera, fingerprint says generator, ELA flags a region, noise looks clean. When that happens we don't average the four readings into a comfortable number — we show you which layers fired, which stayed silent, and which had nothing to measure. If the conflict is too large to resolve, the verdict becomes INCONCLUSIVE and we tell you exactly why. A detector that always produces a confident answer isn't detecting harder. It's just hiding the disagreement from you.

A used phone, a camera, a piece of furniture — anything where you're paying a stranger on the strength of a picture. Two things go wrong here: the whole image was generated, or a real photo had its flaws painted out with AI (scratches, cracks, missing parts). Layer 3 catches the second kind better than the first, because the unedited part of the photo gives it a baseline to disagree with.

If it comes back likely AI-generated, ask for a new photo — a specific angle you choose, next to something with today's date in it. If it comes back edited, that's not automatically a scam; people clean up photos all the time. It is a reason to ask what was changed, and to keep the payment inside a platform that can reverse it.

And one thing this tool cannot tell you: a photo that passes every check can still show you a different item than the one that ships. Camera original means this file looks unmanipulated, never this listing is honest.

Before you sign a lease on a flat you've only seen photos of, book a venue, or hire someone you've never met. The failure mode here is subtler: instead of a fake object, you get a real-looking image that isn't of the place or person you're dealing with — a stock photo, someone else's storefront, a generated interior, or a genuine photo from a genuinely different address.

Detection helps with only part of that. It can flag a generated or spliced image; it can't tell you whose flat it is. So pair the result with a second request: a live video walk-through, or the same room photographed from a second angle that you specify. Ask for the angle after you see the first photo — that's the part a borrowed image can't satisfy.

A camera original reading here is worth having, and it settles less than people assume. It tells you the file wasn't faked. It doesn't tell you where it was taken.

A news photo, a chart, a screenshot from a group chat, a social card doing the rounds. This is the case where ten seconds of checking does the most good, because forwarding is what spreads it — and because most of what circulates in this form has already lost the evidence.

Read the verdicts this way. Likely AI-generated — don't forward it; if you do say something, say what the tool said ("this scored as likely synthetic") rather than what you concluded ("you've been fooled"). Edited — AI inpainting is the one people get wrong constantly: the base photograph is real, one region isn't. Say it that way — calling the whole image "AI-generated" is the part that starts an argument you can't win. Camera original doesn't make a claim true; a real photo can be real and mislabelled. And inconclusive gets treated exactly like any other unverified image: wait for a source, or just don't pass it on. Nothing bad happens if a real photo 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, most of what these four layers read is gone. If you're going to check an image at all, check the version you received — not the one you saved out and re-uploaded.