← All guides Guides

Is this photo AI? A 60-second check

A photograph landed in front of you and something feels off. Here is the fastest reliable way to decide, what to look at first, and when to stop guessing.

· 8 min read · Best AI Image Detector

Run the image through a pixel-level check, then read the region map rather than the headline number. A whole-frame result answers whether the picture was generated; the map answers whether part of a real photo was replaced, which is the more common case.

Start with what you already know

Before opening any tool, spend ten seconds on the things a tool cannot see. Where did the image come from? Who sent it, and what do they gain if you believe it? Was it posted somewhere with a history, or did it appear from an account created last week?

Provenance settles more cases than pixel analysis does. A photograph that has been on a company website for three years is almost certainly what it claims to be. One that arrived in a direct message alongside a request for money deserves scepticism whatever any detector says about it.

This matters because a detector answers one narrow question: does this frame carry the statistical texture of something a model produced. It has no opinion on whether the picture shows what the sender says it shows, and those are different questions that people routinely merge.

Sixty seconds, in order
  • Ask where it came from Source and motive settle more cases than pixels do, and cost nothing.
  • Get the original file if you can A screenshot of a photo has lost most of what a check reads.
  • Run a pixel check About four seconds once the model has loaded.
  • Read the region map before the score One hot tile in a cool frame is a different finding from a uniformly warm image.
  • Check what the file declares A signed credential settles the question outright when one survives.
The order that resolves most images fastest. Each step is cheap, and each one can end the enquiry.

Get the best copy you can

The single biggest thing you control is the quality of the file you test. Detection reads texture at the pixel level, and every time an image is re-saved, resized or passed through a messaging app, some of that texture is thrown away.

A screenshot is the worst case. It replaces the original encoding entirely with a fresh one made by your own device, which is why a screenshot of a genuine photograph often scores higher than the photograph itself. If a screenshot is all you have, treat any result as weak evidence rather than an answer.

Ask for the file as it left the camera or the app that made it. In practice that means the attachment rather than the preview, the download rather than the right-click save, and the original rather than the version somebody forwarded twice.

Run the check

The analysis itself is the quickest part. A model reads the frame as a whole and returns a calibrated number from 0 to 100, then reads it again as a grid of overlapping regions and scores each one separately. Both numbers arrive together in a few seconds.

Nothing about this requires an upload. The model runs inside the browser tab, which means the picture you are checking never leaves your device, and a sensitive image stays sensitive. That matters more than it sounds when the thing you are checking is a passport, a medical scan or somebody else's private photograph.

Read the map before the number

The headline score is the part people quote and the least informative part of the result. It compresses an entire image into one figure, which works when the whole frame is synthetic and fails when it is not.

The region map is where the useful finding lives. A real photograph with one object added or removed produces a cool frame with a single hot tile, and the whole-frame score may sit comfortably in the low forties while that tile reads eighty-seven. Quoting the forty-two would be exactly wrong.

What each shape of result actually means
What you seeMost likelyWhat to do next
Low score, flat mapAn ordinary photographNothing further
High score, flat mapGenerated end to endTreat as synthetic
Low score, one hot tileA real photo with one editLook at that region
Middling score, flat mapHeavy compression or filteringAsk for the original
Middling score, scattered heatInconclusiveDo not decide on this alone

The fourth and fifth rows are the ones worth memorising. A number in the middle is not a weak yes; it is the model reporting that it cannot separate the two explanations, and the honest response is to get a better file rather than to round the number in whichever direction you already believed.

Check what the file says about itself

Separately from the pixels, an image file can carry a signed record of how it was made. Content Credentials are cryptographic rather than statistical: when one is present and valid, it is not an estimate, it is a signature.

Most images do not have one. Almost every platform strips this data on upload, so the absence of a credential tells you nothing at all. Its presence, on the other hand, can end the question in a single step, which is why the scan is worth running even though it usually finds nothing.

When sixty seconds is not enough

A quick check is right for deciding whether to trust a marketplace listing, whether a profile photo is worth a second look, or whether an image in a group chat is worth forwarding. The cost of being wrong is low and reversible.

It is the wrong instrument when the cost of being wrong lands on a person. A hiring decision, an insurance claim, a disciplinary process or anything a lawyer will read needs corroboration that a single automated score cannot supply: the original file, the account of how it was obtained, and a human who has looked at it.

The practical dividing line is whether anyone else will have to rely on your conclusion. Checking for yourself is a judgement call you can revise in a minute. Writing the answer into a file that somebody else acts on is a different act, and it needs the original file and a second pair of eyes behind it.

Questions people ask

How long does a check actually take?
About four seconds per image once the model has downloaded. The first check of a session is slower, because the model has to arrive before anything can run. After that it stays in the browser cache, so subsequent checks start immediately.
Do I need to upload the picture?
No. The model and the runtime are downloaded to your browser and the analysis happens there, so the file never leaves your device. That is the reason this approach suits identity documents, claim photographs and anything else you would not paste into a stranger's website.
What score means it is definitely AI?
None of them. The scale is a likelihood with a published decision line at 65, not a switch. A 95 is strong evidence and a 12 is strong evidence the other way, but neither is proof, and the honest reading of anything between 45 and 65 is that the model could not separate the two explanations.
Can I check an image I only have as a screenshot?
You can, and you should discount the result. A screenshot re-encodes the picture with your own device's compression, which overwrites much of the texture the model reads. Screenshots of genuine photographs frequently score higher than the originals for exactly that reason.
Does it work on images from a phone camera?
Yes, though modern phone processing pushes scores upward. Night mode, portrait blur and computational photography all rewrite pixels after capture, and a heavily processed phone photograph can land in the middle of the scale while being entirely genuine.
What if the result contradicts what I can see with my eyes?
Trust neither on its own. Visible errors in hands, text or reflections are getting rarer with each generation of model, so the absence of them proves little, and a detector can be wrong on an image that looks obviously fine. Where the two disagree, get a better copy of the file and look again.