How to tell if an image is AI-generated

You received a picture and you want to know whether a person photographed it or a generator made it. People who say they can tell by looking are right less often than they think, and the free "AI detector" sites rarely tell you how often they are wrong. This guide sorts the evidence by how far it can be trusted, shows what we measured when we tested it, and says plainly where every method stops. Our AI Image Checker runs the same checks in your browser.

Five kinds of evidence, strongest first

The rungs are different in kind, not just in strength. The top two are statements that a generator or its software put into the file. The third is records that a camera writes. The fourth is a guess from how the picture looks. Never let a guess outvote a statement, and never read the fifth rung ("nothing found") as "real".

What the file says

Many generators leave their settings in the file: Stable Diffusion front ends write the prompt and seed into PNG text, ComfyUI saves its whole node graph, and services such as ChatGPT, Adobe Firefly and Microsoft's Image Creator attach Content Credentials that name the tool. Google marks some images with an IPTC label saying they were edited with AI. How each of these is stored is in our guides on reading AI image prompts and on Content Credentials and watermarks.

Such a statement is the best evidence there is, with one limit: it says what the file says. A file can be edited, and a label can be copied onto a real photograph. In practice that is rare, and a statement that names a generator together with its settings is hard to fake by accident.

The catch is how seldom it is there. In our test set of 142 AI pictures, 34 (24%) carried a statement of any kind. The other 108 had been stripped, never had one, or came from a service that writes none. Facebook, X, messaging apps, screenshots and most image editors remove the metadata when they re-save a picture, so the more a picture has been passed around, the less the file can tell you.

Invisible watermarks

A watermark in the pixels survives a metadata strip. The one we can read is the mark that the Stable Diffusion reference code adds (an open-source scheme): we search for it in the picture and, when it is there, say so with certainty. It is fragile. It survives only in the unmodified output: a JPEG saved at quality 95 or lower removed most of it in our own tests, and many popular front ends never add it. Marks from Google (SynthID), OpenAI and others cannot be read outside their makers' own detectors, so for those a checker such as ours is blind, and "no watermark found" means nothing.

Camera records point the other way

A camera writes a lot that a generator has no reason to write: the make and model, the lens, the shutter time, aperture and ISO of that shot, the moment it was taken to the fraction of a second with the time zone, a block of private notes from the manufacturer, often the place. In our test, 64 of 107 real photos carried enough of these records to count as strong evidence, against 2 of 142 AI pictures, and both of those were Firefly files saved by Photoshop. So strong camera records are a good sign that a picture started as a photograph.

They are not proof. The records can be copied onto any picture, and a real photo can be changed afterward, even with AI: Apple's Clean Up tool, for example, declares the edit in the file, and our checker reports such a picture as partly AI rather than as a plain photo. A picture with no records is also common among real photos, because stripping is so widespread (43 of our 107 real photos, 40%, had too little to count).

What a detector model sees

When the file says nothing, a detector model can look at the pixels. Ours is a SigLIP 2 vision model fine-tuned by its author to separate AI pictures from photographs (Apache-2.0), shrunk to 57 MB so that it runs in your browser. We tried five public detectors on the same 249 pictures. Four scored between 0.63 and 0.72 on a standard measure of separation (AUC, where 0.5 is a coin toss and 1.0 is perfect); this one scored 0.90. Its author reports 99% accuracy on its own test pictures, and that figure does not carry over: on ours it flagged 85% of the AI pictures and also 25% of the real ones.

The chart shows where it works and where it does not. It caught every picture from eight services and a third of Firefly's, which looks photographic and is built to avoid the usual tells. On real pictures it was wrong most often for anime art and paintings: all five of our anime pictures were called AI, and half of the paintings. Drawings, graphics and illustrations are the weak spot of every detector we know of, so the checker gives no number at all for a picture that is mostly flat color.

Putting the score in context

We combine the detector with a few measurements (the picture's size and the noise in its pixels) into one score from 0 to 100, and we tested the combination by holding out whole generators and whole kinds of real picture, so the numbers below describe pictures from sources it had not seen. Held-out AUC was 0.90 for original files and 0.85 for copies that had been shrunk to 1080 pixels and re-saved as JPEG, the way a social site would. Here is what happened to the pictures in each band:

Two cautions come with these shares. First, 57% of our test pictures were AI-made, which is far more than most people meet. If you check a pile of mostly real photos, a high score is wrong more often than "one in nine". Second, the test pictures all came from Wikimedia Commons (its categories of AI-generated pictures with a named generator, and of photos taken with named cameras), which is not the same as the pictures that reach you.

Noise: a clue that fails both ways

It is tempting to read noise as a tell, and people have noticed that some AI pictures are grainy. We measured it: the noise level is the spread of what is left when a 3 × 3 median is subtracted from the picture, in the middle 1024 pixels. Real camera photos averaged about 2.0, phone photos 1.3 (phones smooth noise away), and more than half of the AI services (10 of 18) averaged under 0.5.

But the two services at the top of the chart are AI: ChatGPT pictures averaged 3.2, more than our real camera photos, and Nano Banana 2.6. So noise cuts both ways, and our blend gives it only a small weight. We cannot say why a service adds grain: it may be a deliberate texture, the way film grain is added to look natural, and a noise level cannot tell grain from a hidden code. We found no way to read a mark in it, and as far as we know none of these services has published one. If you hear that heavy noise is a watermark, the honest answer is that nobody outside the service can check it.

Size and shape

Generators work at set sizes: 1024 × 1024, 1536 × 1024, 832 × 1216 and a few dozen others, nearly always with both sides a multiple of 64. A picture that arrives at exactly one of those sizes is a clue, and it carried the most weight of any measurement in our blend. It only counts for the original: any resize, crop or re-save by a site changes it, and large pictures (a modern phone photo is 12 megapixels or more) are rarely generator output at all.

What none of this can tell you

• That a picture is real. "No sign of AI" is not "photographed". An AI picture can be stripped, re-saved and fall in the low band; 14% did.
• That a picture is fake. A score is a measure of how a picture looks to a model, and about a quarter of real pictures look AI-made to ours.
• Edits. Retouching, a swapped face or a filled-in background in a real photo is often declared by the tool, but a human edit leaves no statement.
• Video and audio. A frame grab is only a picture, and a deepfake video needs tools made for it.

How to use a result

1. Get the original file if you can, not a screenshot or a social copy.
2. Run it through the checker and read the reasons, not only the headline.
3. If the file says AI made it, you have your answer for that file. If it carries camera records, treat it as probably a photograph.
4. If it says nothing, look at the context: who posted it first, whether older or larger copies exist, and whether a reverse image search finds it elsewhere.
5. Word your conclusion carefully: "this file says it was made with X" or "a detector rates it 85 out of 100" are findings. "This is fake" and "this person lied" are not, and a detector alone cannot support them.