Why AI Detectors Say Your Real Photo Is AI (And What We Measured)

You took the photo yourself, and an AI detector says it was generated. It happens, and it is not rare for certain kinds of pictures. Before we added an AI check to DiffALL we measured exactly where it goes wrong, on 1,511 images — each one twice, once as the original and once as a compressed copy like the ones WhatsApp sends.

How often real pictures score high

On ordinary photographs, the false alarm rate was low. Real images landing in the “strong signs of AI” range (85 or more on our 3–97 scale):

So the risk is concentrated. A holiday snap is rarely flagged. A polished digital illustration, a 3D render or a photo pushed hard in an editor is flagged often enough that a high number alone means very little.

One of the twelve real photographs we keep as a test set — a night shot — scored 91. It is a real photo.

Why it happens

A detector does not know what a camera is. It has learned the statistical fingerprints of images that came out of generators: unnaturally clean surfaces, very even noise, smooth gradients, a particular kind of sharpness.

Real pictures can have those too:

• Digital art and 3D renders are made on computers, often with smooth brushes and perfect gradients. They share more with a generator’s output than a photo does.
• Heavy editing — noise reduction, skin smoothing, strong sharpening, HDR — removes exactly the camera grain a detector expects from a photograph.
• Phone processing stacks and smooths several frames at once, especially at night, which is why low-light shots are a known weak spot.
• Small or repeatedly compressed images leave the model less to read, so its estimate wobbles more.

The opposite mistake is common too

A low score does not clear a picture either. In the same test, about 9% of AI images scored under 25, our “no signs found” range. And when we pasted an AI-generated patch over a tenth of a real photo, the result reached “strong signs” fewer than one time in eight. Detectors look at the whole image, and a small edit barely changes the whole.

How to read a high score

1. Check what the file itself says first. Many AI tools sign their output with Content Credentials, and a real camera can sign its photos too — see how to check if an image was made with ChatGPT, Gemini or Firefly. That is evidence; a model’s estimate is not.
2. Ask what kind of picture it is. A high score on a digital painting or a heavily retouched portrait is much weaker evidence than the same score on a plain snapshot.
3. Find the original. The first upload, the camera file, the edit history. A person who really took a photo usually has the original, with its camera details still inside it.
4. Never use a score as proof. Not for grading a student, rejecting an artist’s work, or accusing anyone of lying.

Why our scale stops at 3 and 97

Because nothing here is certain. The DiffALL AI check reports an estimate from 3 to 97, never 0 or 100, and shows what the file declares separately from what the model sees, so you can tell which one a number came from. Below 25 we found no signs; from 85 the signs are strong; in between, the honest answer is “not conclusive”.

We set the “strong signs” line at 85, not lower, on purpose: at 75 more real pictures crossed it. Missing an AI image is a mistake; telling someone their real photo is fake is a worse one.

If the check got your picture wrong, the result asks whether you know the truth. Those answers are how we find the kinds of pictures it still misreads.