You took the picture. You know it is real. The detector scored it 71 and called it likely AI. That is a false positive, it is more common than the marketing on most detection sites suggests, and it has causes you can usually identify.
This is the failure mode worth understanding, because it is the one that hurts people. A generated picture that slips through means a check did not help. A genuine picture wrongly flagged can cost somebody a commission, a listing, or their reputation.
What the model is actually measuring
A camera sensor produces noise with a particular character. The lens, the demosaicing step that turns sensor readings into colour pixels, and the encoder that writes the file each stamp their own regularities on top. Those layered irregularities are what a capture looks like statistically.
A diffusion model never touches any of that machinery. It assembles a picture from learned statistics, and what it leaves behind is smooth and internally consistent in places a real optical chain never is.
Which means the question a detector answers is narrower than the question you asked. You wanted to know whether a camera took the picture. What comes back is how closely the surviving texture resembles a training set. Those two line up most of the time, and where they come apart is where the false positives live.
Nothing about that is fixable by a better model alone. A photograph whose fine texture has genuinely been removed no longer carries the evidence, and no amount of training recovers information that is not in the file.
The six usual causes
In rough order of how often they turn up on genuine photographs that score badly.
- Phone computational photography. Night mode, portrait mode and multi-frame merging all combine several exposures and denoise aggressively. The result is a real photograph of a real scene, assembled by software that removes exactly the sensor noise the model reads.
- Beauty filters and skin smoothing, including the ones applied automatically by a camera app before you ever see the picture.
- Upscaling. An upscaler invents plausible texture to fill in detail that was never captured, and invented texture is precisely what the model reacts to.
- Repeated re-saving. Every time a picture is recompressed it loses a little more of the fine structure. A picture forwarded through three apps has been through three rounds of that.
- Screenshots. A screenshot is a fresh file with none of the original encoding history, rendered by a compositor rather than a camera.
- Heavy noise reduction in editing, whether applied by hand or by a one-click preset.
Why the middle band exists
Scores between 45 and 65 are labelled inconclusive rather than being rounded to the nearer side. That band is where the causes above collect, and calling it either way would be dishonest.
- No AI signal 0–20
- Probably real 20–45
- Inconclusive 45–65
- Likely AI 65–90
- AI-generated 90–100
A detector that reports a confident verdict on every picture is not more accurate than one that admits uncertainty. It is the same model with the uncertainty hidden from you.
Getting a fairer reading
-
Find the original file
Not the copy from a chat app, not a screenshot, and not the version a website re-encoded on upload. Straight off the camera or phone, at full resolution. This single step moves more results than everything else combined.
-
Turn off the processing you can control
If you are shooting something you may need to defend later, use the plain photo mode. Night mode and portrait mode both work against you here.
-
Read the region map, not just the number
Smoothing tends to raise the whole frame together. If every region reads moderately high and none stands out, the pattern is consistent with processing rather than with an inserted object.
-
Keep the evidence you have
The original file, the date it was taken, and anything that places you at the scene. Provenance settles what a score cannot.
How often this happens
On the benchmark used to characterise the model, balanced accuracy is 91.3 percent under controlled conditions. Roughly one picture in eleven lands on the wrong side of the decision line even there, and real pictures off the open internet are harder than benchmark pictures.
The decision line sits at 65 rather than 50 for this reason. It deliberately trades some sensitivity for a lower rate of accusing genuine work, because the two errors do not cost the same. A missed generated picture usually means a check failed to help. A wrongly flagged real one can damage somebody who did nothing at all.