How Do AI Image Detectors Work — and How Accurate Are They Really?

Someone pasted your work into a website. The website printed a number. Now you are arguing with a client, a moderator, a teacher or a marketplace about whether you are a fraud — on the authority of a black box that has never explained itself once.

It is worth knowing what that box actually does. Not to game it, but because once you understand what it measures, you understand why its verdicts break in exactly the cases where they matter most.

This is the mechanics spoke of our AI image detection pillar guide.

Two different things share the name "AI detector"

Before any accuracy discussion, separate them, because their reliability differs by an order of magnitude.

Everything in the rest of this article is about the third row. The first two are covered in C2PA Content Credentials explained and what SynthID is and whether it can be removed.

The five signals a classifier actually scores

A classifier does not perceive fingers, faces or plausibility. It measures texture at a level you cannot see.

1. Frequency-domain artefacts

Generative models build images through repeated upsampling. That process tends to leave faint periodic structure in the frequency spectrum — a regularity that camera optics and sensor sampling do not produce in the same way. Transform an image into the frequency domain and those grids can stand out clearly even when the picture looks flawless.

2. Noise-floor coherence

A real sensor produces noise that varies with exposure, with brightness across the frame, and per colour channel. Synthetic images are frequently either implausibly clean in shadow regions or carry noise that is too uniform to have come from photons hitting silicon.

3. Texture regularity

Hair, grass, fabric weave, gravel, skin pores, distant crowds — anything with fine repeated structure is expensive for a model to keep statistically consistent at every scale. Detectors look for texture that repeats too neatly or dissolves into mush at a particular zoom level.

4. Edge and gradient behaviour

Real lenses are imperfect. They add chromatic fringing, corner softness, vignetting and depth-of-field falloff that follows optical rules. A suspiciously perfect edge, or bokeh that does not respect a coherent focal plane, reads as synthetic.

5. Compression history

Camera JPEGs carry a plausible quantisation history. A freshly generated PNG carries none. Some detectors weight this heavily — which is also why simply re-saving a real photograph a few times can move its score.

Why accuracy claims fall apart in the real world

Vendors quote accuracy on curated benchmarks: clean generations versus clean photographs, often from the models the detector was trained against. Real images do not arrive that way.

The false-positive trap. Heavy retouching, AI denoise, AI upscaling, strong noise reduction and aggressive sharpening all strip the sensor-level evidence of authenticity. A photographer with a polished workflow produces images that look statistically synthetic — and gets accused of exactly what they did not do.

The false-negative trap. Compression, resizing, cropping and screenshotting rewrite the same statistics in the other direction. An AI image that has passed through a messaging app twice is a materially harder case than the original file.

The base-rate trap. Even a detector that is right 95% of the time is wrong 5% of the time. Run it over a hundred thousand genuine uploads and you have thousands of wrongly flagged real photographs. At platform scale, a "good" error rate is still a lot of falsely accused people.

The moving-target trap. Detectors are trained on the artefacts of yesterday's models. Every new generation of image model erases some of the tells the previous generation left behind, so detector performance decays with time unless it is constantly retrained.

Treat a classifier score as a smoke alarm, not a conviction. It tells you to look; it does not tell you what you found.

What this means if you were flagged unfairly

If a genuine photograph was labelled AI, arguing about the score is the weakest possible move. Argue with the provenance layer instead, because that layer is verifiable.

1. 1Read what your file actually declares. Open it in our Remove AI Metadata tool and read the findings list: C2PA manifests, EXIF Software strings, XMP creator tools, IPTC digital source type, PNG text chunks.
2. 2Check whether an editing step confessed on your behalf. A generative fill or AI denoise pass may have written a provenance record into a photo you genuinely shot — that record, not the pixels, is often what triggered the label.
3. 3Keep your original. Camera raw files and the untouched original with full EXIF are the strongest evidence you have that the frame came from a sensor.
4. 4Compare a before/after report. Our tool scores detection likelihood locally on both versions so you can see which signals were container metadata and which were pixel-level.

Which pixel-level factors move a heuristic score

Our own local heuristics are deliberately transparent, because a score you cannot interrogate is worthless. Broadly, these are the directions of travel:

• Missing camera metadata raises suspicion, because genuine photographs usually carry a device fingerprint.
• Perfectly round dimensions like 1024x1024 or 1536x1024 correlate strongly with model output presets.
• An unnaturally low, uniform noise floor looks synthetic; plausible per-channel variation looks photographic.
• PNG delivery of a photographic subject is unusual for camera output and common for generation tools.
• Presence of a C2PA manifest or generator tag is not a heuristic at all — it is a declaration, and it dominates everything else.

Read next: how to read your before-and-after detection report →

Practical guidance, without the fantasy

Three things are true at once, and any article that only tells you one of them is selling something.

• Statistical detection is genuinely unreliable, and you should push back hard when a single score is used against you.
• Provenance metadata is genuinely reliable, and it is the layer you can read and control yourself.
• Robust invisible watermarks are genuinely durable, and no browser tool can promise to defeat them.

Work with the layer you can verify, be honest about the layer you cannot, and stop treating a percentage from an unaccountable website as a fact about your work.