AI detector photo analysis — how detection actually works
Detectors do not recognise “AI”. They measure a handful of physical and documentary signals, then weigh them against each other.
The five signals a detector can look at
Every credible detector builds its answer from a small set of measurements, and each measurement has a different failure mode. Metadata and provenance records are documentary evidence: they say what the file claims about its own origin, and a signed credential is among the strongest evidence you can get — but most cameras and most generators still attach nothing at all. Watermarks work the same way in reverse: a hit is meaningful because the mark had to be written at generation time, while a miss tells you only that this particular provider did not mark this particular file.
The remaining signals are statistical. Frequency-domain analysis looks for the regular patterns that upsampling leaves behind, because generators typically build an image at low resolution and scale it up, which introduces periodic structure a camera never produces. Noise-residual analysis asks a different question: does the fine-grained texture behave like sensor noise, which is stronger in the shadows and differs from frame to frame? A generated image often has noise that is too uniform, or noise that sits on top of the image instead of inside it. Finally, semantic checks look for physical impossibilities — a reflection that does not match its subject, a shadow pointing the wrong way, text that dissolves into shapes.
| Signal | What it measures | Reliability |
|---|---|---|
| Metadata and provenance | EXIF blocks, XMP packets and signed C2PA content credentials — who or what claims to have made the file. | Strong when present and signed; meaningless when absent |
| Watermarks | Invisible marks embedded at generation time by a specific provider, readable only by that provider's verifier. | Decisive on a positive hit; a negative result proves nothing |
| Frequency-domain fingerprints | Periodic patterns left by upsampling and deconvolution, which show up as regular peaks in the frequency spectrum. | Moderate — degrades with resizing, cropping and re-compression |
| Noise residuals | Whether the fine-grained noise in the image behaves like sensor noise, which is stronger in shadows and varies per frame. | Moderate — denoising and HDR stacking erase the same signal |
| Semantic and physical consistency | Whether shadows, reflections, hands, text and geometry obey the physics of the scene they appear in. | Weakest and most subjective — needs a human reviewer |
Why real photographs get flagged
False positives are not a bug in one tool; they follow from what the tools measure. A photograph that has been denoised loses exactly the noise residual a detector looks for. A long exposure smooths water and clouds until the texture statistics resemble a generated render. Focus stacking, HDR merging and heavy noise reduction all push a genuine photograph toward the statistical profile of a synthetic one. Add a screenshot or a re-export that strips metadata, and a real photo can arrive looking as bare as a generator output.
The opposite error is just as common. A generated image that has been resized, cropped and re-compressed a few times loses the frequency fingerprints that gave it away, and if its metadata was cleared along the way, a metadata-based check finds nothing to report. This is why a single score is such a poor basis for a decision, and why the honest answer to “is this AI?” is usually a list of observations with a confidence level attached, not a verdict.
A detector is not a verdict. It reports what its signals found. Use it to decide where to look next — the file's provenance, the person who sent it, the original capture — not to close the question.
What this means when you receive a file
Work from the strongest evidence down. Ask for the original file rather than a re-export, because the original still carries the metadata and credentials that an export has lost. If the file claims a camera, check that the claim is internally consistent — make, model, lens and exposure should describe a real combination. Then look at the weaker statistical signals as supporting context. If you want to see what a specific file carries before you form a view, you can inspect its metadata in your browser without sending it to anyone, or compare the available detector options and their trade-offs first.