How AI Image Detection Works

Neural classifiers: pattern recognition

The first family of methods is what most people picture: a model trained on large sets of generated and real images, learning to separate them.

This works well and generalises reasonably to generators it has not seen, because diffusion models share underlying statistical behaviour. Its weakness is that it is a black box — it gives you a number without a reason, and it degrades on image types that were rare in training.

We run three separate networks rather than one. They carry the largest share of the final score, but they can be outvoted.

Frequency analysis: the shape of the noise

Convert the image into the frequency domain and you get a map of how detail is distributed across scales. Cameras and diffusion models produce measurably different maps.

A photograph shows the fingerprint of optics and a sensor: falloff away from the focal point, characteristic high-frequency noise. Generated images tend towards an unnaturally even distribution.

This is one of the more compression-resistant signals, which is why it matters for images pulled from social media.

Noise fingerprints

Every camera sensor produces a characteristic noise pattern, consistent across the frame. It is invisible and very hard to fake convincingly.

Generated images have no sensor, so any noise present is synthetic and statistically flat. This also catches deepfakes: a swapped face carries a different noise floor from the body around it.

Error Level Analysis: the compression history

Every time a JPEG is saved it loses information in a predictable way. If one region has been through a different number of saves than the rest, it shows up as a different error level.

This is the classic manipulation test — powerful for finding an edited region, and unreliable on its own, because resizing and format conversion produce similar effects. It carries a modest weight in our score for exactly that reason. It is what drives our Photoshop detector.

Metadata: cheap, decisive, and usually absent

Metadata is the only method that can end the question outright. Some generators sign their output; local Stable Diffusion often stores the prompt and seed; cameras record make, model and settings.

It is also the most fragile. Every messenger and social platform strips it, and every field can be forged. So it carries real weight when present and contributes nothing when absent — you can check it yourself for free in our EXIF viewer.

Why the combination matters

Each method above has conditions where it is blind. Neural classifiers struggle with unfamiliar styles. Frequency analysis weakens under heavy compression. ELA is confused by resizing. Metadata is usually gone.

Run them together and the blind spots stop overlapping. More importantly, you get *disagreement* — and disagreement is information. Nine methods agreeing is a different situation from one weak signal carrying the score, which is why every result shows the full breakdown rather than a single number.

It is also why an honest detector reports “inconclusive” sometimes. If the methods genuinely disagree, saying so is more useful than picking a side. Our comparison of detectors goes into how to judge that.

Key points

• Detection combines neural classifiers with forensic signal analysis.
• Frequency and noise analysis survive compression better than metadata.
• ELA finds edited regions but is unreliable alone — resizing confuses it.
• Metadata can be decisive when present and is usually absent.
• The value is in combining methods: their blind spots do not overlap.
• Disagreement between methods is information, not a malfunction.

Frequently asked questions

Check an image yourself

Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.

More from the blog