How AI Image Detectors Actually Work (Without the Jargon)

AI image detectors don't look for weird hands. They measure noise statistics, frequency patterns and physical consistency. Here's what that means.

Most explanations of AI image detection either wave at "machine learning magic" or list visual tells like six-fingered hands. Neither tells you what the software is doing. The actual mechanisms are understandable without any maths, and knowing them tells you exactly when to trust a result and when not to.

Start with how the images are made

Nearly every current image generator is a diffusion model. The training process teaches it to reverse noise: take a clear image, add random noise in steps until it is pure static, then learn to undo each step. Generate by starting from static and denoising toward something coherent, steered by your prompt.

That process leaves fingerprints, and every detection method below is a way of measuring one.

Signal 1: noise that is too well-behaved

A camera sensor produces genuinely messy noise. Photon arrival is random, sensor sites vary microscopically, read electronics add their own noise, and the result varies with ISO, temperature, and exposure. Critically, this noise is spatially uneven, different across the frame, correlated with brightness in specific ways.

Diffusion output has noise too, but it is the residue of a denoising process, not physics. It tends to be more uniform across the frame and statistically smoother than any sensor produces.

Detectors model expected sensor-noise distributions and measure deviation. This is the oldest and still one of the strongest signals, and it is why detection survives an image being stripped of metadata.

Signal 2: the frequency domain

Any image can be decomposed into frequencies: low frequencies are broad gradients, high frequencies are fine detail and sharp edges. Run a Fourier transform and you get a spectrum.

Real photographs have a characteristic falloff, natural scenes distribute energy across frequencies in a fairly consistent way, shaped by optics and the sensor's anti-aliasing filter.

Diffusion models, because they generate through repeated upsampling and convolution, frequently leave periodic regularities in the spectrum. Grid-like artefacts invisible to the eye but obvious in a frequency plot. Some architectures leave a signature distinctive enough to suggest which family of model produced an image.

This is also why heavy JPEG compression hurts detection: compression discards high-frequency information, taking part of the evidence with it.

Signal 3: physical impossibility

The most conceptually interesting category. A diffusion model has no model of the physical world, no light transport, no geometry, no anatomy. It reproduces correlations from its training data. So it gets things statistically right and physically wrong:

• Lighting direction. Shadows that imply light from two incompatible directions in one scene.
• Reflections. Mirrors, windows and eyes showing content inconsistent with the scene, or reflections whose geometry does not resolve.
• Perspective. Vanishing points that do not converge; architecture subtly out of true.
• Depth of field. Blur unrelated to distance, a background element sharp while something at the same depth is soft.
• Repeated structure. Crowds, foliage and brickwork with near-identical repeating patches.

Detectors trained on these features effectively learn what physics permits. This is also the layer your eyes can check, which is why the visual signs guide remains useful alongside software.

Signal 4: watermarks and provenance

Distinct from the above because it is deliberate. Google's SynthID embeds a signal into the pixels themselves, designed to survive cropping and re-compression. C2PA Content Credentials attach a cryptographically signed provenance record.

When either is present, it is far stronger evidence than any statistical estimate, because it was designed to be detected rather than inferred. The catch is coverage: only some tools implement them, and C2PA metadata is strippable. We covered both in Google SynthID explained and C2PA Content Credentials.

Why "AI-modified" is harder than "AI-generated"

A fully generated image has synthetic statistics everywhere. Detection is comparatively easy.

An authentic photo with one generated element, an inpainted object, a swapped face, a removed person, has genuine sensor noise across most of the frame and synthetic statistics in one region. A whole-image score averages the two and can land comfortably in the middle, reading as "probably real."

This is the practically important case. Faked insurance photos, doctored listing images, and manipulated evidence are usually edits, not generations. It is why per-region output matters more than a single number, you need to know where the anomaly is. GPTOne's AI image detector returns a per-region breakdown for exactly this reason, and we wrote about the mechanics in AI-modified image detection.

What weakens every method

Be honest about the limits, because they are real:

| Factor | Effect |

|---|---|

| Heavy compression | Destroys high-frequency evidence |

| Screenshots | Re-encodes and resamples; discards original noise |

| Small images | Fewer pixels, less statistical signal |

| Multiple platform hops | Each re-encode degrades the signal further |

| Newer model architectures | Detectors trained on older families may not recognise them |

| Photo-realistic filters on real images | Can push authentic photos toward synthetic statistics |

The last row is the false-positive mechanism worth remembering: an authentic photo through aggressive computational photography or a heavy filter can look synthetic to a detector, because both processes smooth exactly the statistics detection relies on.

What a confidence score means

A detector returns a probability, not a verdict. An 80% result means the statistics resemble the synthetic side of the training distribution at that strength, not that there is an 80% chance the image is fake in any strict sense.

Use scores as evidence weight, combined with the other checks. For anything consequential, a claim, a story, a moderation decision, a single number should never be the whole case. We unpacked this in what an AI image confidence score actually means.

The practical stack

1. Metadata and C2PA, free, occasionally decisive (how to read EXIF)
2. Reverse image search, settles provenance when the image has history
3. Pixel-level detection with per-region output, the AI image detector, which works after metadata is gone
4. Your own eyes on hands, text, reflections and repetition
5. Source verification, who sent it, and can they produce the original file?

The free tier gives you 20,000 credits with no card. Paid plans from $7.99/month cover regular verification work.

The bottom line

AI image detectors measure noise statistics, frequency-domain regularities and physical consistency, not weird hands. That is why they still work on stripped, re-compressed files, and also why compression, screenshots and heavy filters degrade them. Treat the score as one weighted input, insist on where the signal is rather than just how strong, and check the regions before you conclude anything.