Guide
How AI Image Detection Works
A plain-English explainer of the forensic techniques behind tools like TrueFace.
AI image detectors look for the tiny fingerprints that generative models, beauty filters and deepfake pipelines leave behind. The human eye misses them, but they are statistically loud — once you know where to look.
1. Frequency-domain analysis
Real cameras record light through a lens, a sensor and a JPEG compressor. That chain produces a very specific frequency signature — small high-frequency noise spread evenly across the image. Generative models (diffusion, GANs) tend to over-smooth the high frequencies and leave periodic artefacts behind. A Fourier transform of the photo makes those artefacts visible as unnatural peaks or grids.
2. GAN-inversion and reconstruction error
For AI-generated faces, detectors attempt to project the photo back into the latent space of a known generator. If the model can reconstruct the face almost perfectly with a small latent vector, the photo is statistically close to something that generator could produce. Real photos resist this inversion — the reconstruction error stays high.
3. Filter and retouching detection
Beauty filters smooth skin, enlarge eyes, shift skin tone and slim facial geometry. These edits leave measurable traces: unnaturally low skin-texture variance, symmetry that exceeds natural human range, and warped local geometry around the jaw and nose. TrueFace measures these per face, not per image, so a group photo where only one person is filtered is flagged correctly.
4. Deepfake-specific cues
Face-swap and lip-sync deepfakes blend a generated face onto a real body. Detectors look at the seams — colour shifts between face and neck, mismatched compression noise inside vs outside the face region, inconsistent lighting direction and missing micro-expressions around the eyes.
5. Why no detector is 100% accurate
Generative models improve every month, and a heavy re-compression (screenshot → social platform → screenshot again) wipes out many of the artefacts detectors rely on. A responsible AI image detector returns a probability, not a verdict, and is best used together with context (where did the photo come from? does the metadata survive?). TrueFace treats every result as a forensic signal — not a courtroom proof.
Try it yourself
Upload any portrait in the TrueFace scanner and you will see a per-face breakdown of the signals above, with a colour-coded verdict. Open the scanner.
Frequently asked questions
How do AI image detectors work?
They analyse frequency artefacts, GAN-inversion error, filter traces and deepfake seams left by generative models. Each signal is measured per face and combined into a probability, not a binary verdict.
Can AI detectors always tell if a photo is AI-generated?
No. Heavy re-compression (screenshot → social platform → screenshot again) can erase the artefacts detectors rely on, and each new generation of models leaves fewer traces. A responsible detector reports a probability plus context.
What is GAN inversion?
GAN inversion projects a photo back into the latent space of a known generative model. If the model reconstructs the face almost perfectly with a small latent vector, the image is statistically close to something that generator could produce.
Can beauty filters be detected in a photo?
Yes. Filters leave measurable traces: unnaturally low skin-texture variance, warped local geometry around the jaw and nose, and symmetry that exceeds the natural human range. TrueFace measures these per face, so a group photo where only one person is filtered is flagged correctly.
How accurate is TrueFace?
TrueFace returns a probability between 0 and 100% per face along with the underlying signals. On clean, uncompressed portraits detection is highly reliable; heavily compressed or edited images are reported with lower confidence.