AI-generated and altered images: how image forensics works
In shortProofVerity runs two image detectors. One asks whether a model made the pixels: generator fingerprints, diffusion artefacts, content credentials. The other asks whether someone changed them: clone stamping, splicing, inpainting, recompression, with the exact region outlined. Both report evidence, not a verdict.
Synthetic: did a model make this?
Generated images carry statistical traces of the process that made them. ProofVerity reads them from the pixels and reports a probability with the signals that produced it.
- Frequency-domain fingerprints left by upsampling and diffusion steps.
- Noise patterns that differ from what a camera sensor produces.
- Physical inconsistencies: lighting direction, reflections, text, hands and repeated textures.
- Content credentials (C2PA) and metadata, read alongside the pixel analysis, never instead of it.
Altered: did someone change this?
A real photograph can still lie. The alteration detector looks for the traces editing leaves behind and outlines the region it affects, with the share of the frame involved.
- Error level analysis: regions recompressed differently from the rest of the image.
- Clone detection: patches copied from elsewhere in the same frame.
- Splicing boundaries: edges where two sources meet, visible in noise and colour statistics.
- Inpainting: areas regenerated by a model inside an otherwise real photo.
- Double compression and resave history.
Content credentials are read alongside, not instead
C2PA credentials can prove where an image came from and what was done to it, when they are present and intact. Their absence proves nothing: most images on the internet have none. Stripped EXIF is a weak signal on its own and is reported as such.
Screenshots, memes and messaging apps
Every hop through a messaging app recompresses an image and erases some evidence. ProofVerity still detects strong signals in recompressed files, but confidence drops and the report says so. When it matters, ask the sender for the original file.
Video stills
A still exported from a video can be checked as an image. Face swaps and inpainted regions fall under the altered and synthetic detectors, with the region outlined. Results apply to the frame examined, not to the whole video.
Reading an image result
- Check which detector fired: synthetic, altered, or both.
- Look at the outlined region and its share of the frame. A 2% clone patch and a 90% generated frame are different stories.
- Read the signals and the confidence note; low resolution and recompression are stated.
- Check provenance: credentials, metadata, earliest appearance.
- Record the decision and export. The evidence appendix travels with the report.
Questions
- Can ProofVerity tell which generator made an image?
- Often the family of generator, sometimes not. When the fingerprint is recognisable it is reported as a signal, never as a certainty.
- Does image resolution matter?
- Yes. Higher resolution preserves more evidence. Below roughly 500 pixels on the short side, results are labelled low-confidence.
- Does it detect deepfake faces?
- Face swaps and inpainted faces are reported by the altered and synthetic detectors with the region outlined. Video is analysed frame by frame as stills.
- How are images billed?
- Four images cost one credit, at any resolution. A document with embedded figures is one credit as a document, up to 100 pages.