Technical documentation

Detection Methodology

The system verifies what can be proven, learns the remaining signals, and refuses to guess outside its calibrated range.

What the score means

The image score represents calibrated AI evidence only when the response says that calibration is available. It is not proof that the depicted event is true or false. The separate decision can be AI-generated, AI-edited, likely authentic, or Uncertain.

Uncertain is a first-class result. It is returned when models disagree, an input is outside the calibration distribution, compatible model artifacts are unavailable, or neither error-controlled threshold is crossed.

Image analysis pipeline

1

Cryptographic provenance

C2PA Content Credentials are parsed and cryptographically validated. Only a valid credential from a trusted signer can make an authoritative origin decision. Unsigned EXIF, software names and missing metadata remain contextual and can never prove either class.

2

Semantic forensic expert

A frozen CLIP-family vision encoder analyzes the full image and several spatial crops. A small detector head learns transferable synthetic-image evidence without retraining the general visual representation. Crop dispersion is retained for the separately trained AI-edit localization decision.

3

Low-level learned expert

A second model receives multiscale spectral distributions, neighboring-pixel relationships and local residual representations. None of these dimensions is interpreted using a hand-written AI threshold. They are learned jointly from content- and format-matched data.

4

Learned fusion and calibration

Expert outputs and disagreement are combined by a model trained only on out-of-fold predictions. Isotonic calibration is fitted on a separate generator/source split. Final real and AI thresholds are selected to meet declared false-negative and false-positive targets on that calibration set.

5

Distribution-shift rejection

Each trained expert stores the feature distribution of its training data. Inputs beyond its calibrated distance are marked out-of-distribution and cannot receive a decisive passive-model verdict.

Training and evaluation safeguards

  • Training, calibration and test sets are separated by generator or real-image source, never by random image rows.
  • The benchmark rejects byte-identical overlap with a supplied training manifest.
  • Real and synthetic examples must be matched for subject matter, file format, resolution and post-processing so models cannot learn shortcuts such as PNG versus JPEG.
  • Results report coverage, selective accuracy, false-positive and false-negative rates, calibration error and group-bootstrap confidence intervals.
  • Performance is broken down by generator and transformation, including resizing, JPEG/WebP encoding, screenshots and social-media laundering.

Known limitations

No universal passive detector exists. New generators, aggressive re-encoding and deliberate attacks can erase or imitate learned signals.

Provenance is positive evidence, not an absence test. Screenshots and platforms can remove C2PA. A missing credential does not imply authenticity.

Unsigned metadata is spoofable. Camera EXIF and generator names are displayed for context but do not control the verdict.

Calibration has a domain. Reported error targets apply to the documented calibration data, not to every possible future generator or image category.

Transparency

Every response states which experts were available, whether their outputs were calibrated, whether the image was out-of-distribution, whether the system abstained, and what provenance was actually verified.

View benchmark results