Product methodology

How Detectiks evaluates an image

Detectiks gives an estimate about an image, not a certificate of origin. This page explains the actual signal groups in the detector and a repeatable way to use its result.

Try a local scan

What the pipeline checks

The server pipeline preprocesses the file, evaluates enabled modules, then combines their scores. Module availability and weights can vary with configuration and input quality. The browser path may use a different available model, so two modes can disagree.

Visual artifacts

Looks for unusual texture, smoothing, and edge patterns. Editing, compression, and unusual photography can produce similar clues.

Metadata

Checks available EXIF/IPTC fields and software signatures. Missing metadata is common on social platforms and is not proof of AI generation.

Physics and frequency

Examines consistency of scene cues and repeated frequency patterns. Small files and screenshots can weaken these signals.

Texture contrast

Compares detail in richer and flatter regions. It is one contribution to a combined estimate, not a stand-alone test.

Model output

An optional image model contributes a score when an appropriate model is available. A model may perform differently on new generators or edited images.

Provenance

Checks for available C2PA information. A valid credential can inform origin; no credential does not mean an image is synthetic.

A useful way to read a scan

  1. Start with the best file available. An original export preserves detail and metadata that a screenshot or repost may remove.
  2. Read the score with the explanation. Inspect which signals contributed and whether the result reports uncertainty or conflicting evidence.
  3. Check independent context. Look for a source, capture history, reverse-image matches, and any verifiable Content Credentials.
  4. Escalate consequential cases. Do not accuse a person or remove content solely because one detector returned an AI label.

What the result cannot establish

A high score is not proof of who created an image, which generator was used, or whether the depicted event happened. A low score is not proof that a scene is authentic. Crops, screenshots, recompression, overlays, and unseen generators can change the evidence. We do not publish a universal accuracy percentage because performance depends on the dataset and decision threshold.

For a deeper account of evaluation metrics, see our benchmark guide. For a practical investigation, see the verification workflow.