AI image detector

Inspect images.
Keep the context.

Submit a JPEG or PNG and review a warning label, raw score, active model and processing source. The result does not prove who created the image.

Open the free checker

Scope before score.

Every result is a warning signal. The evidence and limitations below define what this route can and cannot support.

Production lane

The live image route uses the pinned OwensLab Community Forensics 384 model with a versioned 0.50/0.80 warning policy.

Input boundary

The public checker accepts JPEG or PNG up to 4 MB and 25 million decoded pixels.

Evidence level

An internal 200-image evaluation reported 90.5% accuracy, 85% synthetic recall and 4% real false positives. It is not independent validation.

Do not infer more than the lane measured.

  • Art, screenshots, compression and unseen generators can fail.
  • A low score does not establish that an image is human-made or authentic.
  • Model and training-data rights remain a commercial diligence item.

Continue with evidence

Benchmark and limitationsRead more →False positivesRead more →Human reviewRead more →