How To Detect AI Generated Images At Scale

Checking a few images by hand is manageable. Checking thousands a day, for a marketplace, a social platform or a newsroom archive, needs an automated pipeline with people reviewing the hard cases.

1. Read provenance data first

It's cheap and precise when present. Parse each upload for C2PA Content Credentials, IPTC Digital Source Type values that indicate AI, generator names in Software fields, and Stable Diffusion settings in PNG text chunks. Open-source libraries such as ExifReader and the C2PA project's own tools handle this. Record the result before your system strips metadata for privacy.

2. Run a detection model

Most images won't have provenance data, so send them to a classifier. Commercial APIs from Hive, SightEngine and others return a probability that an image is AI-generated, often with a guess at the generator. Test several on a sample of your own images before choosing; accuracy varies a lot by image type.

3. Set thresholds and routes

• High score: label automatically or hold for review, depending on your policy.
• Middle scores: send to a human review queue.
• Low score: allow, but keep the score for audits.

4. Review and measure

Have reviewers record decisions and reasons. Regularly test the system on images you know the answer for, including real photos that have been heavily edited, to measure false positives and missed AI images. Models drift as new generators appear.

5. Handle appeals

People whose real work is flagged need a way to appeal. False positives cost trust fast, especially for photographers and artists.

For individual reviewers

A browser tool such as AI Image Detector (made by Patrick Bushe, who runs this site) is useful for spot checks during manual review: it reads metadata on any image you right-click and, if you add your own SightEngine or Hive key, can send it for a Deep Scan. It isn't built for bulk processing.

See AI image detection for content moderation and keeping false positives low.