PhotoCheck API

AI image detector

Find out whether a photo was made by a generator. You get a score from the pixels and a read of what the file itself declares.

Drop a photo to check it

or paste one with ⌘V / Ctrl+V

JPEG, PNG, WebP or GIF up to 10 MB. Checked in memory, never stored.
Have an API key?

With a key, checks here spend your credits instead of the free daily allowance. It stays in this browser.

Two kinds of evidence

PhotoCheck checks whether an image is AI-generated in two independent ways. A trained classifier scores the pixels from 0 to 1, and a metadata scan reads any provenance the generator attached. When a file declares it was generated, that is near-conclusive. When it declares nothing, the score is the evidence.

The classifier is umm-maybe/AI-image-detector (CC BY 4.0), run on our own server. We picked it after testing three openly licensed detectors on images with known provenance. It was the only one to separate all of them cleanly. What that test did and didn't show.

What the score can't do

  • Prove anything. It's a probability from a model, and models are wrong in both directions.
  • Keep pace with every new generator. Expect more mid-range scores on models released after it was trained.
  • Judge screenshots and heavy filters well. Both blur the texture it reads.
  • See invisible watermarks like Google's SynthID, which only the vendor can decode.

AI detection API

POST /api/v1/ai-detect returns the score and the metadata scan for one credit. Use /analyze to add face detection at no extra cost.

Endpoint reference

bash
curl https://photocheckapi.com/api/v1/ai-detect \
  -H "Authorization: Bearer $PHOTOCHECK_API_KEY" \
  -F image=@suspect.jpg

Questions

How does the AI image detector work?

Two ways at once. An image classifier (umm-maybe/AI-image-detector) scores the pixels from 0 to 1, and a metadata scan looks for labels generators leave on purpose: C2PA Content Credentials, IPTC digital source types, and Stable Diffusion or ComfyUI generation data.

What score means AI-generated?

0.8 and above: strong signs. 0.5 to 0.8: some signs, look harder. 0.2 to 0.5: probably a real photo. Below 0.2: no signs. If the file declares it was generated, that outranks any score.

Can it detect AI edits in a real photo?

Only sometimes from the pixels, because most of the image is real. When the editor recorded the edit in C2PA, as Photoshop's generative fill does, the metadata check reports it as a composite with AI content.

Is it free?

Yes, 20 checks a day on this page with no account. The API version costs one credit per image, from $0.0075.