Inspect the files. Reproduce the results.

Original image metadata research with downloadable inputs, exact hashes and recorded outputs, released under CC0 so anyone can repeat it. See precisely which creation records a file stores and what survives an export.

Published datasets and analysis

Every collection ships its method, input files and observations together, so you inspect the result and repeat the check in minutes.

Reproducible image experiment

Image Metadata Lab

Inspect saved prompts, conflicting records, ComfyUI graphs and export behavior. Download the original PNGs, verify their hashes and reproduce the recorded output.

7 worked cases · 10 original filesControlled, code-drawn images with deliberately attached creation records.Download observations and hashes (JSON)
Reproducible image experiment

What survives an image export?

Compare six operations across three constructed creation-record formats. Inspect field retention and exact pixel equality, then check your own original and exported files.

18 export testsVersioned local export tests; record formats describe the attached data, not the origin of the pixels.Download observations and hashes (JSON)
Reproducible image experiment

Platform metadata survival study

Which AI and provenance records survive a platform upload? Five test files carrying C2PA, IPTC, EXIF/GPS, saved-prompt and ComfyUI records, uploaded and downloaded again.

5 test files · 1 platform testedDated uploads through each platform's official upload path; untested platforms are listed as not yet tested.Download observations and hashes (JSON)
Reproducible image experiment

Free AI image detectors tested on 24 images

12 AI images and 12 camera photos from Wikimedia Commons, uploaded unchanged to five free AI image detectors and isGenAI. Every verdict, false positive and free limit, with files and hashes.

5 test files · 1 platform testedLicence-clean Commons files; each tool used as an anonymous visitor until its free limit.Download observations and hashes (JSON)

For researchers: test isGenAI, cite it, get access

Free research access. The check API and the MCP server need no key for single checks. If you are benchmarking AI detectors, provenance tools or C2PA adoption and need more AI scans, email h.a@skyvaults.io with your project and we will raise your limit at no cost. Three conditions, the same ones other detector vendors ask for: tell us before you publish, let us publish a short summary of your results on this page, and cite isGenAI by name and URL.

Evaluate us, including when the result is bad. We invite independent tests of the text detector, the image checks and the AI replies estimate. There are no publishing conditions: send us the link and we will list every outside evaluation on this page, favourable or not. No outside evaluation has been published yet.

Open test material. The Image Metadata Lab files, export tests and platform survival study above are CC0, with hashes and reproduction scripts, and the fixtures are mirrored on Hugging Face.

How to cite isGenAI.

@misc{isgenai2026,
  title        = {isGenAI: AI detector and origin-evidence checker for images, text, code and links},
  author       = {{isGenAI}},
  year         = {2026},
  howpublished = {\url{https://isgenai.com}},
  note         = {Reads C2PA Content Credentials, IPTC AI labels and saved prompts, and gives an AI-likelihood score. Accessed: <date>}
}

Name the date of your checks and the endpoint or page you used, since detectors and providers change over time; the changelog records each change.

From a published result to your own check in three steps

Start with a supplied original, inspect the recorded fields and repeat the documented operation. Keep the file hash and software versions with your result and you have a citable reproduction. The examples measure metadata reading and file transformations, not AI-detection accuracy or authorship, and every case states that scope.

Run the image metadata API examples · Follow the worked guides · Read the checker methodology