Frequently asked questions

You upload an image. A frontier vision model analyzes it against a prompt designed to surface AI-generation signals. You get back a verdict (AI Generated, Real Photo, or Inconclusive), a confidence bucket, a written rationale, and a list of specific visual signals the model noticed.

Accuracy depends on the type of image and the generator. Against the labeled test set we evaluate against, the active prompt's accuracy is reported on every change — we don't ship prompts that don't improve the score. No detector is perfect, especially against adversarial post-processing or brand-new generators.

Photographs and photo-realistic images are the primary target. Illustrations, paintings, and obvious art are out of scope — the model will likely return Inconclusive on those.

JPG, PNG, WebP, and GIF up to 5MB. Larger images get downsized server-side before analysis. EXIF metadata is stripped before the image leaves your browser.

A frontier vision LLM accessed through a managed inference platform. The specific model is configurable per environment and recorded with every check result so accuracy reports stay reproducible.

Classifiers give you a number. We want a reason. A vision LLM can commit to a verdict AND explain itself — and the explanation is the constraint that keeps the verdict honest. We tune the prompt against a labeled eval set on every change.

It means the model found signals pointing in opposite directions and isn't willing to commit. Treat Inconclusive as a real verdict, not a failure — it's better than a false positive.

If you're signed in, yes — we store them with your check history so you can revisit results. If you're anonymous, the image is associated with the rate-limit row only and rolls off when the row does. You can delete your account from the user menu at any time to remove everything. See our privacy policy for the full details.

No. The terms of our inference provider guarantee customer inputs are not used to train the foundation models we call. We don't train any models on user submissions ourselves.

We use a small set of established third-party service providers, each for a specific function — AI inference, image storage, database, authentication, payments, and (for free-tier users) advertising. None receive your image or detection results for any purpose other than performing that function. See the privacy policy for the category list; the named sub-processors are available on request.

Pro is $10/month — see the pricing page for what's included. Cancel anytime from the user menu — Pro stays active through the end of your billing period.

Vision-model inference costs real money per request. Ads keep the free tier free without forcing you to sign up. Pro removes them.

Not via self-serve yet. If you have a real need (volume, integrating into a workflow), email us.