A free image checker for fact-checkers & verification desks
Upload a photo and get the probability it was AI-generated — not a yes/no verdict. Free, no sign-up, and the image is deleted right after analysis. Below are our real accuracy numbers — the good and the bad — and what it can't catch.
1. What it does
Upload or paste an image and it returns a single number: the estimated probability the picture was produced by an AI image generator rather than captured by a camera. It's trained on a large, diverse set of real and AI-generated images and refreshed as new generators appear. When an image carries Content Credentials (C2PA) or camera metadata, that's factored in as a corroborating signal. No account, and the file is securely deleted the moment you get a result — we don't store or share your images.
2. How accurate is it — the real numbers
Measured on images the tool had never seen before, as of July 2026. We publish the weak spots too — most tools don't.
How often it catches AI images, by generator
| Generator | Caught as AI |
|---|---|
| DALL·E 3 | 90% |
| Stable Diffusion XL | 90% |
| Stable Diffusion 2.1 | 88% |
| Flux.1 | 86% |
| Midjourney v6 | 85% |
| Stable Diffusion 3 | 78% |
| Stable Diffusion 3.5 — newest, weaker | 62% |
| Artistic / stylized AI — see limits | 63% |
Overall, on a balanced mix of modern photorealistic fakes and real photos: ~89% accuracy.
How often a real photo is correctly kept as real
| Real content type | Kept as real |
|---|---|
| Screenshots & UI | 100% |
| Artwork, textures, x-rays, faces, interiors | 99% |
| Everyday photos | 95% |
| Older / varied web photos | 86–90% |
In plain terms: a genuine photo is wrongly flagged as AI roughly 1–14% of the time, depending on the image.
Prefer to verify yourself?
Tell us you're evaluating and we'll lift the usage cap so you can batch-test your own set of real and AI images — no key, no sign-up. Independent results beat any vendor's self-reported table. Get in touch →
3. What it can not reliably detect
This is the part most tools leave out. Please treat these as real limits:
- Brand-new generators. New models appear faster than any detector can keep up (Google's newest, for one, isn't yet covered) — the very latest images may slip through until the model is refreshed.
- Artistic or heavily stylized AI (painterly, dramatic, or heavily prompt-engineered) is harder than photorealistic fakes — the tool is tuned toward realistic images, which matter most for misinformation.
- Heavily compressed, re-saved, or screenshotted images. Social platforms strip and recompress images, which weakens the signal.
- Partial or hybrid edits — inpainting, a face-swap onto a real photo, or a real scene with one AI element. It judges whole-image AI-vs-real, not localized manipulation.
- It returns a probability, not proof. A high score is a strong reason to investigate; a low score is not a guarantee the image is authentic.
4. How to use it in a verification workflow
Treat it as one fast triage signal, not a ruling:
A high score means
Strong reason to dig deeper — pair it with reverse image search, provenance (Content Credentials), and human judgment before publishing.
A low score means
No obvious AI fingerprint — but not a clean bill of health, especially for compressed, cropped, or newest-generator images.
It's built to sit alongside the tools you already use — reverse image search, metadata / Content Credentials, InVID-WeVerify, Forensically — as the quick first pass. For provenance specifically, our metadata & C2PA checker reads Content Credentials and EXIF directly.
5. Privacy & getting in touch
Every uploaded image is analyzed and then securely deleted — no storage, no sharing, no account required. If you're evaluating the tool for a newsroom or classroom and want the usage cap lifted, or want to flag a case where it got something wrong, get in touch — that feedback is exactly how it improves.
wedetect.ai is an independent, free tool. Accuracy figures as of July 2026; we keep this page current as the model is updated.