How this AI photo detector works
This AI photo detector looks for three different kinds of evidence and shows each one separately, because they are not equally reliable. Mixing them into one number would hide the difference between a signed record from the generator and a statistical guess.
1. Content Credentials (C2PA)
Content Credentials are a signed record that travels inside the image file. OpenAI (ChatGPT and the GPT image models), Google (Gemini and Imagen), Adobe Firefly, Microsoft and a growing list of cameras and editing apps add them. When a manifest says the image was made by a "trained algorithmic media" source, that is the generator itself stating the picture is AI generated. The detector verifies the signature and shows who signed it and when. See the C2PA checker for the full manifest history.
2. Metadata
EXIF, XMP, IPTC and PNG text fields often name the software that produced an image. Midjourney, Stable Diffusion front ends such as Automatic1111 and ComfyUI, NovelAI and Google AI all leave recognisable tags or embedded prompts. Camera fields such as make, model, exposure time and lens are listed too, but only as weak evidence: they can be copied, and most social apps strip them.
3. The detection model
When the file carries no labels, a neural network looks at the pixels. It is a compact vision transformer trained on millions of real photos and images from thousands of generators, compressed to about 22 MB so it can run in the browser with WebAssembly. It returns an AI score from 0 to 100, tuned so that real photos rarely reach the AI band. The first check downloads the model once; later checks start straight away.
Is this image AI? Reading the verdict
The answer comes in three levels, followed by the reasons behind it:
- AI-generated (embedded evidence): Content Credentials or metadata state that a generator made the image. This is the strongest result the tool can give.
- Likely AI-generated: no labels, but the model score is high. Treat it as a strong hint, then check the source.
- Not sure: the score sits in the middle, or the evidence points in different directions. Many edited photos and recompressed AI images land here.
- Likely a real photo: nothing points to generation and the model score is low. New generators can still pass as real, so this is a likelihood, not a guarantee.
Below the verdict, each line shows its own result: whether credentials were found, which metadata tags matched, and the model's AI score. "Show details" lists the signer, actions, software and camera fields so you can judge the evidence yourself.
How to check if a photo is AI generated
- Get the best copy you can. The original download keeps its credentials and metadata. A screenshot or a forwarded image keeps only the pixels.
- Drop it into the detector above, paste it with Ctrl+V, or paste an image link.
- Read the verdict and its reasons. Embedded evidence settles the question; a model score is a likelihood.
- Look at the picture itself. Hands, text, jewellery, reflections, ears, teeth and repeated background patterns still trip up generators.
- Check where it came from. A reverse image search often finds the first post, and an account that only posts flawless images with no behind-the-scenes shots is a warning sign.
The step-by-step guide to check if an image is AI generated covers the visual clues in more detail.
What the AI photo detector is good at, and where it struggles
| Image | What to expect |
|---|---|
| Fresh download from ChatGPT, Gemini, Firefly | Usually caught by Content Credentials, with the tool named. |
| Stable Diffusion, ComfyUI, NovelAI PNG | Usually caught by the embedded prompt or workflow. |
| Stable Diffusion-family art (anime, illustration), no metadata | Model catches most of it. |
| Newest models (GPT Image, FLUX.2, Seedream, Midjourney v7), no metadata | Hardest case. The model misses many of these; expect "Not sure" or a false "real". |
| Real photos | Rarely flagged as AI (about 1 in 100 in our tests). |
| Phone photo with filters or AI editing | May score in the middle. Partly AI edits are a grey zone. |
We publish measured accuracy by generator in the accuracy report and update it whenever the model changes. We would rather show where the detector fails than claim a number no detector can reach.
Private by design
Many AI image checkers upload your photo to a server, and some even publish scan results as public pages. Photos of people, documents and private moments should not need to leave your device to be checked. Here the image is read, analysed and discarded inside the browser tab. There is no upload, no account and no stored history.
Who uses an AI image detector
- People checking a viral picture before sharing it, from breaking news photos to celebrity images.
- Dating and marketplace users checking whether a profile photo or a product shot is real.
- Teachers and contest judges looking at photography or art submissions, as one input next to the original files.
- Journalists and moderators triaging user-submitted images before deeper verification.
- Artists and buyers asking whether a piece of digital art was made with AI. The AI art detector page explains what changes for illustrations.