Quick start
- Add your pictures. Drop one picture, several, or a folder anywhere on the page, or click the box. JPEG, PNG, WebP, GIF, HEIC and AVIF files work.
- Read the headline. Each picture gets a card with a verdict in plain words and a color: red when the file says an AI made it, amber for mixed signals, green for a camera photo or no clear sign of AI.
- Read the reasons. Under the verdict, What this is based on lists every finding, one line each: what the file says, camera records, the detector's reading, the size and the noise.
- Check how sure it is. Open How sure is this? for the test numbers behind a score, and What the file contains for the metadata that was found. Copy the report or download it as JSON.
The first picture that needs the detector downloads it once (57 MB) and your browser keeps it. A picture whose file already says an AI made it needs no detector and is answered at once.
What it checks, and how certain each check is
| Check | What it can tell you | How far to trust it |
|---|---|---|
| The file's own statements | A prompt or settings from Stable Diffusion front ends, a ComfyUI workflow, NovelAI data, Content Credentials (C2PA) naming a generator, an IPTC or XMP label such as "trained algorithmic media", the Software tag of an AI service | As certain as the file. It names the generator, so it is hard to fake by accident, but a file can be edited. We read the credentials but do not verify their signature. |
| The invisible Stable Diffusion watermark | The small mark the Stable Diffusion reference code writes into the pixels (both the version-1 and the SDXL message) | Certain when found. Found only in unmodified output: a JPEG save at quality 95 or lower removes most of it, and many front ends never add it. |
| Camera records | Make and model, lens, shutter time, aperture, ISO, the time to the fraction of a second, manufacturer notes, GPS, an HDR gain map | Strong evidence of a photograph (64 of our 107 real photos had it, 2 of 142 AI pictures), but copyable, and a real photo can be edited afterward. |
| The detector model | How much the picture looks like the AI pictures a model learned from | An estimate. On our test pictures it flagged 85% of the AI pictures and 25% of the real ones. |
| Size and noise | Whether the size is one generators give (1024 × 1024, 832 × 1216 and others), and how much camera-like noise the pixels carry | Weak on their own. Size only counts for the original file; noise cuts both ways (see below). |
How the verdict is decided
The checks are not averaged. They are used in order:
- The file says an AI made it. The verdict is "AI-made: the file says so", with the generator named and 100% shown as certainty about what the file says. If a prompt is stored, you can copy it.
- The file says AI was used on part of it (generative fill, Apple Clean Up, an extended canvas): "Partly AI". It does not say how much changed.
- The file carries strong camera records and no AI statement: "Looks like a real camera photo". If the detector disagrees strongly, the verdict is "Mixed signals".
- Otherwise an estimate: a score from 0 to 100 combining the detector, the size and the noise, shown with the share of our test pictures in that band that really were AI-made.
The score is not a probability. The four bands from our held-out test (whole generators and kinds of real picture left out) were: under 35, 14% of those pictures were AI; 35 to 65, 67%; 65 to 80, 81%; 80 and over, 89%. Because 57% of our test pictures were AI-made, these shares are high compared with a pile of mostly real photos. The numbers, and the detector's results per generator, are in the guide to telling if an image is AI-generated.
What it cannot check
- Watermarks of Google (SynthID), OpenAI and Meta. They can only be read by their makers' own detectors. A picture with only such a mark and no metadata looks unmarked here.
- Stripped files. Social sites, messengers and screenshots remove metadata. Only about one AI picture in four in our test still carried a statement.
- Drawings, anime and flat graphics. The detector called all five real anime pictures we tried AI, so for a picture that is mostly flat color no number is shown.
- Small pictures. Under about 400 × 400 pixels the measurements are not meaningful, so only the file is read.
- Edits and video. A real photo retouched by hand carries no statement, and a video needs tools made for video.
Camera records, and why they count the other way
A camera writes much that a generator has no reason to write. Finding it shifts the verdict toward "photograph": the card lists exactly which records were found. It does not make the picture real. The records can be copied onto any file, and tools such as Apple's Clean Up declare the edit separately, which is why such a picture is reported as partly AI and not as a plain photo.
Noise: a hint that fails both ways
The pixels of a camera photo carry fine noise, and many AI pictures are smoother. We measured it on the originals: real camera photos averaged a noise level of about 2.0, phone photos 1.3, and more than half of the AI services under 0.5. But ChatGPT pictures averaged 3.2 and Nano Banana 2.6, noisier than our camera photos, so the checker gives noise only a small weight. It cannot tell grain from a hidden code, and we know of no published pixel watermark for those services.
How long it takes
On our test laptop a picture took about two seconds in Chrome and in Safari's engine once the detector was downloaded, and a 45-megapixel photo about three. In Firefox the same model ran about ten times slower (around 17 seconds a picture), which is how that browser runs ONNX Runtime in our measurements, not something the page can speed up. A picture whose file already says an AI made it takes well under a second.
Your pictures
The checks run in your browser: the metadata engine, the watermark search, the pixel measurements and the detector model all run on your device. Like every tool on this site, the page also sends a copy of each picture, and its result, to our server (see the privacy policy), so do not add pictures you would not want stored. Nothing is sent to any outside AI service.
Short on time? Is this AI or not? gives the quick version of the same check, and our five-minute checklist says what to do with the answer.
Related tools: the Image Metadata Viewer shows everything in a file, the Image to Prompt tool describes a picture in words, and our guide to Content Credentials and watermarks explains the signals in a file.
Frequently asked questions
Can it tell for certain that a picture is AI-generated?
Only when the file says so: a prompt, a workflow, Content Credentials or an AI label, or the Stable Diffusion watermark. Then the answer is certain for that file. Without a statement it can only estimate, and in our test that estimate was wrong for about one in four real pictures.
Why does it say "no clear sign of AI" for a picture I know is AI?
Because AI pictures can be stripped of metadata and some services are built to look photographic. 14% of the AI pictures in our test scored in the lowest band. That band never means the picture is real.
Does a camera name in the file prove the picture is real?
No. Camera records can be copied onto any picture, and a real photo can be edited afterward, even with AI. They are good evidence, not proof, and a picture with no records says nothing either way.
Does it detect SynthID or ChatGPT pictures without metadata?
Not by a watermark: those marks can only be read by their makers. The detector model flagged all eight ChatGPT pictures in our test, but that is a judgment of how they look, shown as an estimate.
Why is there no number for my anime picture?
Because the detector is unreliable on drawings and flat-color art: it called every real anime picture we tried AI. For a picture that is mostly flat color the page shows what the file says and no estimate.