Find out whether a picture was made by AI or taken with a camera. Our AI image detector reads the file’s metadata, inspects its pixels and runs a detection model, then shows you the evidence.
Drop images here to check themChoose photos to check
Up to 20 at a time · JPG, PNG, WebP, GIF or TIFF · up to 10 MB each
FreeNo sign-upImages are never stored
An AI image detector, sometimes called an AI image checker, estimates whether a picture was made by generative AI or captured with a camera. FreeAIDetector does it for free. Add a photo, a screenshot or a piece of art, and it combines three independent checks into one likelihood score: provenance metadata, pixel forensics and a trained vision model.
Every result comes with its evidence: the labels found in the file, how the pixels behave and what the model thinks. You can judge the call yourself instead of trusting a bare number.
Three steps and no settings. The AI image detector starts checking the moment you add an image.
Drag and drop up to 20 images, choose them from your device, paste one with Ctrl+V, or paste an image link.
Your browser reads the file’s metadata and measures its pixel noise. If that isn’t conclusive, a vision model trained on real and AI images takes a closer look.
Each image gets a likelihood score and a verdict: likely real, uncertain or likely AI. Open the details to see exactly what was found.
No single test catches everything, so the AI image detector stacks three. Each one covers what the others miss.
Many AI tools label what they make. The detector reads C2PA Content Credentials, the IPTC DigitalSourceType field, EXIF software tags, and the generation settings Stable Diffusion and ComfyUI write into PNG files.
IPTC: trainedAlgorithmicMedia
Software: Adobe Firefly
Steps: 30, Sampler: DPM++ 2M
Your browser measures fine grain, sharp detail and flat areas, and checks for camera data. It tells a screenshot from a photo, but texture alone can’t prove an image is AI, so it adds context rather than deciding.
Texture: moderate grain
Flat areas: 3%
Camera: Canon EOS R6
When the file carries no labels, our server runs a classifier trained on real photos and generated images. It judges the pixels themselves, so it still works after metadata is stripped.
Model score: 91% AI
Confidence: high
Result cached for repeat checks
The score is the estimated chance that an image was generated by AI. Every score falls into one of three bands.
Camera-like signals and nothing that points to AI. Most unedited phone and camera photos land here.
Weak or mixed signals. Common for screenshots, heavily edited or compressed images and illustrations. Check the evidence and where the image came from.
Strong signs of generation: an AI label in the metadata, or a confident model score. A metadata label alone puts the score at 98%.
Next to the score, confidence (high, medium or low) tells you how solid the evidence behind it is.
When a file carries a generator’s signature, the detector names the tool. Images from newer or unlisted generators are still scored by the model.
Anyone who needs to know whether a picture can be trusted.
Check viral images and reader-submitted photos before they’re published or shared.
Review image assignments and learn how synthetic media gives itself away.
Look twice at product photos and listings that seem too perfect to be real.
Find out whether a profile picture was generated before you trust the person behind it.
Screen entries for photo and art competitions that require human-made work.
Spot AI-generated images in reports and posts, one careful check at a time.
Your images stay yours. The detector is built to check them without keeping them.
Treat a result as evidence, not a verdict on a person, and weigh it against where the image came from.
A missing label proves nothing.
Screenshots and social networks strip metadata from real and generated images alike.
Edits and compression blur the signals.
Heavy filters, resizing and repeated re-saving can push a score in either direction.
Content Credentials are detected, not verified.
The tool notices a C2PA manifest but doesn’t validate its signature, and it can’t read invisible watermarks such as SynthID.
New generators keep arriving.
Detection is retrained over time, but no detector keeps up perfectly with every new model.
Accuracy, privacy, file formats and what to do with a result.
Yes. There is no sign-up, subscription or credit system. Fair-use limits keep it fast for everyone: you can check up to 20 images at a time, and deep scans have a generous daily allowance.
When a file carries an AI label in its metadata, the result is close to certain. Without one, the score comes from a vision model, which handles typical AI images well but can be misled by heavy editing, compression or unusual styles. The browser’s pixel check adds context but never decides a result on its own. That is why every result shows the evidence behind it.
Yes. It recognizes the metadata these and other generators leave in their files, and when that metadata has been removed, the vision model still scores the pixels.
Yes, with less certainty. Screenshots and most social networks strip metadata, so the result relies on the pixels and the model. Expect more “uncertain” results for small or heavily compressed images.
No. Metadata and pixel checks run in your browser. When a deep scan is needed, the image is processed in memory on our server and discarded right after. It is never saved or used to train models.
JPG, PNG, WebP, GIF and TIFF, up to 10 MB each. You can also paste a direct link to an image.
The signals are weak or point in different directions, for example an image with no metadata where the model isn’t confident either way. Look at the details, where the image came from, and whether other copies of it exist.
Heavy retouching, beauty filters, AI-powered phone processing, upscaling and very clean studio shots can all look synthetic to a detector. If you think a result is wrong, let us know through the contact page. Reports help improve it.
Still have a question? Contact us
Sourced explainers that go with this detector.
Five checks that still work, from the labels inside the file to a reverse image search, and why hunting for glitches no longer does.
Read the guideThe four ways a detector can tell, from labels and watermarks to classifiers and statistical tests, and why each one is an estimate.
Read the guideIndependent studies, a regulator’s order and a test on a public benchmark: what detectors get right, where they fail, and how to read an accuracy claim.
Read the guideDrop in a photo and get a verdict, with its evidence, in seconds. Free, and no sign-up.
Check an image