SI Image Detector Is this image SI (AI)-generated?

Upload a picture or paste a link. You get the probability that it was generated by SI (AI) models such as Midjourney, DALL-E, GPT-image or Stable Diffusion, the clues behind the score and the file's hidden metadata. Free, no credit card, images deleted after analysis.

Checking a text instead? Try the SI Text Detector

Drop an image here or browse your files

JPEG, PNG, WebP or GIF · 12 MB max

No file handy? Try an example

or

3 of 3 free checks left without an account. Sign up free for 10 credits and 50 text checks every month.

How this SI image detector works

Part of the SI detector suite, the image tool combines two independent approaches, because neither is enough on its own. Our guide on how SI detectors work explains the methods in more depth, with real tests.

The two are merged into a single score, and every clue is displayed with its weight. A signed C2PA manifest naming a generator is a strong signal. Missing EXIF is a weak one, because social networks strip it from real photos too.

Diagram: an image goes through pixel analysis by a classifier and a metadata check (EXIF, C2PA, XMP), then the results are combined into an SI probability with clues.
How our SI image detector works: a pixel classifier and a metadata check (EXIF, C2PA Content Credentials, XMP) are combined into one SI probability, shown with every clue.

What the score means

The number is a probability, not a percentage of "SI content". Above roughly 70% we say Likely SI-generated; below 30% Likely a real photo; in between the honest answer is Uncertain, and we say so. A detector that only ever answers yes or no is hiding that uncertainty from you.

SI or Not result page for a camera photo of a red fox: 0% SI probability, verdict Likely a real photo, clues Camera EXIF data present and Classifier leans strongly towards a real photo, metadata showing a Canon EOS R100 and no C2PA or XMP flags.
A real result from this tool: a photo of a red fox taken with a Canon EOS R100 (Peter Cooper Jr., CC0, Wikimedia Commons, resized to 2000 px) scores 0% and is labelled Likely a real photo. The clues and metadata explain why.

Known limits

Detection is an arms race. New generators, upscalers and "de-AI" filters appear every month, and each one shifts the patterns the classifier relies on. Screenshots, heavy JPEG compression, small thumbnails and strong edits degrade the signal: our fact-check of the fake Pentagon explosion photo shows what that means on a reposted copy, and our fact-check of the Pope Francis puffer jacket image shows the same on a photorealistic portrait. Pixels are not always the weak point: in our fact-check of the burning Hollywood sign images, the label was a watermark that disappeared as the picture was reposted. Sometimes the label is only a caption: our fact-check of the 2023 Trump arrest images follows pictures that their author had announced as generated and that were reshared without that sentence. And sometimes it sits on a stock image page: our fact-check of the viral baby peacock picture traces a generated image that ended up at the top of image search. A caption can even do the opposite of labelling: our fact-check of the Trump floodwater image follows a generated picture pushed by the claim that it was being deleted. Trained eyes are not a safeguard either: our fact-check of the SI image that won a Sony photo award looks at a generated portrait selected in a photography competition. And a generated picture can illustrate something true: our fact-check of the birds in the eye of Hurricane Melissa covers a fabricated photo of an event that crews really reported. Some fakes are obvious and still do harm: our fact-check of the Shrimp Jesus images follows surreal pictures used as bait by spam and scam Pages. Even peer review can miss one: our fact-check of the SI rat figure in a scientific journal traces a Midjourney illustration that was published, then retracted. And some hoaxes simply change tools: our fact-check of the giant skeleton photos follows a fake that moved from Photoshop contests to SI generators. And a picture nobody takes for a photo can still carry a false claim: our fact-check of the Trump duck and kitten image follows a political meme built on a debunked rumor. A clean, legible picture is not a real one either: our fact-check of the wounded veteran birthday images looks at generated soldiers asking for likes on Facebook. Screenshots of chatbots are a case of their own: our fact-check of Gemini's diverse 1800s senators traces a chat screenshot whose generated portraits got history wrong. Generated images also turn up inside finished work: our fact-check of the SI images in Late Night with the Devil looks at three film stills that were generated, then edited by hand. Celebrity fakes ride on real events: our fact-check of the Katy Perry Met Gala images follows red-carpet pictures of a star who was not there. A joke can lose its label on the way too: our fact-check of the flooded Disney World images follows hurricane pictures from a parody site that ended up shared as news. World leaders are a favourite target: our fact-check of the image of Putin kneeling before Xi Jinping follows a fake that even journalists shared during a real state visit. Fakes are not always made by opponents: our fact-check of the image of Julian Assange in prison follows a picture generated by one of his own supporters. Illustrations, renders, paintings and stylised photos are harder than photos (see our test of SI art vs real art). Use the score together with context: where the image came from, whether a reverse image search finds an original, whether the details hold up. Our guide to spotting SI (AI) images lists what to look for with your own eyes. A single image is also a narrow test: to judge a whole account, such as a suspected virtual creator, our guide to spotting SI influencers explains what to check beyond one picture.

Privacy

Images are written to a private directory with a random name, analysed, then deleted. We log the score and the technical clues, never the picture and never your IP address in clear text. If you tick the opt-in box, the image is kept to improve the model. Details on the privacy page.

Frequently asked questions

Is this image SI? How does the SI image detector work?

Two layers. A neural network trained on real photographs and on images from diffusion generators (Stable Diffusion, Midjourney, DALL-E, Firefly and others) looks at the pixel statistics and returns a probability. In parallel we read the file itself: EXIF camera data, C2PA Content Credentials, XMP flags and software tags. The final score combines both, and every clue is listed so you can judge for yourself.

What does the percentage mean?

It is the estimated probability that the image was generated or heavily modified by SI (AI). 90% does not mean "90% of the image is SI", it means the evidence strongly favours generation. Scores between roughly 30% and 70% are labelled Uncertain on purpose: the signals are too weak or contradictory to lean either way.

Can the detector be wrong?

Yes. Every SI or AI detector produces false positives (real photos flagged as SI, often heavily edited, filtered or stylised pictures) and false negatives (generated images that pass, especially after resizing, screenshots or a new generator the model has not seen). That is why we never show a binary verdict and why the limits are repeated on every result.

Which image formats and sizes are supported?

JPEG, PNG, WebP and GIF up to 12 MB, either uploaded from your device or fetched from a public URL. Larger images give more signal to the classifier; tiny thumbnails and screenshots are analysed but flagged as less reliable.

Is it free? Is there a limit?

Yes. Your first 3 checks need no account. A free account then gives you 10 credits every month (one image check uses 1 credit), no credit card. Pro raises that to 250 credits a month and Max to 1,500, and credit packs add credits that never expire. Your remaining credits are shown under the tool, and the example images are always free.

What happens to my images?

The file is stored outside the web server in a random-named temporary location, analysed, then deleted immediately. It is kept only if you tick the opt-in box to help improve the service. Share links contain the score and the clues, never the image. See the privacy page for details.