ChatGPT Image Detector

Add the picture to the ChatGPT image detector above. If it is an original file saved from ChatGPT or OpenAI's image API and it still carries the Content Credentials OpenAI signs into its images, your browser reads them and answers Likely AI-generated on the spot. No detection model is called and none of your 3 free checks is used. If the credentials are gone, as in a screenshot, a detection model judges the pixels instead.

Content Credentials are a signed record in the C2PA format. The record names the tool and the signer, says how the image was made, and is bound to the file's image data, so the checker can see whether the file changed after signing. It settles the answer only when the signer is on the C2PA trust lists we check and the file is unchanged; a company name that merely appears in a file isn't enough. You can also read a file's credentials on their own. The model step uses Sightengine, with Hive as a backup if Sightengine fails, the same engines as the AI image detector on our home page.

All 8 original GPT Image files in our test carried OpenAI-signed Content Credentials saying they were AI-generated, so the checker settled every one of them without a model. The 2 screenshots carried no credentials and went to the model, which called both AI. That made 10 right out of 10 for the ChatGPT image detector, but on a small sample: 10 images, tested on October 13, 2026.

We made 6 of the originals ourselves through OpenAI's image API (via a reseller that relays requests to it), with GPT Image 2 and GPT Image 1.5. Their credentials name OpenAI Media Service API as the tool and are signed by OpenAI OpCo, LLC, which is on the current C2PA trust list. The other 2 are GPT-4o images from a public dataset. Their credentials name Sora as the tool and are signed by OpenAI, which is on the older interim trust list that the official C2PA list replaced. The checker accepts both lists and tells you which one the signer is on.

We didn't generate anything inside the ChatGPT app itself. OpenAI said in 2024 that images from ChatGPT and its API would carry Content Credentials, so a file saved straight from ChatGPT should look like our API files. That is our expectation, not something we measured. The accuracy report has the same test for every generator.

A screenshot of an image from ChatGPT carries none of OpenAI's credentials, because your phone draws a brand-new file. Neither of our 2 GPT screenshots kept them. The checker then sends the picture to Sightengine, and both screenshots scored above the 0.8 line it uses for Likely AI-generated. The lowest score any of our 10 GPT images got was 0.86.

You can try both paths with the sample buttons under the checker. The first sample is one of our GPT Image originals, a fake news photo of a flooded street, and its answer comes from the credentials. The second is a screenshot of the same picture in a social feed: no credentials, so the answer comes from the model score we recorded in our test, which was that 0.86. Hive, our backup engine, caught 9 of the 10 GPT images at its recommended 0.9 threshold. For any screenshot, our guide to viral pictures explains how to find a better copy first.

OpenAI runs its own check at openai.com/verify. It looks for two OpenAI signals: the same C2PA credentials our checker reads, and an invisible SynthID watermark that is part of the pixels and may survive some changes. This ChatGPT image detector doesn't read SynthID, so OpenAI's tool can sometimes confirm an image from ChatGPT that we can only score.

The two tools answer different questions. OpenAI's tells you whether a file carries OpenAI's own signals, and OpenAI says it doesn't detect images from other companies' models. Ours gives a likely answer for an image from any generator, credentials first and a detection model after that. OpenAI also says a no-signal result doesn't prove an image wasn't made with its tools: the metadata may have been stripped, the watermark degraded, or the image made with an older model or before the signals existed. If the picture matters, run both. Google offers a similar SynthID check for Nano Banana images.

A Likely AI-generated answer from the ChatGPT image detector, backed by OpenAI's credentials, tells you which company's service produced the file and that the file hasn't changed since. It doesn't tell you who typed the prompt, whether the person sharing it made it, or whether it is being passed off as a real event. Many images from ChatGPT are honest illustrations.

• Who made it. OpenAI's documentation tells developers not to infer a prompt, an account or an individual creator from a verification result. Our checker can't either.
• What happened later. If the file is edited after signing, the credentials no longer match it. The checker then stops relying on them and lets the model decide.
• How much is real. ChatGPT can also edit a photo someone uploads. A credential that says AI-generated may not show how much of a real photo is underneath.
• The context. Whether a picture is a joke, an ad or a hoax depends on how it is used, not on how it was made.

A ChatGPT image detector answer usually falls into one of three cases. Each points to a different next step, and none of them is a verdict on the person who shared the picture.

• Likely AI-generated, from credentials. The strongest case. The reasons name the signer and say which trust list it is on.
• Likely AI-generated, from the model. Usually a screenshot or a re-saved copy. It's a probability, so quote the range.
• Can't tell or Likely a real photo. The model wasn't convinced. If you still suspect ChatGPT, look for the first post or the original file and try OpenAI's verification tool.

Other generators leave different traces. Midjourney signed none of its files in our test, so its pictures always go to the model.

Not for an original file with OpenAI's credentials: those are read in your browser and settle the answer without a detection model. A screenshot or a re-saved copy needs the model, which uses one of your 3 free checks a day, unless someone already checked the same file.

Only when credentials are present: they name OpenAI as the signer and the tool that made the file. Without them, the model judges whether the picture looks AI-generated, not which service made it. Hive, our backup engine, sometimes guesses a generator from the pixels, and the result marks that as a guess.

No. OpenAI's own verification tool checks for it. Our checker reads Content Credentials and metadata, then asks a detection model.

It may still have been. OpenAI has said that most social platforms remove metadata from uploaded images and that a screenshot can remove it too, so a missing credential tells you nothing either way.