ChatGPT & DALL·E image detector

OpenAI labels its images — if you have the original file

OpenAI is a member of the C2PA coalition and has attached Content Credentials to its generated images since early 2024, starting with DALL·E 3. The credential is a signed manifest stored inside the file. It records that the picture was created by an OpenAI tool, using an action marked with the IPTC digital source type trainedAlgorithmicMedia, and it is signed with a certificate issued to OpenAI.

When GPTTrace finds that manifest, the verdict is marked “(provenance)”: the file itself says where it came from. You will also see who signed it and which actions were recorded, and you can open the full manifest in the C2PA viewer.

The limitation is equally important. Content Credentials live in metadata, and metadata is fragile. Screenshots, most social networks and many messaging apps remove it. So the absence of credentials never proves that an image is real — it only means the easy check is unavailable.

When the credentials are gone

For stripped copies, GPTTrace uses three weaker signals.

Image size

OpenAI’s models output a handful of fixed sizes: 1024×1024 squares, 1536×1024 or 1792×1024 landscapes and their portrait versions. An image at exactly one of those sizes with no camera data is suggestive. Re-sharing often resizes images, though, so this is only a nudge.

The look

Images from ChatGPT’s native generator became known for a warm, slightly yellow colour cast and very clean, even lighting, and for strikingly accurate text in signs, labels and posters. None of that is proof, but it is why so many viral “photos” of shop signs and documents turned out to be ChatGPT output.

The neural classifier

The classifier that runs in your browser was trained on output from thousands of generators. It looks at texture and noise statistics rather than content, so it can flag an OpenAI image even when the scene is mundane. Its confidence drops with heavy compression and small sizes, which is reflected in the result.

Where ChatGPT images cause problems

Because ChatGPT is the most widely used assistant, its images turn up everywhere: product listings with impossible photos, fake screenshots of conversations, invented historical pictures, school projects and insurance or refund claims with doctored evidence. In each case the best defence is the original file. Ask for it. A person who really took a photo can usually send the camera original, with its make, model and timestamp intact — and GPTTrace will show those fields too.