How To Identify DALL-E Generated Images

OpenAI's image models, from DALL-E to the image generation built into ChatGPT, leave a few traces you can check, at least on files that haven't been altered.

Content Credentials

Since February 2024, OpenAI has added C2PA Content Credentials to images made with DALL-E 3 in ChatGPT and its API, and it does the same for newer ChatGPT image generation. The credentials record that OpenAI's tool created the image. Upload the file to Content Credentials Verify to read them.

The old DALL-E 2 stripe

Images from DALL-E 2 carried a small strip of five colored squares in the bottom-right corner. It was easy to crop, and newer models don't add it.

Visual traits

There's no reliable "DALL-E look," but common traits include:

• Clean, evenly lit compositions with a slightly illustrated quality, even in "photos."
• Text that's mostly right but has odd letters or spacing; newer models handle text much better.
• Symmetrical, centered framing.
• Very smooth surfaces and skin.

These overlap with other generators, so they can't tell you which tool was used.

Check the file in your browser

AI Image Detector (made by Patrick Bushe, who runs this site) reads an image's metadata when you right-click it and choose "Detect AI Image." On an original DALL-E or ChatGPT file it should spot the Content Credentials entry and flag it. It doesn't read the full credential record, so use Content Credentials Verify to confirm what it says.

Why the clues disappear

Screenshots, cropping, re-saving and posting to most social networks remove Content Credentials. OpenAI itself has said the metadata isn't a complete solution for this reason. If the file has been shared around, a reverse image search to find the earliest copy is often more useful.

See Midjourney metadata and Stable Diffusion images.