DALL·E image detector
last updated 7 june 2026
DALL·E images (now generated through ChatGPT and GPT-4o) carry C2PA Content Credentials that explicitly mark them as AI-made — and verifai reads them on-device in one click, flagging credentialed DALL·E images right on the page. Where that metadata's been stripped, verifai marks the image "unverified" rather than guessing; a hosted deep-scan for those cases is planned.
OpenAI attaches C2PA "Content Credentials" to DALL·E and GPT-4o image output, so an unaltered DALL·E image can be verified as AI-generated by reading its provenance metadata. The catch: screenshots, re-saves and most social platforms strip that metadata, leaving you with visual tells and a detector.
verifai reads each image's embedded credentials on the page in one click. When a DALL·E image still carries its credentials, that's an authoritative flag; when they've been stripped — common for images already on the web — verifai says "unverified" rather than pretending, and the planned hosted deep-scan is what will cover those.
Check Content Credentials first
If you have the original file, check it for C2PA Content Credentials (for example via the Content Authenticity Initiative's verify tool). Present-and-valid credentials are strong confirmation. But remember the asymmetry: credentials present = reliable yes; credentials absent = unknown, not no.
Visual tells when metadata is gone
DALL·E output shares the common diffusion artefacts:
- garbled text and lettering within the image
- hands and small anatomy errors
- a smooth, slightly illustrative rendering style
- lighting and reflections that don't fully cohere
frequently asked
Do all DALL·E images have a watermark?
DALL·E images carry C2PA Content Credentials metadata rather than a visible watermark, and that metadata is frequently removed when images are screenshotted or re-uploaded. So it can't be relied on for images already on the web.
How does verifai detect DALL·E images?
verifai reads each image's embedded C2PA content credentials on-device. DALL·E stamps these in, so credentialed DALL·E images are flagged reliably; stripped images show as "unverified." Text uses an on-device heuristic, and a hosted image deep-scan is planned.