Does ChatGPT, Codex and the OpenAI API (images) add Content Credentials?

If you publish ChatGPT, Codex and the OpenAI API (images) output in the EU

Marking outputs in a machine-readable way is the duty of the tool’s provider under Article 50(2) of the AI Act. If you use the output for work and it is a deepfake, the visible disclosure is yours (Article 50(4)), whatever OpenAI adds to the file. What that means for your content: AI images.

Scope and limits

• The article says coverage varies by product, model, export path, file type and creation date.
• Its content-type table lists images and audio; it has no video row.
• It does not name gpt-image or DALL·E; it says “ChatGPT, Codex, and the OpenAI API”.
• OpenAI’s help centre blocks scripted fetches, so the text was read from an Internet Archive snapshot of the same page taken on 20 September 2026.

If the credentials are removed

C2PA data travels inside the file, so a re-encode, screenshot or upload that does not carry it across leaves an image with no credentials. A watermark or fingerprint can help find a record again, if one was stored (see the C2PA specification’s soft bindings). Proofcan is independent of OpenAI: for images you mark with Proofcan it writes a C2PA manifest, an invisible watermark and a registry record. It does not change or vouch for how OpenAI marks its own outputs.

Check a file yourself

The Content Credentials checker reads the C2PA data in an image in your browser and shows who signed it, the generating tool and whether the signature still matches. It reads what a file declares; it does not detect AI from pixels, and a file with no credentials proves nothing either way. See why metadata gets stripped.