OpenAI has embedded C2PA Content Credentials in DALL·E-generated images since 2023, and ChatGPT's current image generation (branded internally as "gpt-image") continues this. When we decoded a real file generated through OpenAI's pipeline, the manifest listed "OpenAI Media Service API" as the claim generator, a software agent of "gpt-image", a chain of recorded actions (created → converted → watermarked), and a cryptographic ES256 signature — a full, signed provenance record, not just a flag.
Unlike Midjourney's plain-text prompt approach, OpenAI's images carry a structured, cryptographically signed C2PA manifest (technically a JUMBF box containing CBOR-encoded data). Based on direct inspection, it typically includes:
c2pa.created, c2pa.converted, and c2pa.watermarked.unbound, each naming gpt-image as the software agent.Because the manifest is both present and signed, platforms like Instagram, Facebook, and TikTok — all of which read C2PA on upload — detect OpenAI-generated images reliably. This is the main reason ChatGPT-generated images get labeled consistently, while Midjourney images (no C2PA at all) often slip through metadata-based detection.
The signature proves the manifest hasn't been altered since OpenAI signed it — it doesn't make any claim about the image's content being true, accurate, or non-misleading. It's a provenance record, not a fact-check.
Use AiEraser's View mode to decode the actual manifest on your own file — claim generator, every recorded action, and signature algorithm, exactly like the breakdown above — then Remove mode to strip it via in-browser canvas re-encoding before you share or publish.
Primary source: direct JUMBF/CBOR decoding of a real OpenAI-generated image file, cross-checked against public reporting that OpenAI added C2PA Content Credentials to DALL·E outputs in 2023 in partnership with the Content Authenticity Initiative.