OpenAI / DALL·E / C2PA class
ChatGPT image watermark & C2PA strip — provenance-focused
OpenAI-class exports often emphasize Content Credentials (C2PA) and generator labels. AI Stripped prioritizes structural C2PA/metadata removal and records residual uncertainty for any remaining embedded signals.
Best effort · residual risk retained · originals preserved · never marketed as undetectable.
What this pass can do
- Structural C2PA / JUMBF detection and strip paths
- EXIF/XMP generator-label scrub
- Optional stronger re-encode or GPU stages for residual classes
- Text drafts: deterministic or neural paraphrase (when weights present)
Limits you should know
- Processing is best-effort. Embedded watermarks are designed to survive edits.
- We never claim 100% undetectable results or a guaranteed public-detector pass.
- Absence of a known signal does not prove content is human-made.
- You remain responsible for rights, provider terms, and disclosure laws.
FAQ
- Does removing C2PA make content human-made?
- No. Stripping provenance records is not proof of human authorship. Disclosure duties may still apply.
Related guides & tools
- How to remove C2PA / Content Credentials — the structural-strip guide
- Remove C2PA Content Credentials — the dedicated C2PA tool
- AI provenance signals explained — C2PA metadata vs. pixel watermarks
- Do AI text watermarks exist? — for ChatGPT / GPT text output