ChatGPT image watermark and C2PA strip
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.
Results are best effort. Residual risk is retained, originals are preserved, and nothing is 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