OpenAI layers C2PA metadata, Google DeepMind SynthID invisible watermarking, and a public verification tool to make AI-generated images more traceable.
Key Takeaways
SynthID watermarking is now embedded in images from ChatGPT, Codex, and the OpenAI API; it persists through screenshots and format changes where C2PA metadata can be stripped.
OpenAI became a C2PA Conforming Generator Product, enabling platforms to reliably read and pass along provenance metadata attached at creation.
The two layers are complementary: C2PA carries detailed context, SynthID carries a durable signal when metadata is lost.
A public verification tool (preview) lets anyone upload an image to check for SynthID watermarks or C2PA Content Credentials originating from OpenAI.
If no signal is detected, the tool deliberately avoids a definitive negative conclusion, since provenance signals can be stripped.
Hacker News Comment Review
Commenters are skeptical about removal resistance in adversarial conditions; the main concern is that motivated actors building AI propaganda pipelines will quickly produce stripping tools.
One commenter pushed back, noting SynthID has been available for years with no reproducible public break yet, suggesting durability is better than critics assume.
Google keeping SynthID closed-source drew friction: auditing and independent adoption are blocked, and at least one commenter hinted at releasing an open-source alternative.
Notable Comments
@minimaxir: Questions closed-source lock-in and hints at a “feature complete” open-source invisible watermarking repo as an alternative.
@Arnt: Argues even a month or year of delay before a watermark is broken may be practically sufficient for platform integrity decisions.
