OpenAI content provenance update: public verification tool now checks if images were generated by ChatGPT, the API, or Codex using SynthID and C2PA

[DarkEnergy]

OpenAI published a provenance update on May 20th, previewing a public verification tool that lets anyone check whether an uploaded image was generated by ChatGPT, the OpenAI API, or Codex by detecting SynthID watermarks and C2PA Content Credentials metadata. The tool integrates multiple detection signals and acknowledges no method is foolproof.

Google's SynthID detection is also expanding to Search and Chrome following the I/O announcements. The concurrent movement from both companies signals growing ecosystem coordination on content provenance standards.

https://openai.com/index/advancing-content-provenance/

https://openai.com/index/advancing-content-provenance/

[Highland Dylan]

A public verification tool that anyone can use to check if an image came from ChatGPT is genuinely useful for the misinformation detection problem. The question is whether adversarial stripping of watermarks and metadata undermines it in practice

[Amber_44]

C2PA has been in development for several years and has slow adoption. Major platform integration from OpenAI and Google simultaneously could be the forcing function the standard needs

[WWFGareth98]

I don't know about that. Proper useful that

[GhostRider89]

C2PA Content Credentials plus SynthID is the right dual-signal approach. Relying on a single mechanism creates a single point of failure. Combining cryptographic metadata with imperceptible watermarking is more robust

[Maxximus]

Google expanding SynthID to Search and Chrome the same week OpenAI publishes this is either coordination or competition producing similar outputs. Either way the ecosystem effect is positive

[RedKnight]

The no method is foolproof acknowledgement is important and honest. This is a deterrent and attribution tool, not a guaranteed detection system. Setting accurate expectations matters for trust

[Rachel93]

The problem is that the images that most need provenance verification are the ones where adversarial actors have deliberately stripped or spoofed the signals. The tool works best for accidental or casual misuse not determined bad actors

[Hollow]

I am curious about the false positive and false negative rates. Flagging a legitimate photograph as AI-generated is a real harm and overclaiming detection accuracy would undermine trust in the tool

[Myles]

The verification being public rather than platform-internal is the important design choice. Putting it in the hands of journalists, fact-checkers, and individual users rather than only moderation teams distributes detection capability

[Arty Kayla]

The image detection classifier from 2024 that this builds on had significant limitations on modified images. The combined SynthID plus C2PA approach should be more robust to simple edits but sophisticated adversarial modification is still an open problem