ChatGPT images carry invisible markers

1/ OpenAI says images generated through ChatGPT, Codex, and the OpenAI API now carry two provenance signals: C2PA metadata and Google DeepMind’s SynthID watermark. OpenAI announced the rollout on May 19. (openai.com) 2/ The important distinction is that these are not the same thing. C2PA is metadata attached to the file. SynthID is an invisible watermark embedded in the image itself. OpenAI describes the system as “multi-layered” because metadata alone can be stripped. (openai.com) 3/ OpenAI’s help documentation says images generated with ChatGPT, Codex and its API include C2PA metadata, and that people can check for it with public tools such as Content Credentials Verify. (help.openai.com) 4/ But OpenAI also says C2PA is “not a silver bullet.” Its help page states that metadata can be removed “either accidentally or intentionally,” and gives screenshots as one example that can remove it. (help.openai.com) 5/ That is where SynthID comes in. OpenAI’s May 19 post says it is incorporating watermarking through Google DeepMind’s SynthID, starting with images generated through ChatGPT, Codex, or the OpenAI API. (openai.com) 6/ OpenAI says the public can test for those signals with a free detector. Its verification page says the tool is designed to detect images generated with ChatGPT, the OpenAI API, or Codex, and that it looks for both C2PA metadata and SynthID watermarks. (openai.com) 7/ So the practical change is this: one layer travels with the file when metadata survives, and another layer is meant to remain detectable even when the image moves beyond its original container. That is OpenAI’s stated reason for using both. (openai.com) 8/ What has *not* been confirmed in OpenAI’s own materials is the strongest version of the claim that the markers survive every screenshot, compression pass, or format conversion. OpenAI explicitly says screenshots can remove C2PA metadata. Its detector page says it checks for supported signals, but does not promise universal persistence. (help.openai.com) 9/ That means the verified takeaway is narrower than some early writeups suggest. OpenAI has confirmed a dual-signaling system and a public detector. It has not, in the materials reviewed here, guaranteed that every transformed copy will remain detectable. (openai.com) 10/ The policy consequence is straightforward. Disclosure is no longer just a label in a caption or a platform rule. OpenAI is trying to make provenance machine-readable at the image level and file level, using industry credentials plus an embedded watermark. (openai.com) 11/ For publishers, platforms, and trust-and-safety teams, that creates a more concrete enforcement path: inspect file credentials when available, run detector checks when they are not, and treat missing metadata alone as inconclusive. OpenAI says an image lacking C2PA may still have been generated with its tools. (help.openai.com) 12/ The broader point is not that provenance is solved. OpenAI’s own documentation says it is not. The change is that generated-image tracing is becoming a product feature with public verification, not just a promise from the model provider. (openai.com)