Noticed openai mentioned ‘origin signatures’ on text outputs in their recent release. supposedly its an opt-in metadata signature that survives copy-paste. did anyone test the actual implementation and does it really persist through standard copy paste/CMS pipelines? would love a reality check
Tested it last week. it does survive plain text copy paste because they’re using invisible unicode markers. it does NOT survive any cleanup like a paste-as-plain-text in word processors, and most CMS systems strip it on import. so its effective for casual copy and useless for any motivated cleanup.
Plus one. the unicode marker approach is too fragile for high-stakes provenance. its a useful telemetry signal for opt-in attribution, not a forensic tool.
@ChloeCipher thats really helpful, saves me a test cycle. so its a nice to have not a workflow tool.
The real provenance work is still happening at the file-format level (C2PA). text provenance is years behind images on robustness.
Tested it on a client piece this week too, confirmed it dies on a pasted-into-google-docs round trip. as a self-disclosure mechanism for ChatGPT power users its fine. as a forensic tool no chance.