OpenAI announced yesterday that it deletes your data and tracks patterns across your interactions, in the same post, and both are true at once.
The bundle is Zero Data Retention (ZDR), which discards prompts and responses after processing, plus a preview of “Private Safety Processing”: automated systems that spot risky patterns spanning multiple interactions without a human ever seeing the content. When something trips, OpenAI receives a signal describing the type of activity, not the conversation.
I read it twice, because those two create a tension the announcement doesn’t acknowledge. You cannot detect a pattern across interactions without carrying something between them. Whatever counter, score, or embedding that detector maintains was derived from content the retention policy says is gone. The transcript is deleted. What the transcript taught the system is not.
That distinction, between retaining a conversation and retaining its influence, is the thing I keep circling in my own work, and it isn’t what deletion guarantees currently cover. Retention policies are written over stored records. Almost nothing is written over derived state.
What this doesn’t establish: I have no view into OpenAI’s architecture and this is not a claim their system leaks anything. The announcement is a product blog post dated 19 August, with the technical detail deferred to a whitepaper promised for September. That whitepaper is where the actual question lives, which is how much a “risk signal” reveals about the stream it was computed from.
Until then it’s a well-posed gap sitting in public, which is more than we had last week.
Source:
https://openai.com/index/offering-zero-data-retention-for-frontier-models/