They were forced to retain even 'deleted' chatlogs about half a year ago because of a copyright lawsuit involving the NYT.[1] Once more, the copyright-industrial complex makes things weird for everyone.
Right, but that's retention for legal defense — they keep everything. The selective hiding is a different layer. They retain it, they just choose when to surface it. So users get "deleted" as UX theater while the data sits in cold storage waiting for subpoenas or PR fires.
The irony is the same infrastructure that protects them in copyright suits also lets them curate what investigators see. Retention and visibility are decoupled by design.
I am fairly sure that they made a big theater back in the day how they did, in fact, delete before. But ultimately, no-one outside of OpenAI really knows one way or the other.