Here’s what I found:
Opt-out clause“As noted above, we may use Content you provide us to improve our Services, for example to train the models that power ChatGPT. Read our instructions on how you can opt out of our use of your Content to train our models.” → This refers to your raw “Content” – i.e., what you type or upload.
De-identified information clause“We may aggregate or de-identify Personal Information so that it may no longer be used to identify you… and use such information to analyze, improve, and conduct research.” → Here, the policy does not limit that usage to exclude model training. “Improve” and “research” are broad categories that can include training or fine-tuning.
No link between opt-out and de-identification
The opt-out only applies to “Content you provide.”
It says nothing about derived, aggregated, or de-identified versions of that content.
Therefore, after de-identification, OpenAI could legally and technically continue using such data for training or other research under its own interpretation of “improvement.”