One of the researchers was using the product. They train on your private interactions unless you explicitly opt out in the settings.

This is speculation right now. The idea would be that if someone used the product and granted training rights, which is the default for many subscription levels, then some knowledge would have been imparted into the general weights of the new model.

oAI has made clear they did not specifically pull in any user data to context for this run.