PRODUCTION SCENARIO
A consumer language-tutoring app on the OpenAI API serves 200,000 end users and wants OpenAI's abuse monitoring to single out one bad actor rather than throttle the whole organization. Its privacy policy bars sending user emails to processors beyond what is needed.
What should each request carry to identify the end user?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: A safety_identifier set to a one-way hash of each end user's account ID
A safety identifier should uniquely identify each user, hashed from the username or email so no identifying information is sent, with a session ID standing in for users who are not logged in. One shared organization ID cannot distinguish the individual abuser, which is the whole point of sending it. An IP address is identifying data and abuse monitoring reads the safety_identifier field rather than transport metadata. A raw email identifies the user and breaches the app's own privacy policy.