OpenAI seems to be silently re-enabling model training on Codex chats

500 points · 149 comments on HN

Points and comments are a snapshot, not live.

Users report OpenAI silently re-enabling data sharing for model training in Codex chats.

A user at a company paranoid about IP leakage noticed that the setting to disable data sharing in Codex had been silently re-enabled for them and coworkers. The setting, called "Improve the model for everyone," can be checked at chatgpt.com/settings/data-controls. A linked HN thread and edit point to prior reports of the same issue. The post asks if this is isolated or widespread.

What commenters are saying

Multiple commenters confirm similar experiences, noting OpenAI also moved or hid usage stats in the menu. A camp argues that trusting checkboxes for security is naive: companies should run their own inference or use cloud gated models. Others call it an ethical issue regardless of cost. Some note that advanced account security with hardware keys locks data sharing off, though an email indicates OpenAI may still train on Daybreak Blue session content even with that setting off. One commenter dismisses it as likely a software bug, not malice.