Codex safety checks allegedly make sessions eligible for reinforcement learning despite opt-outs

A post argues that making good faith a cultural norm creates openings for exploitation. In reply, another user alleges that Codex’s “command safety” or “auto-approver” counts as a safety evaluation, making an interaction eligible for reinforcement learning even when the user opts out of contributing data. The reply further claims that responding to a command-approval request submits the session for training, not just reinforcement learning.