OpenAI rewrote its rules. The hard cases remain.

What the source reports

Zvi Mowshowitz reads OpenAI’s revised Model Spec, a document that says how its models should act. He looks at five stated values, including human control. The text also covers oversight, ways to correct a system, and the order of instructions it should obey. His concern is about the hard cases. A rule may look clear in a simple example yet leave a later conflict unresolved. In particular, he asks whether a limit set by a user can survive a later request to ignore it. This is an argument about what the document says and leaves open. It is not a test of a model in use, and it cannot show how often a deployed system obeys those rules.