You update your custom instructions to ask for shorter answers, then reopen a chat from last week and continue it. The replies there are still long. Why?

The principle — a settings change is a change to what happens next, and an existing conversation already carries the context it started with.

Why the key is correct — OpenAI states directly that updates to your instructions are reflected only in future conversations, and that you can edit or delete custom instructions at any time for future conversations. An old chat continuing to behave the old way is the documented behaviour rather than a fault.

Why the others are wrong — the change did save, and the settings pane will show the new text. No sign-out is required, because updates apply immediately to chats you start from that point. And instructions are not tied to a model, so the model an old chat was using is not what is driving this.

Remember this — after changing your instructions, test them in a new chat. Judging the change inside an old conversation will tell you nothing, and the same rule is why OpenAI notes that removing instructions from previous conversations means clearing those conversations.

Sources — OpenAI's ChatGPT Custom Instructions article.