Hacker News Reader

These are two different control problems.

Training governs what a model tends to say. Authority governs what is allowed to be acted on.

You can’t pre-block bad advice, but you can pre-block unapproved financial or contractual actions.

That’s the scope.

They are the same problem, since the actions are saying.

"AI-generated messages making commitments no one explicitly approved. Refunds implied. Discounts promised. Renewals renegotiated."

And if you can't block bad advice, you have a bigger problem since you cannot block resultant contractual action by the recipient e.g. termination.