The same origin model is a one-size-fits-all cage.

Very simple to reason about and create, but a poor fit for any given real world situation.

You can imagine a system that constructs a bespoke policy cage to allow just what a user actually wants in that situation.

Only composed experiences that fit within that cage are allowed to run.

Humans could construct the cage themselves by constructing an intricate policy matrix.

However, this is time consuming and finicky.

LLMs can be used to rough in a "good enough" cage with reasonable assumptions about characteristics almost everyone would agree on.

Then from that good enough starting point a human could wrinkle and complicate the rules to better tune them for them.

The power of the LLM thus sets not a ceiling on what is possible, but a floor.

A human could have done all of the work, but that likely would be below the cost/value curve and not viable.

The LLM's starter answer provides a floor that lifts significantly more use cases above that viability threshold.