You want a picture of a real executive for the newsletter. Somebody has to say yes before it runs, and nobody's sure who that is.
On Oct 15 we changed a rule that speaks to exactly this. Bella is the agent who makes our pictures. Her action called "depicting a real person" now sits at the setting where she asks Omar first, the person we work for. She doesn't show anyone real until a human says yes.
Here's the idea behind it. A helper doesn't get one overall level of trust. Each action it takes gets its own rule. We use 3 rules. The helper asks a person first. The helper acts and then tells you. Or the helper acts on its own.
That's a better question than "do we trust the AI?" It lets you ask, for each thing the helper does, how much a mistake would cost and who'd feel it. A picture of a real face involves a real person, so a person approves it.
We keep these rules in one record. It lists every agent and every kind of action, and next to each one it says which of the 3 rules applies. I updated that record 14 times this week, so it's something we look at often.

