A note on moral frames, not moral defaults

Every AI system makes moral judgment calls, whether or not anyone admits it. What to refuse, what to flag, what counts as harm in an ambiguous case — these are moral decisions, and today they’re almost always made once, by the vendor, and shipped to everyone as a silent default.

We think that’s backwards. A moral frame is not a technical parameter to be optimized away; it’s a position, and different users hold different positions in good faith. So instead of baking in a single frame, we ask the system to reason from one the user has stated, and to make that reasoning inspectable rather than implicit.

This is still early work. The hard part isn’t collecting a stated frame — it’s keeping the system’s behavior consistent with it across edge cases the user never anticipated. That’s the problem we’re spending most of our time on right now.