Notes on building AI that defers to you

“Aligned to your values” is an easy phrase to write and a hard one to build. Most systems that claim it actually mean something narrower: aligned to a set of values the vendor picked on your behalf, with a thin layer of customization on top.

We’ve been experimenting with a different default: ask the user to state a worldview and a moral frame explicitly, and have the system reason from that declaration rather than from a house default. It’s slower to onboard — there’s no invisible answer already loaded in — but the reasoning that follows is legible in a way that default-aligned systems rarely are.

The open question we’re still working through is how to do this without asking every user to write a philosophy essay before their first query. Early answer: start from a small set of concrete questions, infer the rest, and let the user correct it as they go. More on this as the approach matures.