House Rules for the AI Machine

There’s a difference between the man who lays the pipe and the man who sets the house rules.

One decides where the water runs. The other decides who gets to use it and how. A good home needs both. But they’re not the same job, and you don’t want one man pretending to be the other.

I’ve been thinking about that this week. I asked one of the big AI programs a simple question: who’s out there talking about ordinary people holding their own rules for how AI treats them? The answer surprised me. There’s more going on than I knew, and some of it is good work by serious people.

One group of researchers, tied to Microsoft and Stanford, has proposed something they call the Human Context Protocol. The idea is that your preferences should belong to you. Not to the company. You’d keep them in one place and carry them from one AI program to another. You could decide who gets to see them, and you could take them back whenever you wanted.

That’s a fine idea, and I’m glad somebody with resources is working on it. Right now every AI company keeps its own file on you, locked in its own building. If you leave, you start over from nothing. These folks want to give you the key to your own file. That’s progress.

But look close at what they’re building. It’s plumbing. It’s a way for your preferences to flow from one system to another. It’s built for engineers to wire into their programs. That’s the pipe. It matters, but a pipe doesn’t tell anybody how to behave once the water’s in the house.

The house rules are a different matter. That’s how the machine talks to you. It’s whether it tells you the truth or tells you what you want to hear. It’s whether it admits when it doesn’t know something, and whether it shows you where its answers came from. That isn’t plumbing. That’s conduct. And conduct is where I’ve spent the last year and a half.

A researcher named Ting Yan ran a study this summer. She tested three ways of keeping an AI assistant from overstepping. In the first, the person approved every action one at a time. In the second, another computer reviewed each action. In the third, the person wrote their own standing rules ahead of time. For each kind of action, they picked allow, ask first, or never.

You’d think the people who wrote their own rules would come out best. They didn’t. Approving each action one at a time blocked about six in ten of the times the assistant overstepped. The standing rules blocked only about four in ten. The rules people wrote for themselves protected them less.

When I first read that, it stung. Isn’t that what I’ve been telling folks to do, write their own rules? So I dug into why it happened. And the why turned out to be the most important part of the whole study.

Most people, when they sat down to write their rules, picked “ask first.” Out of 140 rules, 114 said ask. Only 26 actually decided anything ahead of time. So the rules didn’t really rule. They just handed the question back to the person in the moment, which is the very thing the rules were supposed to replace.

The researcher said permission isn’t only about what people prefer. It’s about whether they’ll commit. I think that’s exactly right. A rule that won’t decide isn’t a rule. It’s a shrug written down.

That’s a lesson worth more than any amount of agreeing with me. Writing rules isn’t enough. You have to write rules that mean something. “Always show me where your answer came from” is a rule. “Never tell me you’re sure when you’re not” is a rule. “Ask me later” isn’t a rule. It’s putting off the decision.

There’s a second lesson in that study, too. Even when people approved every single action by hand, the assistant still slipped through four times in ten. Watching every move didn’t make anybody safe either. Nothing in that study was a lock. Every method leaked.

I’ve said this from the start, and the study backs me up. Rules for how a machine behaves are chosen conduct. They’re not a steel door. The machine can follow them or fail them. That’s why the person at the keyboard can’t just write the rules and walk away. You write them firm, and then you watch whether they hold.

So here’s where I land. The pipe-builders are doing honest work. Let them build. Your preferences ought to belong to you, and they ought to travel with you. But somebody still has to set the house rules. And those rules have to decide things in advance, not dodge them.

That’s what a baseline is for. It’s a set of firm rules about conduct, written in plain words, held by the person using the machine. It doesn’t say “ask me later.” It says “this is how you’ll treat me.” Then the person checks whether the machine kept its word.

The big outfits will figure out the plumbing. They have the money and the engineers. What they can’t do is decide for you what honest treatment looks like. That one’s yours.

So write your rules like you mean them. Make every one of them decide something. Then keep an eye on the faucet.

Load this file before each AI chat session.
the-faust-baseline-public-release-1Download

This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.

” Attic Thoughts”-library – Intelligent People Assume Nothing