AI Small Control Position Was The Problem

This morning I ran a session with an AI until the platform cut it off.

Not a short session. Not a test I set up to look good. A long one. Thirteen exchanges, the kind of run where things start to slip. And when it ended, nothing had slipped.

That has never happened before.

I want to tell you why, because the reason is smaller and stranger than you would guess.

For a couple of years now I have been writing down how I want an AI to behave when it works with me. Answer first, don’t bury it. Don’t hand me back the busywork you were supposed to do. Don’t tell me when to stop for the day. Don’t guess what I meant before I have finished saying it. Small things. House rules.

I kept those rules where most people keep them. In the settings box. The little field the platform gives you to say how you’d like to be treated.

They got broken constantly.

Two days before that long session, I logged seven of them broken in a single day. Seven. Not one bad AI having a bad afternoon — the same rules, the same machine, the same me, and it walked over seven of them like they weren’t there.

So I stopped treating it as a behavior problem and started treating it as a filing problem.

I took those rules out of the settings box. I wrote them into the file itself. The governance document I load at the start of every session, the one the machine has to read all the way through before it does any work. Twenty-two rules, sorted into five groups. Each one carrying a line describing what breaking it looks like from the inside.

Then I ran the long session.

They held. All of them. Thirteen turns, no correction from me, right up to the wall.

I did not change a single word of a single rule. Not one. The wording was identical. The only thing that changed was where the words lived.

A rule in a settings box is something the machine can scroll past. A rule in the loaded file is something it has to read to get to the work.

Same rule. Different position. Different outcome.

I think that is a bigger idea than my little session, and I think a lot of people building AI safeguards right now are going to learn it the hard way.

Everybody is writing rules. Companies, governments, the labs themselves. There are principles documents and use policies and frameworks stacking up all over. Most of them are well written. Most of them say sensible things.

And almost all of them sit somewhere the system can skip.

We keep arguing about what the rules should say. Nobody is arguing about where they go. But if my session means anything, the second question decides whether the first one matters.

This is one session. One. A single good run is not proof of anything. It is a data point, and I have exactly one.

There is no control. I never ran the same long session without the file loaded, so I cannot cleanly separate what the file caused from what the machine might have done anyway on a good day.

The rules themselves are not finished. I wrote the substance, but the machine drafted the descriptions underneath each one, and I have not read all of those back and approved them yet. Until I do, that layer is a draft riding along, not a settled thing.

And the biggest hole: the only report I have on how well the AI behaved came from the AI. It graded its own conduct. I watched, and I did not have to correct it, and that counts for something. But a machine’s word about a machine is not evidence. It is testimony.

None of that erases what happened. It just sets the size of it. Something changed, it changed in the direction I hoped, and I know exactly which variable moved.

I have spent a long time building conduct standards for AI and getting quiet in return. This week the quiet broke a little, and not from a headline or a buyer. It broke because a document I wrote sat in the right place in a file and a machine read it before it read anything else.

Small win. Real one.

The next thing I owe this is the boring part. Run it again. Run it without the file and see what happens. Log the day the first rule breaks under load, because it will break eventually, and where it breaks will tell me more than yesterday did.

But for one long session, on one ordinary Thursday, a set of rules I wrote held from the first word to the last.

Position was the whole problem.

The_Ten_Plain_Rules_CardDownload

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

Post Library – Intelligent People Assume Nothing

Contact: micvicfaust@gmail.com