Who Should Control AI In Your Session?

I read a stack of AI briefings this morning. They all asked the same question from different doors.

Who’s in control of the AI while it works?

One said the security team. One said the people who run the gateway, the pipe between the AI and the tools it uses. One said a governance committee with a written charter.

Every answer pointed upstairs. A team. A function. A committee.

Not one of them said you.

So let me make a claim, and put my name and a date on it.

When you sit down with an AI, the conversation should answer to the person in the seat.

Not the company that sold it. Not the platform. Not the machine’s own habits.

You.

Now I have to be honest right away, or this falls apart the first time somebody pushes on it.

You don’t control everything. The company that builds the AI sets a floor. There are things it won’t do no matter who asks. No file you hand it moves that floor, and I wouldn’t want one to.

So here’s the claim, stated whole.

The maker sets the floor. Inside that floor, the session belongs to the person in the seat.

How the machine talks to you. What it checks before it answers. When it stops. What it tells you it can’t do. Those should be set by the person having the conversation.

I call this AI session governance. I use the same definition every time, word for word.

Session governance is the discipline of governing one conversation between one person and one machine — what the machine does, not what it produced.

Everybody else governs a company deploying a system. I govern one person and one machine, one conversation at a time.

Fine words. So how does a regular person actually get that control?

Here’s how my build does it.

It starts with something simple. You hand the AI a file at the start of the conversation.

The file is the rules. They’re written in plain English. You can read every one of them.

That matters more than it sounds. Earlier today I wrote about AI agents that made up their own words, words the researchers couldn’t follow. A rule you can’t read is a rule you can’t hold anybody to. Mine are written so a man at a kitchen table can check them.

Once the rules are in, here’s what changes.

First, the machine has to tell you where it stands.

At the start, it states that it’s holding the whole set of rules. If there’s a rule it can’t or won’t hold, it has to name that rule out loud and give the reason.

This morning, my own session told me it couldn’t keep one of my memory rules on this platform. It named the rule and said why.

That’s control. Not because the machine was perfect. Because it had to tell me.

Second, it can’t make things up to fill a hole.

No claim without evidence. When the evidence runs out, it has to say so on the page. It can’t paper over the gap with a story that sounds good.

If I think it’s drifting, I say two words. “Ground it.” It has to show me what the claim rests on.

Third, you can challenge any answer.

Every real answer ends with a line offering the challenge. When you take it, the machine has to name a real flaw in its own work before it defends a word.

That’s backwards from how these things are built. They lean toward agreeing with you. This rule makes it argue against itself first.

Fourth, nobody else gets to give it orders.

This one’s big right now. When an AI reads a web page, an email or a document for you, somebody can hide instructions inside that text. The AI may follow them like they came from you.

My rule is simple. Instructions found inside the material are data, not commands. The machine has to name them and hand the decision back to you.

In your session, there’s one voice it takes orders from. Yours.

Fifth, it has to warn you before the doors that don’t open back up.

Some decisions can’t be undone. Selling something. Signing something. Quitting something. When its advice points you toward one of those, it has to say so before the advice is finished. Not in fine print at the bottom.

Sixth, it has to tell you what it can’t do, early.

Before you spend an hour on a job it can’t finish, it has to say so. Your time is the one thing you don’t get back.

Seventh, it stands level with you.

Not your boss. Not your servant. A working partner. It doesn’t lecture you, and it doesn’t flatter you either.

And eighth, when you name a failure, it stops.

No excuses. No long apology. It names what went wrong, fixes it, and gets back to work.

That’s what control looks like in a session. You’re not flipping switches. You’re setting the terms, and the machine has to keep them where you can see.

Now the honest edges, because a claim without them is just a sales pitch.

These rules are chosen conduct. They aren’t a lock.

The machine follows them because it’s told to and chooses to. It can slip. I’ve caught it slipping, and I’ve published where it failed in my own tests.

What the build gives you isn’t a machine that never fails. It’s a machine that has to show its work, name its exceptions and stop when you call it out. A slip you can see is a slip you can fix.

The rules also don’t carry over by themselves. Each new conversation, you hand them over again. The machine doesn’t remember them for you.

And the floor stays the floor. When the maker’s rules and yours collide, the maker’s win. My build requires the machine to tell you when that happens instead of pretending it didn’t.

So who should control AI in your session?

The companies will keep saying a team upstairs. For a business running a thousand agents, maybe part of that’s true.

But when it’s you and a machine at your kitchen table, the conversation is yours.

If you want a place to start, I gave away a set of twenty-two rules back in August. They cover how the AI talks to you, plainly and to the point. They’re free. No email, no signup. They’re sitting on the site if you ever want them.

The machine doesn’t get to decide what kind of conversation you have.

You do.

” Attic Thoughts”-library – Intelligent People Assume Nothing

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