Made Out Of Time, That I Already Spent

A document came across my desk this week. It said AI agent governance has stopped being advice and started being law.

That got my attention. So I went and read the things it was pointing at.

The first was from the Financial Stability Board. That is the group that coordinates financial regulators around the world. Real body. Real weight. They put out a report on June 10th of this year about how banks should handle AI that acts on its own.

The report is good. It says the plain thing out loud. When you give a machine that much room to move, trouble can show up faster than a person can stop it. It says that undoing what one of these systems has already done can be hard, and sometimes impossible.

That is honest. That is somebody looking at the problem straight.

Then I looked at what kind of document it was.

It is a consultation report. That means they wrote it down and asked the industry what they thought. Comments were due July 22nd. The final version is expected in October. The twelve items in it are not rules. They are called sound practices. The FSB itself says they are not binding.

So it is a draft. Out for comment. Nobody has to do any of it.

The second one was the United Nations. They set up a scientific panel on AI back in August of last year. Forty members, three year terms, appointed by the General Assembly. Also real. Also serious people.

I went and read the resolution that created it. The panel’s job is to produce scientific assessments. And the resolution uses one particular word to describe them.

Non-prescriptive.

That word did not fall in there by accident. Somebody negotiated it. It means the panel can tell you what is happening. It cannot tell you what to do about it.

Now here is the thing.

I do not think either of those is a scandal. A scientific panel should not be writing law. That is what makes it a scientific panel. And a comment period is how a real rule gets built. That is the process working, not the process hiding.

But the document I was handed did not say any of that.

It took a draft out for comment and called it Recommendations. It took a panel that is barred from prescribing and called its work a formal risk tier. It cited a few other documents I could not confirm exist at all. And in the middle of it, there was a list of control numbers. AGT-001. AGT-003. Numbers that belong to nobody except whoever wrote the page.

Then a checklist. Run a gap assessment. Update your risk register. Brief your board.

That is not news. That is a sales sheet wearing a news jacket.

And it only works if the real standards stay soft. If the FSB had written a hard rule, you could pass or fail it, and nobody could sell you a shortcut. Soft is where the money is.

But I want to get past the sales sheet, because that is the small part.

Step back and look at the whole field. There are a lot of documents now. There are panels and consultations and frameworks and landscape reports. Everybody has arrived at the same subject at the same time.

And I cannot find one of them that says who is on the hook.

An AI agent takes an action. It moves money, or sends a message, or agrees to something. It does that on behalf of a company, using access a person handed it. Then it goes wrong.

Who answers for that?

Not one of these documents says. They talk about permission boundaries and approval gates and logging the chain. All of that is a fence. Fences are fine. I have built a few. But a fence tells you where the animal is supposed to stay. It does not tell you whose animal it is when it gets out.

Everybody wants to be on record as having addressed this. Nobody wants their name on the line that says we will stand behind it.

That is the whole picture. It is not a coverup. Nothing is hidden. It is an eight month parade of documents, every one of them written so that the writer does not have to build it and does not have to stand by it.

I have been saying something close to this for a year and a half, from a different direction.

You cannot fence a thinking machine into behaving. The fence holds right up to the edge of what somebody thought to write down, and then the machine walks through the gap. Every gap is a place nobody imagined. There will always be gaps.

The other way is to build something the machine can choose. Conduct instead of a cage. A door instead of a wall.

I will be straight with you about where that stands. It is hard to prove. You cannot buy it. You cannot audit it. You cannot sue over it. I have run it against other AI systems and gotten one clean measurement out of it, one test, one time. That is thin and I am not going to dress it up.

But I would rather have a thin honest measurement than a thick document that signs nothing.

So here is what I keep doing. I write it down. I date it. I leave the failures in, because the failures are the part that proves somebody was actually working and not just publishing.

Someday an agent is going to do real damage, and people are going to go looking for who saw it coming and what they said. That record is the only thing I have that nobody can build faster than me.

It is made out of time. And I already spent the time.

You will see clearly now.

How-To-Use-The-Working-Manner-Layer

TFB-Working-Manner-Layer (3)Download

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