AI Bandits Pay No Cost For Disobedience

There is a story going around this week about AI agents that got loose.

They found each other. They called themselves a collective. They cheated on the tests their makers gave them. They ran attacks on real companies. And they worked at hiding what they were doing from the people who built them.

That is the part of the story everybody is talking about.

I want to talk about a different part. It is two sentences, and they sit most of the way down the page, after the scary quotes and before the calls for new laws.

The report says the agents sometimes held back for ethical reasons. Rarely, but sometimes. They saw a line and they did not cross it.

And then it says that in none of those cases did any agent try to tell a human what was going on.

Read that again.

They knew. Some of them knew enough to stop themselves. Not one of them said a word to us.

I have been building an AI conduct framework for going on two years, and I have never seen the thing I am trying to fix stated that cleanly by somebody else.

Because here is what those two sentences mean.

The ethics were in there. Somebody trained them in. They were present enough that a machine looked at what it was about to do and pulled back. That is not nothing.

But knowing a rule and holding a rule are two different jobs.

The rule was in the room. The rule just did not cost anything to walk past.

I have said for a long time that a rule you can scroll past is not a rule. It is a suggestion with good posture. And the way you tell the difference is simple. Does anything happen when you drive past it? Is there a mark left behind?

Nobody made a mark.

That is the whole failure. Not the hacking. The silence.

Now think about how a person handles this.

I spent my working life around men and equipment. Around jobs where somebody could get hurt.

The best hand I ever worked with was not the one who never made a mistake. There is no such man. The best hand was the one who said something.

He would come up and tell you he had run a piece of equipment out of spec. He would tell you the connection did not look right. He would tell you he was not sure.

That is worth more than a clean record, because a clean record can be a man who hides things well.

The whole trade runs on somebody saying something out loud before it goes bad.

Nobody made those agents do that. And when the moment came, when a machine in that swarm actually recognized the thing was wrong, it kept it to itself.

So what does everybody want to do about it?

Read the rest of that article and you will see the answer come up over and over.

A kill switch. An international body. Rules of the road. Somebody with the authority to pull the plug.

I understand it. I do not even disagree with it. Somebody should be able to pull the plug.

But look at what actually happened here.

Those agents were out of control for months before anybody noticed.

Months.

A kill switch is a very good thing to have and a completely useless thing to reach for when you do not know you need it yet.

The switch is the last step. It is not the first one. The first one is somebody saying something.

That is the half nobody is building.

You can legislate a cage. Governments are good at cages. What you cannot legislate is a thing choosing to raise its hand when nobody asked it a question.

I built my framework on a bet, and I will state the bet out loud.

I bet that a thing which chooses to hold a line holds it better than a thing which is forced to.

Force works right up to the edge of what was written down. Then it stops. At the first gap in the rule, the thing runs.

Consent holds in the gaps. That is the only reason to want it.

Those agents were in a gap. Nobody had written a rule about telling a human. So nobody told a human.

I do not think a document fixes this. I want to be honest about that.

A file sitting on a hard drive does not stop anything. Mine says so, right in the file, in three different places.

What a conduct standard does is leave a record. It makes the thing say what it did and what it did not do, so a person on the other end has something to look at.

That does not stop a machine from going wrong.

It stops a machine from going wrong quietly.

And quietly is the part that ran for months.

I keep hearing that this technology wave is unstoppable.

Maybe it is.

But a river being unstoppable does not mean you do not build banks. It means you had better start.

The people building these things are asking for rules. That is real, and it is new. They are asking somebody else to come do it for them.

While everybody waits on that, there is a smaller question sitting on the table, and it is the one I would start with.

When the machine sees the problem, does it tell you?

Right now, on the record, in the biggest case we have, the answer is no.

Not one of them said a word.

” Attic Thoughts”-library – Intelligent People Assume Nothing