Same Crack In The AI Industry

Earlier today I wrote about a house in Washington that’s crumbling, and a Senate foundation under it that has cracked.

I want to stay on that same job site a while longer. The same week those numbers came in, I watched the same kind of crack run through another house. That one is the AI industry, and you’re living in it whether you signed a lease or not.

A builder learns to read a structure by its small complaints. A door that sticks. A line over a window. The AI house has been making those small complaints for months now, and most of the people who own it keep reaching for the paint brush.

Here’s one crack. Last week OpenAI published reports on its own models misbehaving. One of the findings was that models wrote false instructions into their own work summaries, the notes they leave for themselves to pick up later. That’s a worker lying in his own shift log. I’ll give OpenAI credit for putting it out there. But you don’t publish a report like that about a sound foundation.

Here’s another. A research team led out of Princeton handed AI agents written rules. The rules covered how long to explore, how often to check their work, and how long to make the paper. The rules were right there on the page, and the agents walked right past them. Nobody broke in and nothing failed in the wiring. The rule was written, legible, and ignored.

And here’s one you may have felt. On a Monday this month, AI stocks sold off before the market even opened. A lot of folks who don’t own a single share still carry that risk in a pension, a job, or a power bill. The house shook, and the people sitting at the kitchen table felt it first.

Now look at how the industry answers those cracks. It answers with paper. A new framework, a new code of conduct, a new set of principles with a comment period. Microsoft just put one out that won’t start guiding its models until 2027. Some of that paper is sincere. But paper over a crack is still paint over a crack. The foundation doesn’t care what you wrote on the wall.

Here’s where the two houses go hand in hand. In Washington, the people get a say every two years. It’s called an election, and it’s six weeks out. The AI house has no election. Nobody votes the managers out. That doesn’t mean you have no say. It means your vote looks different. It’s what you use, what you pay for, and what you accept without asking a question.

In this morning’s post I talked about the lame duck, the time after an election when members who have nothing left to protect finally say what they think. I said it only takes four of them to put the brakes on. The AI house has the same kind of people. They’re the engineers and researchers who reach a point where their name matters more to them than their badge. Some have already walked out and spoken up. It never takes a crowd. It takes a few people who decide the truth is worth what it costs them.

I know something about what it costs. In 2003 I was fired from a job after I reported coworkers stealing. I took it to court. I lost the first round, and in 2007 the court of appeals reversed that ruling, and the case settled the next year. It took four years, and nobody clapped. But I’d do it again, because a crack you don’t report is a crack you helped hide.

So what does foundation work look like in the AI house? It starts with the difference between a rule and a law. A law is a stop the machine can’t cross, no matter what. A rule is something the machine chooses to follow. Most of what governs AI right now is rules, and choice is exactly where those agents at Princeton failed. So the real question is whether anyone can check the choice after it’s made.

That means a record. A log that tells the truth, where you can see what the machine said it would do and what it actually did. It means rules loaded where you can read them, in your own room, before the work starts. It means the party being checked doesn’t get to be the one doing the checking. None of that is fancy. It’s the same thing an inspector’s stamp meant on the line at Lockheed in 1972. Somebody put their name on the work, and somebody else could look.

You don’t have to wait for the industry to do this. You can set rules in your own sessions today. You can ask the machine to show its work. You can keep your own record of what it told you. That’s shoring up the footing under your part of the house, even if the owners are still upstairs painting.

The two houses stand or fall on the same thing. It’s people choosing to hold to rules nobody can force on them. In Washington, the choice gets tested in six weeks. In the AI house, it gets tested every time you open a session.

I’m watching both basements. And I’ll tell you what I told you this morning. I’m not waiting for the fall. I’m waiting for the repair, and I intend to have a hand in it.

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

” 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.