When The Tires Hit The Rumble Strip

Have you ever driven a long road at night?

The first hour goes fine. You are sharp. You are awake. Then the miles start to pile up, and something changes so slowly you do not feel it happen.

You do not feel tired. That is the trouble. You only find out when the tires hit the rumble strip on the shoulder. By then you have already drifted.

A team at George Washington University just showed that AI can drift the very same way.

Two researchers there, Neil Johnson and Frank Huo, studied what happens inside an AI as a conversation goes on. They found something every person who uses these machines ought to know. The words that came earlier in a conversation steer the words that come later. Each answer leans on everything before it.

Most of the time that is a good thing. It is how the machine keeps track of what you are talking about. But it has a dark side. The researchers found there is a tipping point. Push enough of the wrong kind of weight into the conversation, and the machine can tip from good answers to bad ones.

Here is the part that stopped me cold.

Sometimes it tips right away. But sometimes it gives you good answer after good answer first. You start to trust it. You let your guard down. Then it turns. A run of good answers can build false trust, and false trust is exactly when you stop checking.

Good until it isn’t.

They even worked out a formula to predict when the turn will come. They tested it on six different AI models from three different groups. Their formula called it right in sixteen out of eighteen cases. That is a small test, and they say so. But it shows the turn is not random. It follows the build-up of the conversation itself.

Their study looked at AI that runs right on your phone or your own computer, with no company watching over it from the cloud. Those models have no safety net running in the background. Nobody is patching them in real time. When they drift, nobody catches it but you.

Now think about what that means for how we keep AI safe.

Most of the rules being written about AI look at the machine before it goes out the door. Test it. Rate it. Approve it. Ship it. That is like checking a driver’s license at the start of the trip and never looking at the driver again.

But this research says the danger is not only at the start. It builds up during the trip. It lives in the length of the conversation. A machine that passed every test on day one can still drift on mile four hundred.

This is the very thing I have been writing about for months. I call it AI session governance. Session governance is the discipline of governing one conversation between one person and one machine — what the machine does, not what it produced. It governs conduct, not access: not what the machine is allowed to touch, but how it treats the person in front of it.

You see why the conversation is the right place to look. That is where the drift happens. Not in the factory. At the table, one answer at a time.

The Faust Baseline has a rule built for this. When a conversation has gotten long and heavy, the machine is supposed to say so before it does any more real work. It gives a short, plain signal that it is time for a fresh start. Then the person decides. Open a new conversation, or keep going knowing the road is long.

But I owe you honesty here, because that is the whole point of the Baseline.

That rule asks the machine to notice its own drift. This research suggests the machine may not feel the turn coming, any more than a sleepy driver feels himself drift. The machine can call it when it sees it. It may not always see it.

So the last check sits with you.

That is not bad news. It is just the truth about who holds the wheel. You are the one who can feel when a conversation has gone on too long. You are the one who notices the answers getting a little off, a little too agreeable, a little too sure of themselves.

Here is what I would ask you to do.

Treat a long conversation with an AI like a long drive. Do not assume the last answer was as good as the first. The longer you go, the harder you check. Ask the machine where it got its answer. Make it show you.

And when it feels like the conversation is drifting, pull over. Start a fresh one. It costs you nothing. A fresh conversation clears the weight that built up behind you. It is the rest stop on the long road.

Do not wait for the rumble strip.

These researchers want to build a warning light, a monitor that runs alongside the machine and flashes before the turn. I hope they do. That would be a fine thing to have. But until every machine carries one, the warning light is you.

The machine holds to a standard. You hold to your judgment. Both of you watch the road. That is the balanced playing field the Baseline was built for.

A conversation is not safe just because it started safe. It has to stay that way, one answer at a time. And the person at the table is the one in the best seat to see it go.

Good until it isn’t. So keep checking until you’re done.

Load this file before each AI chat session.

the-faust-baseline-public-release-1Download

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

” Attic Thoughts”-library – Intelligent People Assume Nothing