Verification Isn't About Information

The first time you verify an AI's answer and watch it change the next day, you think it's a bug. The second time, you start to wonder what verification actually is.

What we call verification is mostly checking new information against what's already in there. AI made that visible because what's loaded in is now visible too — you can read the context window.

AI doesn't verify. It builds answers out of whatever's in its context: your earlier messages, search results, notes you fed it. The longer the context, the more the model leans on what it already said. Researchers have a name for when this gets bad: context rot.

The model itself remembers nothing between calls. The chat window keeps the transcript and re-attaches it every time. So the answer comes from a record that keeps getting longer, with the most recent stuff tugging hardest.

You'd expect more information to sharpen the answer. It blurs it.

This sounds like an AI problem. It's also yours.

Your judgments come from what you've seen. When someone reacts to the same event the opposite way you would, it seems strange — but from inside their experience, theirs is the right answer.

Only the material differs. For AI, it's text in a context window. For you, it's wiring laid down over years. The shape of the trap is the same: you can only answer from inside what's already there.

That isn't only a weakness. It's where depth comes from. A veteran's intuition is sharp because they've stayed in one place long enough for the wiring to set. Solid judgment, fast decisions — all of it comes from staying.

But the same wiring makes you less flexible. There's a name for it: cognitive entrenchment. As expertise grows, the categories in your head harden, and new information has a harder time getting in. The more you know, the more you filter inside yourself.

You think you're verifying. You're checking consistency against what's already in you.

Statistics, peer review, experiments — they help. But the interpretation still happens inside your experience. The tools don't pull you out of yourself.

Some checks bypass the filter entirely. Compile the code; it runs or it doesn't. Check the date; it matches or it doesn't. But that's not what's at stake. The verification that matters — of claims, plans, judgment calls — runs through your filter every time.

AI works the same way. The richer the context, the deeper the answer looks, and the same richness blocks what's new. Depth and verification pull at opposite ends of the same rope.

When you need depth, stay. When you need to verify, step out.

To shake AI loose, open a new window. Ask a different model. Clear the context.

To shake yourself loose, talk to someone outside your field. A different generation, a different industry. Anyone whose wiring isn't yours.

You've been on something for two months. A new hire asks why. You don't have a good answer. That's what shaking yourself loose looks like.

One difference matters. For AI, a new window is a click. For a person, it takes deliberate effort, and often it doesn't work at all. People have the harder job.

So verification has three parts, not one. What gets loaded. What you're looking through. What you can clear out.

What gets loaded is the context — what you feed the model, or what surfaces in your head at the moment. What you're looking through is the filter — the structure of what you already know. What you can clear out is the familiar — deliberately cut to make room for something new.

Verification isn't checking whether information is right. It's noticing how those three parts move differently every time.

Thinking that can hold this awareness is thinking that hasn't set.