AI can find your knowledge. It can’t fix it.

I started in psychology, then spent years in learning design and enablement. The longer I do this work, the more I think knowledge management is mostly a memory problem, and memory is a people thing.

Picture a new hire asking the company’s AI assistant how a process works. It answers instantly, totally sure of itself, with a doc from two years ago. The process changed last spring and nobody updated the page. She follows the steps, and a week later someone tells her they were wrong.

The AI didn’t fail. The knowledge was stale.

We remember where, not what

Psychologists Betsy Sparrow, Jenny Liu, and Daniel Wegner found that when people expect to have future access to information, they remember less of it and remember where to find it instead. We treat the internet like a coworker who holds the facts for us.

An AI assistant takes that one step further. People stop carrying the detail and trust the search bar. That’s efficient until the source is wrong, because then nobody catches it. Nobody’s holding the detail anymore.

On a distributed team this gets worse. There’s no hallway to overhear a correction, which I wrote about in the hallway piece. So the written version is all people have.

Sharing is a trust decision

I used to think knowledge sharing was a tools problem. It’s mostly a trust problem. Researchers studying knowledge hiding, where someone holds back what a coworker asked for, found that distrust predicted every form of it. Companies had invested in sharing software, and people still held back.

That tracks with everything I’ve seen. A page is only as good as the willingness to put the real answer on it, and that depends on whether people feel safe doing it. I wrote about the asking side in the safe-to-ask piece.

How I’d run it

I’ve built a knowledge hub for support engineers. The platform was the easy part, and keeping it true was the work. This is what I’d do again.

• Every important page gets an owner and a review date. If nobody owns it, nobody trusts it.
• One source of truth per topic, and I retire the rest. Old versions are what AI quietly serves back.
• Capture knowledge inside the work, like at the end of a launch or when a question comes up for the third time. If documenting is its own project, it won’t happen.
• Make it easy to say “this answer is wrong.” A flag button turns every confused reader into a free quality check.
• Start with one team and one body of content. Fixing a hundred pages is a lot easier than fixing ten thousand.

Where people stay in the loop

Automating this makes sense for a lot of it, but you have to decide where a person belongs. The AI world has three standard terms for how close a person stays to the decision. The difference is whether a person approves before the AI acts (in), keeps watch while it acts (on), or only checks afterward (out).

Set it by action, based on how hard the action is to undo.

The point

Knowledge has a life cycle: you create it, find it, trust it, and retire it. AI speeds up finding. The rest still takes people deciding who owns what and feeling safe enough to say when something’s wrong.

AI changes where people look. It doesn’t change whether what they find is right.

Happy Learning, KP