The Knowledge AI Can't Generate

By April 2025, an analysis of nearly a million newly published web pages found AI-generated content in 74 percent of them. Fifty-seven percent of all online text, by one estimate, has now been generated or translated using AI tools. Automated traffic exceeded human traffic on the web for the first time last year — 51 percent bots, 49 percent people.

These numbers keep surprising me, not because they're alarming, but because of what they imply about value. If AI can generate most of the world's content, the stuff that retains value is exactly the stuff it can't. And almost nobody building knowledge tools is thinking about this clearly.

What AI can replicate and what it can't

Ask an AI to write a summary of best practices for one-on-ones with direct reports. It will produce something thoughtful and reasonable. Ask it to explain how transformer attention mechanisms work. Same. Ask it for five frameworks for thinking about product strategy, or a primer on the psychology of habit formation, or an explanation of how to structure a difficult email. All of it: passable to excellent, available in seconds, essentially free.

That's most of what people currently capture in their knowledge bases. Summaries of books they read. Notes from articles. Frameworks someone shared in a newsletter. Techniques for things they want to get better at.

All of it can now be regenerated on demand. The moment you can prompt your way to a usable version of that knowledge faster than you could find your own notes on it, your notes stop being an asset.

But here's what doesn't regenerate: what you decided, and why. Who you know, and what you know about them. What you tried that didn't work. What you told someone you'd do. The pattern you noticed in your own projects that no one else would have data on. The preference you've developed through experience that you've never written down.

That knowledge is yours in a way that a summary of a book never was. It can't be synthesized from public data because it was never in public data.

The inversion nobody talks about

For most of knowledge management's history, the bottleneck was access to information. The Dewey Decimal System, the index card, the wiki, the search engine — all solutions to the same problem: how do you find the thing you know exists somewhere?

That bottleneck is gone, or very nearly. Information retrieval is becoming essentially solved. The bottleneck has moved.

The new bottleneck is provenance. Not "where do I find an explanation of X?" but "what do I know about X, specifically — from my own experience, my own context, my own decisions?" That question can only be answered from a private record, and private records are what almost no AI tool can touch.

This is the inversion: public knowledge is abundant, free, and quickly regenerable. Private knowledge is structurally scarce. Not because it's locked away — just because it hasn't happened to anyone but you.

The record that can't be faked

There's a man I know who worked in private banking for a decade before AI tools became common. He kept brief notes after every client meeting. Not elaborate — a few lines. Who was worried about what. What they'd mentioned about their kids. What risk tolerance they'd shown in past downturns. Nothing a compliance officer would flag, just context.

Those notes made him genuinely useful in a way that generic advice couldn't replicate, because they were anchored in what actually happened. Not what might have happened. Not what usually happens. What this person, in this situation, had actually told him.

That's the value structure that personal knowledge has always had, but it becomes more visible in contrast with a world flooded with generated content. The AI can write ten plausible versions of the email you might send. It can't tell you that the recipient mentioned last time you talked that they were feeling stretched thin and not to add to their plate. It can't tell you that you made a similar decision in 2023 and later regretted it for reasons X and Y.

Not because it lacks the reasoning ability. Because it lacks the record.

What this means for how we build

Most personal knowledge tools are built around capture. The bet is that if you make it easy enough to write things down, valuable knowledge will accumulate. And then search — the assumption being that the bottleneck is retrieval speed.

I think this is wrong, or at least incomplete. The bottleneck now is structure. Not for your benefit as a reader, but because AI can only work with your private knowledge if it's in a form that supports querying. Prose notes from three years ago, stored in a folder somewhere, aren't usable in any meaningful sense. A person record with a relationship, linked notes, a preference entry that says "prefers direct feedback, not in front of the team" — that's something an AI can reason from.

The format determines whether your private knowledge remains locked or becomes legible. And legibility is the prerequisite for everything useful that follows.

A bet worth making

I realize this can sound like an argument for Harbor, and I suppose it is, indirectly. But the underlying claim doesn't depend on any specific tool. It's about where value is moving.

Generated content is getting cheaper faster than any of us expected. A year from now it will be cheaper still. The information that retains value is the information that can only come from having lived through something — specific, contextual, private. Not because it's hidden, but because it belongs to you and to no one else's dataset.

Building a knowledge base right now feels a bit like buying oceanfront property as the tide is changing direction. Most of what you used to store there is washing away. But the part that's genuinely yours — the decisions, the relationships, the hard-won context — that's not going anywhere.

The question is whether you have somewhere to put it that can actually be used.

---

Asgeir Albretsen is the founder of Harbor.