Knowledge collapse, explained: when AI makes you learn less

Ask a leader today: “Are your teams faster with AI?” The answer is yes, without hesitation. Ask the next one: “And your organization, is it learning faster?” The silence that follows is the subject of this article.

The individual gain is spectacular and immediate. A salesperson drafts a proposal in ten minutes, an analyst clears a file in an hour, a consultant scopes an engagement in the time it takes to drink a coffee. But as work shifts into a private conversation with an AI, something vanishes in the background: the trace. The trade-offs, the methods, the in-house reflexes (everything that used to be written down, re-read, passed on) evaporate into threads no one else will ever see.

What the research says

In February 2026, three MIT researchers (Daron Acemoglu, Dingwen Kong and Asuman Ozdaglar) published a paper with a blunt title: “AI, Human Cognition and Knowledge Collapse.” Their model starts from a simple, almost obvious observation once stated: a good decision rests on two complementary kinds of knowledge.

• General knowledge, shared, accumulated by the community: the principles, the patterns, the good practicesPracticeA unit of know-how captured in Marylink: not a document but an executable structure (content, prompt, rules, style). no one reinvents alone.
• Context-specific knowledge, individual, situated: what you know about this client, this file, this situation.

The two are complements, not substitutes: one without the other isn't enough. And here is the central mechanism. Learning costs effort, and that effort produces two things at once: a private signal about your own context, and a “thin” signal that adds to the community's stock of general knowledge. It's this second, almost-free byproduct that grew the collective asset every time someone searched, doubted, explained.

Agentic AI gives you the context-specific knowledge without the effort. And without the effort, common knowledge stops replenishing itself.

The model's conclusion is sharp: when human effort becomes sufficiently “elastic” and the accuracy of AI recommendations crosses a threshold, the economy can tip into a knowledge-collapse regime: a steady state where general knowledge ultimately vanishes, even though every personalized recommendation stays excellent. That's the trap: everything looks fine at the individual level, and the collective stock collapses anyway.

Why this concerns you, even if you're not an “economy”

The paper reasons at the scale of an entire society. But the same mechanism replays, identically, inside an organization, faster, because a company is a small, very dense community.

In your world, “general knowledge” has a name: it's your know-how. The way you qualify a lead, structure a proposal, run an engagement, handle a dispute. Before AI, that knowledge settled (imperfectly, slowly) into documents, templates, team conversations, the memory of the veterans. Every colleague who solved a problem left behind a trace the next one could use.

With a personal AI assistant, that deposit stops. The problem is solved, brilliantly, in a private thread. Then the thread closes. Tomorrow, the colleague next door will retrace the same path from scratch. The knowledge didn't circulate; it never even existed outside one person's head (and ChatGPT account). And the day that person leaves, it all leaves with them.

The remedy is not to slow AI down

That's the reflexive mistake, and the paper rules it out. Reducing AI accuracy is not the answer: collective welfare is not a monotone function of accuracy, there's an optimum, and crippling the tool destroys value without fixing the root cause.

The authors' recommendation, by contrast, is perfectly clear, and it's the sentence that should keep every leader awake:

A greater aggregation capacity for general knowledge (more effective sharing and pooling of what humans produce) raises welfare unambiguously, and increases resilience to knowledge collapse.

In other words: the problem isn't that AI is too good. The problem is that there's no mechanism to capture and make common what it helps each person produce. At the scale of an organization, that mechanism is exactly what needs to be built.

Marylink, or aggregation capacity at company scale

This is Marylink's reason for being. Not one more assistant, not one more wiki: the layer that aggregates. When a colleague finds, with their AI, an approach that works, Marylink captures it, structures it, has it endorsed by a referent, and makes it reusableReuseThe same practice serving many times, across many spaces, the key measure of the Practice Graph's value. and executable by the whole team, from their own tools, via the Practice Graph and the MCPMCPModel Context Protocol: the open standard that connects an AI assistant (Claude, etc.) directly to your Marylink space. protocol.

• The individual keeps their speed gain: nothing is taken away.
• The organization recovers the “thin” signal that was being lost: every use enriches the common asset instead of bypassing it.
• Knowledge becomes a governed, measured, transferable asset, not a dependency on a few people's memory.

The research names aggregation capacity as the lever that “raises welfare unambiguously.” Marylink is the operational implementation of that lever. The theoretical debate about knowledge collapse has a product answer, and it exists today.

To understand the underlying mechanics (how know-how becomes a typed, governed, agent-executable “practice”), see the science behind Marylink and the piece on “the missing layer.”