Cognitive Compression

Overheard the phrase in a San Francisco coffee shop. Two people at the next table — one research, one product — kept circling an idea without landing on it. Then one of them just said it: cognitive compression.

The term already exists in cognitive science. Lai and Gershman formalized the mechanism in 2024: under time pressure, your brain simplifies its decision rules. You stop weighing tradeoffs and start pattern-matching to the nearest familiar action. Judgment compresses. Decisions get brittle.

What caught me is that AI is reproducing this — at organizational scale — and almost nobody is naming it.

Here's the standard story: AI makes individuals faster. Tasks get cheaper. Throughput goes up. All true — at the individual level. But organizations aren't individuals scaled up. They're coordination systems. And coordination doesn't fail the same way speed does.

Give every person a faster tool without redesigning how judgment flows between them and you don't get a faster org. You get compression at the seams — where one person's output becomes another person's input. Reviews that were substantive become rubber stamps. Decisions that were deliberate become reflexive. The system speeds up where it didn't need to and compresses judgment where it needed more of it.

This is where harness debt stops being a technical concept and becomes a design failure you can feel. A harness is the set of constraints and context boundaries around an AI agent. Good harness absorbs complexity — carries the organizational context so the operator doesn't hold it all in working memory. Bad harness, or no harness, pushes that load onto the human. The operator becomes the guardrail. Under load, the operator compresses.

This isn't a training problem. You can't train your way out of a structural deficit. If your system needs every operator to hold the full context of every decision an agent might make — you have a harness problem. Cognitive compression is the symptom.

Directional Delegation — delegating outcomes-under-constraints, not tasks — is what prevents this. It works because it separates what should compress from what shouldn't. Execution compresses. Judgment doesn't. The delegator holds direction and constraints. The agent — human or AI — handles the path. What matters stays with the person who has context to judge it. How to get there compresses freely.

That's the distinction most of the AI-at-work discourse keeps missing. Not whether to use AI. Which cognitive operations you're compressing — and whether you chose them, or the tool did.

Cognitive compression is going to become a common term. Right now it sits at the intersection of cognitive science, org design, and AI deployment — exactly the seam where the interesting failures live. The question is whether it enters the conversation as a productivity-blog buzzword or as a diagnostic with teeth. I'd prefer teeth.