AI as Exoskeleton, Not Crutch: How to Stop Offloading Your Brain

type

Post

status

Published

date

Mar 23, 2026

slug

cognitive-uploading-inverted-search

summary

Most people are using AI to get more done, not to get more capable. It’s fine to offload low‑value work, but for high‑stakes thinking we should also use AI to upload our brains: strengthening our reasoning rather than eroding it.

The real risk isn’t AI, it’s efficient forgetting

There’s a version of AI use that quietly hollows you out and accelerates forgetting.

We paste in a prompt, paste out an answer, hit submit, and two weeks later we can’t explain a word of what we ‘produced’ or why it fell flat. It perhaps looked like mastery from the outside but to the informed, in the words of T.S. Eliot, it’s really just “shape without form, shade without colour.”

That’s cognitive offloading: using AI as a shortcut around understanding and losing out in the process. You ask it to write the paper instead of helping you write the paper. Instead of investing in your future you find yourself starting to forget how to learn. You ask it to summarise the meeting just to tick a box on your task list; instead of sharpening your thinking for next time, you feel more disconnected from what was discussed and from your work. The output or time saving may be fine, but if you learn nothing and that ‘saved time’ isn’t reinvested, what have you actually gained? If that continues, you may not notice the growing incongruity across your outputs, different LLM styles and sessions pulling in different directions, or the slow organisational drift away from its core value and voice.

In professional settings, this has a natural ceiling. A client meeting, a board presentation, or a public Q&A expose the gap between what you seemed to know and what you can actually defend. Where success is measured purely in documents shipped, the gap goes unchallenged and slowly hardens into culture.

Cognitive uploading: AI as a second brain

There’s a different pattern of AI use that does the opposite job: it increases your capacity over time rather than replacing it.

Author Steven Johnson, now Editorial Director of Google’s NotebookLM, calls this “cognitive uploading.” Instead of treating AI as a ghost‑writer, you treat it as a container for your own thinking; a second brain that extends your memory and makes your past work searchable in useful ways.

The practice itself predates AI. It’s what good researchers and strategists have always done: keep a disciplined archive of notes, quotes, observations, and fragments from books, meetings, and conversations. AI changes what you can do with that archive.

You can talk to it. You can ask what you were working on six years ago that’s relevant to a decision you’re making now. You can surface connections that would never survive a keyword search because they live in the gaps between topics rather than in the topics themselves.

From the outside, the document you ship may look identical to one that came from pure offloading. The difference is in what you retain and what you’re capable of doing next time.

Inverted search: asking what’s missing

One of the more powerful ideas in this space is what Johnson calls *inverted search*.

Standard search finds what’s there. Inverted search finds what isn’t.

The basic move is simple:

Load a body of material into an AI‑augmented notebook: a literature review, a year of customer interviews, a stack of board papers, or three years of internal strategy decks.

Instead of asking “What does this say?”, ask “What’s missing?”

What questions hasn’t anyone asked yet?
Which perspectives are absent?
What assumptions do all these documents quietly share?
For example: “Given three years of our product strategy docs, what assumptions about our ideal customer do they all share, and what new insight might that reveal?”

For researchers, this is a way to find genuinely original angles instead of just recombining the same ideas with new adjectives. For strategy and advisory work, it’s a way to surface blind spots in the analysis before you commit to a direction.

The AI isn’t “being creative” here in any mystical sense; it’s mapping the shape of what’s already in the room, including the negative space.

This only works if you have a knowledge base worth interrogating. That, in turn, depends on the unglamorous groundwork of actually accumulating and curating your thinking over time.

Directed writing: staying cognitively on the hook

Writing is where AI offloading is most tempting and most dangerous.

The default pattern is: “Write a 1,500‑word article on X,” wait for a fluent draft, then tidy it up. Editing output you didn’t generate feels like work, but it’s easier than doing the thinking that would have produced the argument in the first place.

Directed writing flips that around.

The practice looks like this:

1. You decide what you’re trying to say.

💡

Outline the argument yourself: the core claim, the key sections, the evidence you need in each one.

2. You brief the AI paragraph by paragraph.

💡

For each section, you tell the model exactly what you want it to do: the point to make, the examples to draw on, the constraints and the audience.

3. You stay responsible for the logic.

💡

If you can’t explain a paragraph back in your own words, you don’t keep it. You either rewrite it or you go back and do more thinking.

In this model, the human does the cognitive heavy lifting: deciding what matters, how the argument flows, and which trade‑offs they’re willing to make. The AI handles the labour‑intensive parts of language generation.

That’s very different to throwing the problem over the wall and editing whatever comes back. The first pattern builds topic mastery. The second lets you fake it.

A simple test: exoskeleton or crutch?

Johnson’s framing of AI as an *exoskeleton versus a crutch* is a useful rule of thumb.

An exoskeleton makes you stronger. After using it, you can do more unaided than before.

A crutch compensates for a weakness you stop trying to address. After using it, you’re less confident without it.

You can apply that test to any AI‑mediated task:

Could I reconstruct the reasoning without the tool?

Have I added anything to my own understanding, or just produced a passable artefact?

Am I gradually expanding what I can do, or shrinking it?

Speed is not the enemy here. Speed that accumulates understanding is an advantage. Speed that produces nothing durable is just efficient forgetting.

If you’re deciding how your organisation should use AI

The real divide isn’t between people who use AI and people who don’t. It’s between people whose use of AI compounds their capabilities and people whose use quietly erodes them.

If you’re responsible for how your team, faculty, or organisation uses these tools, the key questions are:

💡

Where are we unintentionally rewarding cognitive offloading and at what cost?

💡

Where could we start using AI as an exoskeleton: extending memory, surfacing blind spots, stress‑testing our thinking?

Those are the kinds of systems‑level questions I help leaders work through in knowledge‑heavy, multi‑team organisations that care more about shared expertise than raw output.

If you want help designing AI practices and policies that build capability instead of hollowing it out, get in touch via stephendmann.com

Acknowledgement: This blog piece was in large part inspired by watching this Youtube video:

It's cognitive uploading" | How Google NotebookLM's Steven Johnson uses AI as a second brain

Leadership Intelligence