What the System Does to the Mind

Self & System · The Thesis

What the System Does to the Mind

A framework for tracking what running your life through an AI does to your cognition, function by function.

Anyone who runs their working life through an AI system is running an unmonitored experiment on their own cognition. And the point is not confined to systems like mine: any use of AI at all changes how you do things, and with it how much each function of your thought gets exercised. Whatever the system does for you, you stop practising. Whatever its workflow demands of you, you practise more than ever.

This framework is a way of tracking that trade: which capabilities a system is doing for you, which it is training, and which are quietly going unexercised. It runs in six steps, one function at a time. The examples are from my own system, because that is the one I can see inside; the method is meant to be general.

FunctionsWhat a mind does: working memory, retrieval, analogical reasoning. Always in blue.

ActivitiesWhat a person does: attending a lecture, capturing an idea, writing. An activity is an aggregate of functions, and it is activities a system changes directly. Always in amber.

STEP 1The inventoryThe 24 functions of a mind, in 8 families. The list the audit runs against. STEP 2The ledgerWhat am I doing less of, what more, and what is new? Activities, observed. STEP 3The mappingWhich functions does each activity touch, and which way does it push them? STEP 4The aggregateEvery function in one table: my direction of use, and the system's verdict. STEP 5The triageThe exposure test: which losses matter. Pure framework. STEP 6The responseMy watch list, and how I practise it.

Step 1 · The inventory

The twenty-four functions

I made this list by reading a lot of other people's frameworks and pulling them together. It is not meant to be new; it is the level of decomposition I find useful for breaking cognition into fundamentals. One rule throughout: a function is something a person does, never a tool.

Hover or tap any function for its definition. This board recurs through the whole audit: Step 3 lights it up per activity, Step 4 stamps it with verdicts, Step 6 rings the watch list.

A footnote on mode: any of the twenty-four can run fast and automatic or slow and deliberate (System 1 and System 2). Mode is a way of running a function, not a function, so it gets this footnote and nothing more.

The second axis: channels

Information also has channels, in and out: four modalities, each with a direction. They are not functions, so they get their own strip, and the audit tracks them the same way.

The channel grid: four modalities, two directions each.

Step 2 · The ledger

What am I doing less of, what more, and what is new?

A step up in altitude: from functions to activities. An activity is a collection of functions exercised together, and it is activities the system changes directly. Observable behaviour only; no judgements yet.

Non-exhaustive: just examples from my life

↓Less

Taking notes by hand in lectures. The session is recorded, transcribed, and the notes are generated.
Reading every article end to end. The pipeline reads and summarises the long tail; I read the pieces that matter.
Writing first drafts of routine prose.
Remembering people's details unaided.
Holding the week's plan in my head.
Listening as the primary way of taking anything in. Nearly everything is transcribed.
Holding many things in mind at once, in one shape: I voice-dump ideas out the moment they occur instead of carrying them.

↑More

Speaking thoughts aloud at the moment they occur, all day, to a machine that needs the narration.
Explaining: saying what I am doing and why, out loud, as I work.
Building frameworks and taxonomies, constantly, with AI in the loop. This page is an instance.
Connecting ideas across domains. The linking layer keeps putting distant material side by side.
Holding many things in mind at once, in another shape: running several workstreams in parallel and switching judgement between unrelated things, tab to tab.
Editing and judging drafts, rather than producing first drafts.
Reviewing machine output: verifying, gating, approving, rejecting.

Step 3 · The mapping

Which functions does each activity touch?

Three worked examples. For each activity: which functions it touches, who runs each one now, and how much I exercise it compared to before the system. Pick an activity; the board lights up with what it recruits.

Before: sit, listen hard, take notes by hand, lose whatever I failed to catch. Now: the session is recorded and transcribed, the notes are generated, and I attend the argument. I concentrate less hard, because I know I have it.

Function touched	Who runs it now	My exercise of it, vs before
Spoken word · in (channel)	Machine record and transcribe	↓ less then re-routed into text
Encoding	Shared the machine makes the notes	↓ less note-making forced encoding; generated notes do not
Working memory	Shared	↓ less I hold less because I know it is held
Sustained attention	Me	↓ less the cost of drifting has fallen
Retrieval · search	Machine notes, retrieval on demand	↓ less I used to try to pull it back from my own memory; it was unreliable
Fluid reasoning (following the argument)	Me	→ same

Before the system, an idea stayed in my head until it faded there. Now I speak it the moment it occurs: capturing by voice is the AI-era form of the same activity. It is transcribed, filed, and connected.

Function touched	Who runs it now	My exercise of it, vs before
Spoken word · out (channel)	Me	↑ more the idea was never spoken at all before
Divergent generation	Me	↑ more the follow-on thought used to die with the first
Working memory	Machine after the first sentence	↓ less I dump early instead of holding the thought
Encoding	Machine transcribed and filed	↓ less
Retrieval · search	Machine	↓ less I used to try to recall the idea from my own head; it was unreliable
Elaboration	Split the machine downstream, and me later	↓ less ↑ more less per idea in the moment; more overall, because an idea that would have been lost now survives for me to elaborate later

Augmentation and training at once. A piece starts as a spoken or typed brain-dump, comes back as structure, and goes through rounds of my judgement against the tool's drafts.

Function touched	Who runs it now	My exercise of it, vs before
Analogical reasoning / structure mapping	Shared built with the tool	↑ more every piece is a framework argued into shape; I build and hold far more structures than I did
Convergent selection	Me judging the drafts	↑ more judging draft after draft is selection at volume
Elaboration	Split between me and the machine	↓ less the tool develops part of every seed; I develop the rest

Legend: ↑ the function runs harder or more often than before the system · → unchanged · ↓ less · re-routed = the information changed channel on the way in. "Better output with the same or less work by the function" counts as unchanged, not more: the measure is exercise, not results.

Step 4 · The aggregate

All twenty-four functions, one board

Done properly, this step would map every activity a brain performs, weight each by time spent before and after the system, and weight each function within each activity. I have not done that here. The board below is a judgement call across the whole system, not just the worked examples above: highly indicative, no more. Where the worked evidence is missing, the call is drawn from how the system runs day to day, and marked as indicative.

Direction · my unassisted use, read off my days

↑ more the function runs harder than before
↓ less it runs less than before
→ same unchanged

Verdict · what the system does about it, read off the machine

extension amplifies a function that still runs in me
replacement performs it in my place
gap fill performs what never reliably ran in me
conversion moves information from a weak channel into a strong one
none does nothing about it

Each function coloured by the system's verdict, carrying its direction arrow where the row is filled. Click any function to read the evidence behind its row.

All 32 cells carry a verdict: 14 read off worked evidence, 18 indicative · extension 7 · replacement 10 · gap fill 1 · conversion 1 · none 13. Indicative calls firm up, or get overturned, as evidence accumulates.

One pattern worth naming: where a row reads none yet more, the workflow itself is the training. The system does nothing about the function and demands it daily.

Step 5 · The triage

Which of the losses matter?

Not every fall is a problem. Retiring a function on purpose is legitimate; losing one by accident while it is still owed in rooms where nothing can assist is not. Each fallen function faces a two-part test.

is the function still needed unassisted, in real time?

noyes

strong before the system

Lend it out

Something to lose, but nowhere it is owed. Delegate, keep the judgement calls, accept the fall in exercise as affordable.

Protect

The exposure quadrant. A function relieved of its daily repetitions by the system, still load-bearing in rooms the system cannot enter. Losses here compound quietly.

weak before

Delegate freely

Nothing to lose and nowhere it is owed. The pure wins of the audit; take them without guilt.

The trap

Weak is not absent. Routing every use around a weak but working function entrenches the weakness, and the function is still owed where nothing can assist.

The exposure test. Care about a function when there is something to lose and somewhere it is owed. The bottom-right quadrant is the subtle one: conversion looks free precisely because the capacity was weak, and it is only free if the weak channel is truly unusable or never needed again.

Step 6 · The response

Deciding what to do about it

The aim is not to use the system less. The aim is to make the retention of exposed functions a decision rather than an accident. The watch list is personal by construction; mine, worked as an example, is working memory and spoken word · in: one function, one channel.

The watch list, ringed: one function and one channel, out of everything the audit turned over.

Working memory

Before the system: It carried the week's plan, held an idea until I could use it, and kept a lecture in mind while it ran.
The assessment: I dump early instead of holding the thought, and I hold less because I know it is held. The aggregate reads it both ways: less in the old shape, and more in a new one, because I jump tab to tab and deliberately run many things at once, and that multitasking is itself the practice.
Where it is still owed: Conversation: holding several threads while someone watches, with no tabs and no capture. The exposure quadrant exactly.
The practice: I write some essays entirely by hand. Repetitions, not perfection.

Spoken word · in (listening)

Before the system: Listening was the primary way anything got in: lectures, meetings, conversation heard once and held.
The fall: I route around the channel. Nearly everything is recorded and transcribed, and the sound re-enters as text: conversion, not exercise.
Where it is still owed: Every live room. A live discussion cannot be joined off a transcript.
The practice: Conversation itself: the speaking repetitions are listening repetitions too.