What Happens to Your Own Thinking

Before any argument about these tools, it is worth being clear that delegating thinking to external aids is ordinary, ancient, and mostly good.

Writing offloads memory. Calculators offload arithmetic. Navigation offloads spatial reasoning. Search offloads recall. Each transfer freed capacity for something else, and no serious person proposes reversing them.

What the research on cognitive offloading, studied by Evan Risko, Sam Gilbert and others, establishes is not that offloading is bad. It is that it has a predictable structure: when we offload a task, our internal capacity for that task develops less or degrades, and we become progressively more reliant on the external aid.

The well-known illustration is the Google effect, described by Betsy Sparrow, Jenny Liu and Daniel Wegner in Science in 2011: when people expect information to remain available, they remember the information less well and remember where to find it better. That is an efficient reallocation rather than a deficit, provided the external store remains available and accurate.

So the question about any offloading is not whether capacity moves outward. It is whether the thing being offloaded was load-bearing for something else you need.

Arithmetic is a good case for offloading, because the number-sense that matters for judgement is largely separable from the mechanics of long division, and calculators are correct.

The argument of this lesson is that composition, analysis and reasoning are a worse case, for a specific reason: they are not just outputs, they are the process by which understanding is formed and checked. Which means offloading them removes the thing that would have told you the output was wrong.

The central claim, and it is worth stating carefully because it is easily caricatured as nostalgia for difficulty.

When you work through a problem, the difficulty is not purely friction to be removed. Some of it is diagnostic. The point where you cannot make the argument work is usually telling you that you do not understand something, and the discomfort is the signal.

This appears across several literatures. Research on desirable difficulties, associated with Robert Bjork and colleagues, finds that conditions making learning feel harder and slower, spacing, interleaving, retrieval practice, produce better long-term retention and transfer than conditions that feel fluent and easy. Crucially, learners systematically prefer and predict better results from the easier conditions, and are wrong.

That last finding is the important one here. The subjective sense of learning well is a poor guide to actually learning, and it points in the wrong direction.

Apply that to generated output. Reading a clear explanation feels efficient. Producing the explanation yourself feels slow and effortful. The research says the second builds the durable understanding and the first builds the confident feeling of it, which is the illusion of explanatory depth from lesson one arriving by another route.

The same argument applies to composition, which the writers cursus in this catalogue makes at length. You find out that an argument does not hold when you try to write the sentence joining its halves. Delegating the sentence removes the discovery.

What this does not argue. It does not argue for doing everything the hard way. Most tasks are not building anything you need, and the difficulty in them is genuinely waste.

It argues for being able to tell the difference, which is the next step.

A test for any task you are considering delegating, which resolves most cases quickly.

Start with the task and ask whether doing it builds a capability you need. Not whether it is enjoyable or whether it feels productive, but whether the doing produces something in you.

If no, delegate freely. Formatting, reformatting, boilerplate, transcription, converting between structures, looking up syntax you will never memorise. Nothing is lost, and this is most work.

If yes, ask a second question: will you be accountable for judging the output?

If you will not be, delegating is still mostly fine. You are consuming a result within someone else's responsibility.

If you will be, and this is the crucial branch, delegating creates a specific problem. Judging output requires the capability that doing the work would have built. You are removing the practice that produces the skill you need to supervise the removal.

That is the trap, and it operates on a delay. It is invisible in the short term, because you currently have the skill from work done before the tool existed. It appears years later, in people who never built it, and in you as it decays.

The practical resolution for that branch. Keep doing enough of it to stay calibrated. Not all of it, and not for efficiency, but as maintenance of the instrument you supervise with.

Who benefits from these tools is not uniform, and the pattern has an uncomfortable implication.

The consistent finding across studies of AI assistance in various domains is that assistance tends to help less experienced people more, compressing the gap between novice and expert performance on the measured task. That looks straightforwardly good, and it is, for the task.

The complication is what happens to the mechanism by which novices become experts.

An expert can evaluate generated output because they have the domain model to check it against. They notice the wrong assumption, the missing consideration, the answer that is technically right and inappropriate here. For them, the tool is genuine leverage: it removes production work and leaves the judgement they already have.

A novice cannot evaluate it, because the evaluation requires exactly what they do not yet have. They receive fluent, confident output in a domain where they have no independent basis for assessment, which the first lesson established is the worst possible position.

And historically the way they would have acquired that basis is by doing the work the tool now does, badly at first, being corrected, and gradually building the model.

So the paradox. The tool most helps the people least able to check it, and it does so by removing the work through which they would have become able to check it.

This is the same open question the architecture and engineering cursus raised about the training pipeline, and it does not have a settled answer. Some argue that reviewing generated work builds comparable judgement; whether reviewing teaches what producing taught is genuinely not known.

What can be said practically to anyone early in a field. The short-term output gain is real and the long-term cost is real, and you cannot detect the second from inside because your output looks fine. That is an argument for deliberately doing some work unassisted while it is cheap to be slow.

A finding worth confronting directly, because it undermines the evidence most people are relying on.

A controlled study published by METR in July 2025 examined experienced open-source developers working on their own repositories, with and without AI assistance. Before the work, they expected to be about twenty-four percent faster with the tools. Afterwards, they believed they had been about twenty percent faster. Measured against the control condition, they were about nineteen percent slower.

The usual caveats apply and should be stated. One study, one domain, experienced people on codebases they knew deeply, which is a setting where assistance has less to offer than it would for unfamiliar work. It should not be generalised into a claim that these tools never help.

What it does establish robustly is the calibration failure. Self-assessment of the speedup was wrong by roughly forty percentage points, in the favourable direction, by people well placed to judge, on their own work, having just done it.

Why that matters for this cursus specifically. Almost everyone's belief that these tools help them is based on exactly the kind of self-assessment this study found unreliable. It feels faster. The output arrives sooner. Those sensations are real and they are not measurements.

And there is a mechanism that explains the direction. Time spent waiting, reading, evaluating and correcting is fragmented and less salient than time spent stuck. Being unstuck feels like progress even when the total is longer.

The practical response is not to stop using them. It is to prefer changes whose benefit is structural and checkable over changes whose benefit is felt: the report that now goes out on time, the backlog that cleared, the thing that previously did not happen. Those are facts about the world. Faster is a feeling, and it is the specific feeling that has been measured and found wrong.

An effect that operates at the level of groups rather than individuals, and that no individual can observe.

A model produces, in effect, the central tendency of what it has read. Two people asking the same question receive substantially similar answers. Across an organisation, a profession or a field, that pulls output toward a common centre.

Why that is a loss rather than a gain in consistency. Progress in most fields depends on variation: someone framing a problem unusually, applying an idea from an unrelated domain, or being wrong in a productive direction. Those are deviations from the centre, and a tool that returns the centre reduces their supply.

The concrete versions.

Strategy documents across competitors converging on the same framing, because they were produced the same way.

Research converging on the questions that are well-represented in the literature, since those are what a model surfaces as important, which entrenches whatever the field already studies.

Hiring, where applications converge and the differentiating signal disappears, so selection moves to whatever remains, which may be worse.

And the compounding version: generated text becomes training material for later systems, narrowing the distribution further.

What an individual can do about it is limited, and it is not nothing.

Generate your own position before consulting anything, so your framing is yours and the tool is checking it rather than supplying it.

Use the tool for what it is good at, which is the typical case, and reserve the unusual angle for yourself.

And treat convergence as a signal. If your output resembles what everyone else produced, that is information about how it was made, and in any competitive context the value was in the deviation.

The general point for this cursus. Some costs are invisible from the inside because they are statistical and distributed. Your document is fine. The distribution of all documents got narrower.

What a reasonable person should actually do, given everything in these three lessons.

Form your view first, then consult. This single habit addresses more of the problems in this cursus than anything else. It preserves your framing against homogenisation, prevents sycophancy from having a position to agree with, keeps the composition work that builds understanding, and turns the tool into a check rather than a source.

Delegate what does not build anything you need. Formatting, transcription, boilerplate, conversion, lookup. This is most work and there is no argument for doing it by hand.

Keep enough of what you supervise. If you are accountable for judging a kind of work, do some of it unassisted, periodically, as maintenance rather than efficiency. This will feel wasteful and is the point.

Be more careful if you are early in a field. The tool helps you most and you can check it least, and the work it removes is how you would have become able to check it.

Use the explain-it-yourself test before relying on your own understanding. If you cannot produce the explanation without looking, you have familiarity rather than knowledge, and you cannot evaluate output with it.

Judge adoption by structural changes rather than by how it feels. The thing that now happens on time, not the sense of moving faster, which has been measured and found badly miscalibrated.

And hold the distinction that runs through all three lessons. These tools are extremely good at producing the typical, well-formed, plausible version of anything. That is genuinely valuable for the large share of work where typical is what you need, and it is precisely wrong for the work where the value was in being right, being specific, or being different.

Knowing which kind of work you are doing is the whole skill, and nothing about it is new. What changed is that the cost of not knowing went up.

Three lessons in one place.

The cues you use to assess writing, fluency, confidence, structure, specificity, worked because they were costly to produce. They are now free and uniform, so they carry no information while still feeling like they do. Processing fluency gives smooth text an unearned truth bonus. Automation bias suppresses checking and strengthens as tools get more reliable. The illusion of explanatory depth means a clear explanation feels like understanding. Sycophancy means asking for a second opinion often returns your own. And presented reasoning raises persuasiveness more than reliability.

All of those operate below deliberate control, and awareness does not remove them, which is why resolving to be careful is not a strategy.

So checking has to be structural and decided in advance. Triage by consequence rather than by suspicion, because suspicion no longer tracks anything. Always verify numbers, citations, names, and anything reaching a decision or a third party. Know that asking the model again is worthless and asking another model is nearly so; only the primary source discriminates. For citations, read them, because saying-something-different is more common than not-existing. Frame requests as questions rather than arguments. And mark what you checked, in the artefact, because verification does not survive editing.

And the effects on you. Offloading is normal and it moves capacity outward, which is fine unless the offloaded thing was how you built the judgement you now need to supervise it. Keep doing enough of what you are accountable for. Be most careful early in a career, where the help is greatest and the ability to check is least. And do not trust your sense that it is working, which has been measured and found wrong by forty points.

The single idea underneath all of it. These systems produce the typical version of anything, extremely well. Most work needs typical. The work that matters usually does not.