What Can the Work Tell You About Yourself?

The work was good

Let me build a scene rather than report one. Give someone a piece of work due on Thursday — an analysis with a recommendation at the end of it, the kind that gets read by people who will act on it. Give her an assistant, and three days.

She makes real decisions along the way: she throws out the first framing because it flatters the wrong people, she notices a figure that cannot possibly be right and chases it down, she overrules a suggestion she can't defend. On Thursday she sends it. It is good work. Everyone says so, and they are right.

The following week someone asks a question one layer beneath the surface of it. Not a hard question — a reasonable one, about why the second approach was ruled out. And she cannot answer it with much confidence. Not because she has forgotten: because she cannot say which parts of the argument she understood and which parts she accepted.

Nothing in three days ever asked her that. The process measured the report. It never measured her.

What the help actually did

The familiar worry runs like this: the help makes the task easy, and the easy version skips the part where you would have learned something. It is a plausible story. A strong causal study complicates it.

In 2023, researchers ran a randomised trial in a Turkish high school — nearly a thousand students, four ninety-minute sessions. Each session had the same shape: a teacher reviewed a topic, the students practised, then sat a closed-book exam on their own. Only the practice differed. One group had textbooks and notes. Two had a laptop running the same application, the same model, the same chat interface. The only thing separating those two was the system prompt, which no student ever saw.1

One prompt told the model to tutor the student and to follow the student's instructions carefully. The guardrail prompt instead supplied the teacher's own solutions and common mistakes, prohibited full solutions, required students to show their work before receiving help, introduced progressive hints and asked for an explanation when a student volunteered an answer.

The results run against the intuition. During practice, the plain prompt raised scores by 48 per cent over the control group; the guardrail prompt raised them by 127 per cent. Then the laptops went away. The plain-prompt students scored 17 per cent below students who had practised with textbooks and notes. The study detected no exam difference between the guardrail group and the control group: neither a positive gain nor a detectable loss.

The larger gain in practice did not come with a worse unassisted result, and its size does not reveal the mechanism. The design of the assistance mattered, but the study cannot isolate which element did the work: the two prompts differed on at least nine things at once.

That is design evidence, about children, in mathematics, measured immediately — not the subject of this essay, but enough to clear away the easy explanation before the harder one arrives.

Three quantities

The harder one comes from a different literature, and it is about adults.

In two studies reported last year, researchers gave people twenty logical-reasoning problems from the American law school admission test. In the second, the cleanly randomised one, 245 people worked with ChatGPT and 207 without, and everyone was told they would be paid a bonus for estimating their own score accurately.2

The assisted group did better: about thirteen and a half out of twenty, against about nine and a half. They also overestimated. They thought they had scored seventeen. Roughly four points high.

The unassisted group overestimated too, by about the same margin (slightly more, in raw points). In this study, AI did not increase the raw calibration error. The same tendency to overestimate was already visible without it. Nor did it blunt the finer-grained judgment: asked after each problem how sure they were, both groups separated their right answers from their wrong ones at about the same rate, better than chance and below the level the field treats as moderate. The incentive produced no detectable improvement in calibration.

One more result, easy to misread. On this task, the human-and-AI pair did not on average beat the model working alone. That comparison does not show that people contributed nothing. It shows only that the combination produced no average advantage over the model alone. More than half the participants nevertheless scored above the model's average across its repeated runs.

Put that together and three things come apart that we habitually treat as one. There is what you and the tool produced. There is your estimate of what you and the tool produced. And there is what you could do without it. The adult studies at the centre of this essay measure the first two. They do not measure the third.

Something in the same family showed up in the Turkish surveys. The students whose exam scores had fallen did not report thinking they had done worse; the students in the guardrail group, for whom no exam difference was detected, reported thinking they had done better. Different instrument, different population, so this is a reading rather than a replication. But it looks as though perception tracked the help more closely than the outcome.

Why the work does not correct it

Why doesn't the gap close by itself? Because the work is not a test of you, and never pretended to be.

An assisted result reports on a combination. It is a perfectly good report: the analysis holds, the code runs, the recommendation survives the meeting. It does not contain a separate measurement of what the person brought to it. There is no line in the document for that.

Consider a preregistered field experiment involving 758 consultants at a global management consulting firm, randomly assigned after a baseline task. On work inside the model's competence, the assisted consultants completed 12.2 per cent more tasks and worked about 25 per cent faster. On a complex managerial task deliberately chosen to sit outside it, they were nineteen percentage points less likely to reach the correct answer than colleagues with no AI at all.3 The two kinds of task did not look different from the inside. That is the researchers' point. Their name for it is the jagged frontier, and nothing in the work announces which side of it you are on.

I want to mark the next step as mine rather than the evidence's, because nothing measured it. It seems to me that the better the combined result gets, the easier it becomes to read that result as information about yourself, because it is the only information in front of you, and because it is genuinely good news.

What is actually being kept, and where

None of this makes offloading a mistake, and the strongest version of the opposite case deserves saying first.

Thinking has always been distributed. Nobody extracts square roots by hand as a discipline of character. We keep notebooks so we can stop holding things in mind, colleagues so we can stop knowing everything, checklists because the mind that wrote the list is not the mind that will be tired at three in the morning. Handing a capacity to something outside yourself is not decline; it is most of what a civilisation is.

Often it is entirely rational. Offloading is safest when the output can be checked without recreating the delegated cognitive process. If the error is bounded (a wrong restaurant, not a wrong dose), if a test suite or a colleague or reality itself will catch the mistake, if that check actually happens rather than existing on paper, and if the capacity you freed goes somewhere you do need it, then the trade is a good one and the anxiety around it is misplaced.

It becomes structurally risky when only the delegated capability could reveal the error.

The Turkish study holds a small, uncomfortable illustration. When the researchers put the question the students actually asked (what is the answer?) to the plain-prompt model, it was right about half the time. Wrong reasoning on four problems in ten. The performance pattern contained little sign that the errors were being productively detected: arithmetic mistakes, which should generally have been easier to identify than logical ones, depressed practice scores by a similar amount. The authors read that pattern as more consistent with copying or disengaged use than with students reading carefully and being misled.4 Nobody watched each of them to confirm it.

There is an older and narrower analogy, used once and lightly. Among fifty regular drivers, heavier lifetime satnav use went with worse spatial memory when they had to navigate without it; thirteen of them, retested three years later, declined more steeply the more they had used it.5 Fifty people, then thirteen. Correlational, and nothing to do with AI. The analogy supports a narrower concern, not a general decline: when a delegated capacity changes, the change may be specific to the ability no longer being exercised. And you find out when you have to do it yourself.

The test the work does not contain

Which is the point, and it is less dramatic than it sounds. The test exists; it is simply separate. You can ask someone to explain the thing back. You can audit. You can run the simulation, the unaided spot check, the closed-book exam. None of that requires anyone to give up their tools. It requires somebody to ask a different question than is the work good?

None of the adult knowledge-work studies in this literature contains a delayed measurement of independent capability. The separate tests we do have come from students, in mathematics and introductory programming. They are not about the woman in the meeting, and they cannot be made to be.

What they show is that the test can be run, and that the answer depends on the task and the design. The Turkish students with the plain prompt scored lower unassisted; for the guardrail group no exam difference was detected. In a study of 69 novices aged ten to seventeen, where every programming task written with an AI code generator was followed by one the student had to modify by hand, the assisted group produced better code and showed no measurable loss on the independent task.6 In an exploratory preprint across five British schools, both comparison groups received tutoring. In the LearnLM condition, every AI-drafted message was reviewed and could be revised or replaced by an expert tutor before reaching the student. On the first unaided problem in the next study unit, the estimated success rates were 66.2 per cent for the supervised LearnLM condition and 60.7 per cent for human tutoring alone.7 A positive new-problem outcome, but from supervised AI tutoring rather than a person alone with an assistant.

Four outcomes across three different kinds of measure: an unassisted exam taken minutes later, a hand-modification task, and a new-problem test on a later topic. They answer different questions — what you can do immediately without the aid, whether you can modify code without the tool, and whether learning carries to a new topic — and a result on one says little about the others. In the studies available here, all of them belong to educational settings, not adult knowledge work.

What the work cannot tell you

So: what can be said, and what cannot.

Nobody has followed an adult's independent capability across years of assistance; the study does not exist, and it would be slow and expensive. The adult evidence at the centre of this essay measures calibration rather than capability, and it does not show that AI made the calibration worse: both groups overestimated their total scores by roughly four raw points.

The Turkish trial cannot say which part of its help worked, because nine things differed at once. One proposed repair, asking people to explain the answer back before accepting it, comes from the researchers themselves and already sits, unisolated, inside the Turkish guardrail prompt. Nobody has tested it on its own.

There is one more thing, and it is the strangest. The people who ran the reasoning studies did ask their participants to estimate what they could have scored without the AI. The number is sitting in a table: about 11.6 out of 20. It was never analysed, and it could not have been set against the same participants' actual unaided performance, because that was never measured. The closest measure found in the adult evidence reviewed here of what people believe they could do unaided is printed in a published paper, uninterpreted.

One sentence available here should be refused. The essay before this one was about children, and the obvious move is to say that the child who never worked alone becomes the adult who cannot think alone. Nobody knows that. These adults are not the children from the first essay grown older. No cohort connects them.

The work shows what you and the tool made together. It cannot show what you could make without it. That is not a loss. It is a different quantity, and it can be measured — by someone who thinks to ask.

Which leaves the experiment nobody has run. Give a person the assistance. Take it away. Pay them to predict their unaided score — and then make them take the test.

Does the estimate correct itself when the system is gone?