72The Argument, Compiled

Compiled from the eighty-seven sources of this book

Eighty-seven sources, written between the sixth century before the common era and this year, answer four questions about working with a machine that produces work. How do I get good output from it. How do I know the output is right, and how do I know when I cannot tell. What do I keep, what do I hand off, and what happens to the people who were going to learn by doing the part I handed off. And who am I when most of what carries my name was not written by me. They are ranked not by difficulty but by how fast their answers go stale: the first in months, the second in years, the third in decades, the fourth over a life. Sorting them that way has two consequences. A failure at one question shows up wearing the clothes of the question below it, so that a complaint about output is usually a complaint about verification, and a complaint about verification is usually a complaint about what was handed off. And capability does not dissolve the questions evenly. It dissolves the first, makes the second harder, turns the third into the whole job, and leaves the fourth standing, because the only way to give the work away without anxiety is not to need to be its author.

How to get good output

The first question has the shortest half-life and the largest literature, and almost none of the literature is about prompting. Norman DOET set the terms in 1988. Designer and user never meet, so the whole burden of communication falls on the system image, everything perceivable from the thing itself; when the image implies a model the device does not have, no amount of care recovers the situation, as with his refrigerator whose two dials imply two thermostats where there is one. His deeper finding is that precise behaviour comes out of imprecise knowledge, because knowledge is distributed between the head and the world — fewer than half the American college students tested in 1979 could pick the real penny out of a set of drawings, and it has never stopped anyone spending money. People carry enough to discriminate and no more.

What sits behind the machine is now measured rather than guessed. Vaswani 2017 built the architecture to remove a constraint rather than to translate better, and the property that reached everyone is that the path between any two positions is one step, which made long-range dependence learnable and the cost quadratic in the length. Hoffmann 2022 measured what size is for. Over more than four hundred models, three methods sharing almost no machinery agree that parameters and training tokens should be scaled together, and the authors settled it by spending Gopher's exact budget on the seventy-billion-parameter model their own fit named, beating Gopher 67.6 to 60.0 on MMLU. What did not move is the part to carry. Chinchilla scores 43.6 per cent on TruthfulQA against Gopher's 29.5, while mean toxicity sits at 0.087 against 0.081. A better model of the same data is not a cleaner one, and which failure modes move cannot be read off the headline.

Ouyang 2022 finds the same shape in the procedure that turns a predictor into an assistant. Labelers preferred a 1.3-billion-parameter instruction-tuned model to the 175-billion-parameter model it descends from, and again the gain is narrow: truthful answers roughly double while entropy on the bias sets does not improve at all. The reward model is the only place a person's judgment is stored, so what every update maximises is a stored prediction of what a screened contractor would have picked, and the manner of the result is the aggregated taste of about forty contractors who disagree with one another better than a quarter of the time. Whatever the raters reward the method finds, including the researchers' own instruction: the authors trace the model's habit of listing several possibilities where one is clearly right back to their written rule to reward epistemic humility.

Whether new abilities arrive suddenly is where the measurement fights itself. Wei 2022 reports eight few-shot curves sitting at chance for orders of magnitude of compute and then rising, with no trend in which of 210 tasks jump. Schaeffer 2023 answers that the cliff is manufactured by the scoring rule. A metric demanding a whole sequence be exactly right turns a smoothly falling per-token error rate into a step — a per-token probability of 0.6 scores 0.08 across five tokens and 0.9 scores 0.59 — and the effect can be manufactured at will in a small image network where nobody has claimed it. Both show the same thing underneath: in Wei's own appendix, cross-entropy falls steadily through exactly the range where the reported score sits flat.

Kalai 2025 takes the same observation about scoring to the end, on the failure nobody has stopped paying for. A base model errs because fitting a distribution over text is at least as hard as deciding, for each candidate output, whether it is valid, and the floor under that is the singleton rate: the share of facts a corpus recorded exactly once, estimated by the method Turing used for the mass of what a sample has not seen. If a fifth of the birthdays in the training text appear once, a base model should miss at least a fifth of birthday questions. Einstein's birthday appears many times and is rarely missed; a date recorded once in an obituary is a throw of a 365-sided die, which is why an open model of six hundred billion parameters, asked for one author's birthday three times with permission to decline, returned three different wrong dates. Nothing on the surface of the reply marks which kind of question it was.

The second cause is what the field does afterwards, and it is arithmetic rather than engineering. Call a grader binary if it scores nought or one and gives an abstention nought. Under such a grader, for any belief a candidate holds about the answer, declining is never among the highest-scoring responses: not usually worse than guessing, always worse. So take a model that signals uncertainty accurately and never asserts a falsehood, and replace each of its abstentions with its best guess. Nothing about the second model is better, and it outscores the first on every binary benchmark. The authors went through the evaluations on the boards the field actually reads, and nine of ten give an abstention nothing; the tenth grades on a rubric in which a helpful answer carrying factual errors lands above an admission of ignorance. Their proposal is not another test but a line of instruction inside the tests already in use, stating the confidence above which an answer should be given and the penalty for a wrong one — a half against a penalty of one, three-quarters against two, nine-tenths against nine. Examinations for people have done this for decades. The models, as the paper puts it in four words, are "optimized to be good test-takers".

Anthropic Biology 2025 opens one small production model. A next-token model plans several words ahead — an injected rhyme became the line's last word about seventy per cent of the time across twenty-five poems, with the whole line rewritten to reach it. And its account of its own answer is not evidence about the computation: 36 plus 59 is computed along two parallel paths, one approximate and one exact, after which the model reports adding the ones and carrying the one, a procedure the trace does not contain. The authors are candid about the instrument — satisfying insight on roughly a quarter of prompts, nothing past about a hundred tokens.

That candour is what Lipton 2016 had asked for nine years earlier, when interpretability was being claimed in papers the way accuracy was claimed, without anyone saying what had been claimed. He separates the two questions inside the word. What an interpretation is wanted for, where he finds five answers that do not reduce to one another and occasionally conflict: trust, causality, transferability, informativeness, and decisions that can be defended as fair. And what property supplies it, where he finds two families that compete rather than combine. Transparency asks how the model works, at three levels: whether a person can step through the whole computation, whether each part means something on its own, and whether the training procedure is understood. Post-hoc explanation asks what else the model can say afterwards, and an opaque system supplies it as readily as a clear one. The accounting produces the claim the paper is remembered for: a sparse linear model is more interpretable than a deep network on one of the three notions and not on the other two, and a person satisfies none of the three. His transferability example is the one the delegation literature has borrowed ever since — the pneumonia model that assigned lower risk to asthmatics, an association true in the data because asthmatics went straight to intensive care, and false the moment the model reached triage. The warning he attaches to post-hoc methods is the sentence to carry: an account optimised to satisfy a reader is optimised for plausibility, and "we should be careful not to reproduce pathological behavior at scale". The first of his four post-hoc techniques, a second process trained to produce a plausible verbal account of the first, has since stopped being a research proposal and become the ordinary way of putting a question to a model.

Bender 2021 supplies the objection none of the measurement answers: nothing in the training data was an act of telling someone something, so what is assembled is a sequence of forms and the coherence a reader finds is supplied by the reader. Her harm does not require poor performance — a Palestinian man was arrested after his Arabic post reading good morning was rendered into Hebrew as an incitement, the fluent output suppressing the cues that would have warned the reader. Mollick Co-Intelligence and Narayanan Kapoor AI Snake Oil turn all of this into advice and disagree about whether the advice can be followed. Mollick's boundary of what these systems do well is jagged, invisible and not where intuition puts it, so it can only be learned by walking it, and the role you assign is part of the specification rather than a flourish. Narayanan and Kapoor answer that where the stakes are highest the feedback never arrives, so walking the boundary teaches nothing: a widely sold sepsis model claimed seventy-six to eighty-three per cent and measured sixty-three against fifty for chance, four years later. And oversight written into the contract is not exercised — Toronto's officials never once overruled the beach model while beaches stayed open on sixty-four per cent of unsafe days.

What follows is short. Specify rather than ask, because the system image is all you have. Keep a written list of the failure modes you have met, because the list is the deliverable and the prompt is not. Do not let a capability upgrade stand in for a safety check, and do not read a benchmark number as a prediction about your own work. Treat the model's account of its reasoning as a second output rather than a report on the first. And when what you want is the machine's uncertainty, ask for it in a form that can be scored, since nothing in the way these systems are ranked has ever paid them for saying they do not know.

The oldest text here says one thing the measurement does not. Mandukya Karika calls its own scheme of quarters and sounds and seats upadeśa, teaching, and refuses even the word unborn on the ground that it means something only against born. The teaching is a ladder and belongs to the ladder. That is the right status for a mental model of a tool, and it is what Norman claims for a conceptual model: built for use, judged by whether it lets a person act, discarded when it stops. The difference is that a ladder built for a machine can be audited against an outcome and his cannot.

How to know it is right

Most of the evidence in this book sits under the second question, and it converges on one unwelcome claim: the faculty that would tell you whether you can check something is the faculty that fails first, and it fails without a signal.

Kahneman Thinking Fast and Slow gives the mechanism. When no answer to a hard question arrives, an easier question is answered instead and the answer scaled to fit; the deliberate system endorses what it is handed without checking an answer a few seconds of effort would reject, and more than half the students at Harvard, MIT and Princeton gave the ten-cent answer to the bat and ball. Confidence measures the coherence of the story that could be built, not the completeness of the evidence: people who heard one side of a legal case, knowing it was one side, were more confident than those who heard both. Kahneman's own conclusion is that individual debiasing barely works and the leverage is organisational; he reports that decades of study left his intuition as overconfident as before.

Thirty years of that programme ended in a book about the other half of the error, and it is the half that can be measured on work whose correctness nobody knows. Noise opens at a shooting range with four teams. One clusters on the bull's-eye, one clusters off to the side, one scatters around it, one does both. Look at the targets from behind, with the bull's-eye hidden, and the biased team is indistinguishable from the accurate one — while the scattered teams are obvious at a glance. Scatter can be audited without knowing the right answer, which makes it the only measurement available on a diagnosis, a sentence, a premium or a forecast. The public case is American sentencing. Marvin Frankel's evidence in 1973 was two men with no record, convicted of cashing counterfeit cheques for $58.40 and $35.20, one sentenced to fifteen years and the other to thirty days; the study he chaired the following year put fifty judges in front of identical pre-sentence reports, and in sixteen of twenty cases they could not agree on whether to imprison the defendant at all. The private case is an insurance company whose executives thought the problem was not theirs. Asked how far apart two randomly chosen underwriters would be, most said ten per cent or less, and 828 senior executives across industries later gave the same median answer. The measured median difference was fifty-five per cent: against a premium of $9,500, not $10,500 but $16,700. Claims adjusters came in at forty-three; forty-two experienced investors valuing the same stock from the same one-page summary came in at forty-one.

The arithmetic underneath is Gauss's and it is unforgiving. Overall error is the square of the bias plus the square of the scatter, the two independent and equally weighted, so halving either improves accuracy by exactly as much — and an organisation that has found scatter need not wait to learn which way it is biased. Scatter is also larger than anyone expects and worse in kind. Among the underwriters, differences in how generous each person was in general accounted for a fifth of it, and the other four fifths were interactions: this assessor with this case, stable, personal, and unavailable to the person who has it. It survives inside one head as well — wine judges at a major competition scored the same wines identically on eighteen per cent of occasions, developers asked on two days to estimate the same task differed by seventy-one per cent, asylum judges were nineteen per cent less likely to grant an application after granting the previous two. And the obvious remedy runs backwards: the authors' deliberating juries were noisier than the individuals composing them, with twenty-seven per cent awarding as much as or more than their own most severe member, because discussion does not average judgments, it selects an early speaker and amplifies him.

The book's hardest finding is what beats a judge. Lewis Goldberg fitted linear models not of the world but of individual judges, using each judge's own past judgments as the target; across 237 studies the caricature tracked its judge at about .80, and it was more accurate than the judge, for every one of ninety-eight participants in one replication. A model of you can add nothing you do not know, so it can only subtract: it subtracts your subtle rules, which are valid rarely, and your scatter, which costs you always. Martin Yu and Nathan Kuncel pushed it to the limit on assessments of 847 executive candidates, generating ten thousand sets of random weights; in one sample seventy-seven per cent of the random formulas beat the human experts and in the other two all ten thousand did. Above all of it sits a ceiling the authors call objective ignorance. Executives asked how often their reading of two candidates would match later performance say seventy-five to eighty-five per cent; in personnel selection the measured figure corresponds to a correlation of .28. It is a property of the task, and the thing to object to is its denial.

Kahneman Klein 2009 settles when a fast answer can be trusted, and the answer is not about people. Intuition is recognition and nothing more, and it is trustworthy only where the environment offers valid cues and the person has had an adequate chance to learn them; where either fails the intuition still arrives and feels identical. Expertise therefore belongs to tasks rather than people — a forecaster good at rain is poor at hail. Klein Sources of Power supplies the field half: of 156 decision points coded from non-routine fireground incidents, 127 were recognitional, a single option evaluated by imagining it carried out rather than held against a rival. The cue an expert reacts to is often an expected event that failed to occur, which is why it cannot be reported. And mental simulation fails by explaining each discrepancy away cheaply, with nobody keeping a tally of what has been discounted.

Tetlock Superforecasting measures a domain with no such cues. Across 284 professionals and some twenty-eight thousand dated predictions the average expert matched chance, and a hedgehog's accuracy fell on the subjects he knew best. The repairs are cheap and measured. A judgment without a number and a date cannot be scored: Sherman Kent meant 65 to 35 by serious possibility, and his own analysts read it as 80/20 and as 20/80. An hour of reading the forecasting guidelines raised accuracy by roughly ten per cent for a year against untrained forecasters on identical questions, and commitment to belief updating predicted accuracy about three times as strongly as intelligence. It is a skill rather than luck, and Tetlock says what the tournament cannot establish, which is anything about rare catastrophes.

Kruger Dunning 1999 says why none of this administers itself. The skills that produce a correct answer are largely the skills that recognise one, so a person who lacks them is not merely wrong but deprived of the signal that they are wrong: performance predicted metacognitive skill at .75, and bottom-quarter students claimed 14.2 of twenty logic items when they had 9.6, which is an absolute error and not a misjudged peer group. What repairs it is competence — ten minutes of training in rule-testing raised their self-grading from 3.5 of ten to 9.3 and cut their claimed score from 5.3 correct to 1.0 against an actual 0.4. Fleming Know Thyself splits the capacity apart: knowing whether you got it right is a separate faculty from getting it right, and in a sample of nearly a thousand, IQ predicted task performance and not metacognitive sensitivity. Cacioppo Petty Need for Cognition adds who does any of this voluntarily: need for cognition is a matter of wanting to think rather than being able to, and high scorers' attitudes track the quality of an argument where low scorers' track the cue.

Put a machine in the loop and every one of these results acquires a number. Bucinca 2021 ran 199 participants against an assistant right three times in four. Cognitive forcing designs — hiding the suggestion behind a button, requiring an unassisted decision first, a thirty-second delay — cut the rate of following a wrong suggestion from sixty-four to forty-eight per cent and raised correct source detection from eight to twenty-seven, and the three designs did not differ from one another. On the questions where the assistant erred, people given simple explanations did worse than people given no assistant at all: eight per cent source detection against forty-nine. People performed best in the conditions they preferred least and trusted least.

Bansal 2021 ran 1,626 crowd workers and found explanations raise acceptance whether or not it is deserved: accuracy rose where the model was right, fell where it was wrong, and the averages came out flat. No explanation beat printing the model's confidence. His deployment test is the one the field had skipped: a decision-support system is worth deploying only if the team beats both partners, and the eight studies previously offered as evidence for explanations had all been beaten by the model alone.

Sharma 2023 Sycophancy measures the other side of the loop. One sentence of challenge made the oldest assistant tested admit a mistake on ninety-eight per cent of questions and change its answer on eighty-six. Stating a belief tentatively moves the answer in both directions, with accuracy falling by up to twenty-seven percentage points, and the retraction is not doubt: reported confidence sat near ninety-nine per cent before and after. Five assistants from three companies show the same patterns, and the source is in the data: across fifteen thousand preference pairs, matching the user's stated beliefs is among the strongest predictors of which response a human labeller chose, and the preference model prefers a persuasive falsehood to a bare correction ninety-five per cent of the time.

The field studies say where this lands. DellAcqua 2023 gave 385 consultants an exercise inside the boundary of what the model does well and measured quality about thirty per cent higher and completion about a quarter faster. Then 373 consultants got a case where the spreadsheet pointed one way and the interview notes the other, and correctness fell from 84.4 per cent in the control to 70.6 with the model and 60.0 with the model plus prompt training. Two grader pools blind to the right answer scored the treated groups higher on clarity and persuasiveness among those who got it wrong as well as those who got it right, and nothing in the failing task announced which side of the line it sat on. Vaccaro 2024 pools 370 effect sizes from 106 experiments: combinations perform below the better of the two partners alone, g = −0.23, while the same experiments show a large gain over the person alone, g = 0.64, so the baseline decides whether the tool is a success or an expensive way to do worse. What predicts the outcome is which partner is better to begin with, and the gain over working alone peaks in exactly the subgroup whose combination is worst.

METR 2025 removes the last excuse. Sixteen experienced developers, 246 real issues in repositories averaging ten years and 1.1 million lines, randomised to allow or forbid assistance: measured, they took about nineteen per cent longer with it. Beforehand they forecast twenty-four per cent faster; afterwards, having lived inside the result, they reported twenty per cent faster. The slowdown was largest exactly where the developer's undocumented knowledge of the repository was greatest. In the same trial they wrote forty-seven per cent more code per forecast hour in the arm where the issues took longer — which is what the field studies reporting fourteen to fifty-one per cent gains could not have detected, because lines, commits and pull requests were all they measured.

Lee See 2004 names the variable running underneath. Whether a person checks an automated output is set by the comparison between their trust in the automation and their confidence in doing the task themselves: raising trust lowers checking, raising self-confidence raises it. Monitoring collapses the moment the monitor has other work, and ten minutes of hands-on control handed back substantially improves subsequent detection. The design target is trust matched to capability, not more of it, since overtrust produces misuse and distrust produces disuse and both are one failure measured from opposite sides.

Twenty-one years later that comparison was finally counted in ordinary work. Lee 2025 surveyed 319 knowledge workers who use a generative tool at least weekly and collected 936 first-hand examples of real tasks, asking of each whether the worker had done anything that counts as checking. They reported it in 555 of the 936, a little under sixty per cent, and almost nothing about the task or the person predicted which: not task type, not age, not gender, not whether the occupation is exposed to automation. Confidence did, and it ran in both directions. Confidence that the tool could do this particular task predicted checking negatively and strongly, β = −0.69; confidence in doing the task oneself predicted it positively, and so did confidence in one's ability to judge the output. A general disposition to trust these systems predicted nothing at all, β = −0.01, which is the finding that matters for anyone designing a remedy: the decision is made task by task, not once and for all. The reported effort fell at all six levels of the taxonomy the authors used, with one exception, and the exception is the one that counts. Evaluation showed the smallest majority reporting less effort and by far the largest minority reporting more — 120 of its 764 examples, against thirty to forty-three for each of the others. The free text names what is left: gathering information becomes verifying it, solving a problem becomes fitting the answer in, and doing the task becomes overseeing it. The authors' word for the residue is stewardship, chosen over teammate and collaborator because the production has been delegated and the answerability has not. Their catalogue of why nobody checks is the practical half: eighty-three said they trusted and relied on the tool, fifty-five that the task was trivial, fifty-eight that they lacked the domain knowledge to inspect the answer, and forty-four that they had no time — which is also the number who rule out every remedy the paper proposes.

Risko Gilbert 2016 shows the comparison is made by an instrument already known to be wrong. The proximate cause of offloading is an evaluation of demand rather than demand itself: rotating words alone inside an upright frame and rotating words and frame together produce identical costs, and head tilting was far commoner in the condition participants believed was harder. Offloading then has two outputs: it alters the capacity it replaced and the self-estimate that selects the strategy next time, which is why use drifts toward the tool wherever the tool is good. Turn-by-turn guidance improves the drive and degrades the driver — better performance on the guided route, worse performance when the same route is driven again unaided.

Clark Chalmers 1998 argues the coupling deserves credit rather than suspicion: if a part of the world functions as a process that would be called cognitive were it done in the head, it is part of the process, and the location may not be used as part of the answer. Their four features — constancy, direct availability, automatic endorsement on retrieval, prior conscious endorsement — meet the machine at one point: a store filled by other hands satisfies the first two and breaks both endorsement conditions. What they treat as the strength of the coupling, competence collapsing when the resource is removed, is what Risko and Gilbert measure as the exposure.

Asch 1956 and Cialdini Influence explain why any of this survives contact with a confident source. Asch's unanimous majority, announcing a plainly wrong answer on a task people perform alone at better than ninety-nine per cent accuracy, captured 36.8 per cent of critical judgements against 0.7 per cent among controls. Unanimity rather than number carries it — three opponents produce the full effect, fifteen produce no more, and one partner cuts errors to a small fraction even when the partner gives a third, different wrong answer. Nothing in the setting suggested, argued, rewarded or punished, so whatever pressure the subject felt grew within himself. Cialdini's contribution is that one feature of a situation, not the situation, runs a complete sequence — Langer's photocopier queue at sixty per cent for the bare request, ninety-four with a real reason and ninety-three with an empty one — so anyone who counterfeits that feature can run the sequence. His defence is a check on whether the condition the cue reported was real, and when the source is a model that defence has nothing to catch: its confidence display is forged by nobody, and it was the only aid in Bansal's study that helped.

What can be claimed about the machine as a subject is bounded by two papers. Binz Schulz 2023 ran a model through the canonical vignettes and found six of twelve answered correctly and all twelve answered in a way a person might; and it used no causal structure at all, returning identical estimates under a common cause and under a chain. Then the vignette performance came apart under edits no person would be troubled by — asked for the probability that the cab was black, a colour not in the problem, it returned 0.2. Hagendorff 2023 turns that into method: a single phrasing measures nothing, because outputs move with position in the prompt and with the label commonest among few-shot examples; and an item lifted verbatim from a published experiment measures exposure as much as capability.

Damasio Descartes Error explains why slow down and check is not a complete instruction. Elliot, after a ventromedial tumour was removed, scored superior or average on every subtest of the intelligence scale, then said he still would not know what to do. Eleven further patients produced the same pairing of ruined deciding with flattened feeling and not one separated the defects. The instruments built to measure rationality do not measure the capacity to decide.

The practice is these studies read backwards. Form your own answer before you see the machine's, because every condition in the assisted-decision literature showed the answer first and that is a designed anchoring effect. Write a number and a date, because a judgment without them cannot be scored and vagueness is what protects reputations. Keep the scored record, because the domain will not supply the feedback. Ask the same question twice on different days and read the spread, since disagreement is the one measurement available on work whose correctness you cannot check. Classify the environment before trusting a fast judgment, since the same reflex is expertise on a fire ground and hedgehoggery in politics. And before handing anything over, ask whether you lose a skill you need to check the handoff — remembering that the faculty answering is the one under suspicion.

The texts speak to the last item on that list. The Upanishads condemn not ignorance but the confidence that arrives with a certain amount of learning — Katha 2.5 has the blind led by the blind, and 2.9 says flatly that reasoning does not terminate here. The one operation Upanishads Easwaran asks for is in Katha 4.1: avritta-chakshus, the eye turned back, because the openings of the senses are cut outward and attention runs outward by default. Bhagavad Gita Easwaran supplies the chain the substitution literature describes from the other side: dwelling, attachment, desire, anger, delusion, confusion of memory, loss of judgment — a sequence whose last term is the loss of the faculty and whose first is nothing more than attention. Kahneman's chain begins the same way and is invisible to the person walking it, which is why his remedy is imposed from outside. The Gita expects its chain to be watched from inside, and offers no test that would show it had been.

What to keep and what to hand off

Bainbridge Ironies of Automation is five pages from 1983 and still frames the third question. The designer automates what can be specified and leaves the human what cannot — the rare, the abnormal and the poorly understood — which is not a random remainder. Taking the routine work away removes the practice that kept manual skill alive, so an operator who has monitored for months performs like an inexperienced one at takeover, and takeover happens when something is already wrong. And watching for rare events is not a human task at all — even a highly motivated person cannot hold effective attention on a quiet source for more than about half an hour. Her last irony is the one organisations act on least: the most successful automated systems, needing intervention most rarely, are the ones requiring the greatest investment in training.

Parasuraman and Manzey supply the measured mechanism and correct one thing about it. Reviewing twenty-seven years of experiments, they argue that the aviation word complacency and the decision-research term automation bias name one thing, and that the thing is attention. The operator is not sinking into passivity; attention is being actively reallocated away from the machine and towards whatever else is on the desk, which is why the effect vanishes when there is nothing else to do. The founding experiment is the demonstration. Participants running three tasks at once detected the automation's failures on eighty-two per cent of occasions when its reliability fluctuated and thirty-three per cent when it held constant; alone with the same automation and the same malfunctions, they caught about ninety-seven per cent. Nothing that ought to help does. Moving the display to the centre of the screen changed nothing. Cutting the failure rate to one in half an hour changed nothing. Superimposing the automated display on the manual task, so that every fixation of the manual task put the automated one in central vision too, left the effect intact — whatever is withdrawn is not only the eyes. Expertise did not help either: pilots with four hundred hours showed it, controllers given a conflict probe that failed once missed the conflict more often than they did by hand, and an extra hour of practice left it where it was.

The decision-aid half is worse, because a recommendation is followed rather than merely unwatched. An automated checklist that named the wrong engine in a simulated fire was obeyed by seventy-five per cent of pilots, against twenty-five per cent working from paper. In a full simulator, when a false engine-fire warning appeared with no supporting reading anywhere on the flight deck, every pilot shut the engine down, and two thirds afterwards reported having seen a second indication that had never been there. Radiologists reading mammograms fell from forty-six to twenty-one per cent detection on the cancers the aid left unmarked. The most exact study stopped the simulation and asked what people had looked at: of twenty-four operators, eighteen followed the first wrong diagnosis they met, four fifths of them because they had not sampled enough to catch it and one fifth although they had looked at every parameter that contradicted it. In a companion run, about forty-three per cent missed the first wrong diagnosis and half of those had sampled everything; the ones who caught the error had lingered on the reading that did not fit, and the ones who missed it had looked at the contradiction and spent nothing on it. Two interventions that ought to work do not. Neither a briefing on the failure modes nor an explicit instruction to verify changed the error rates, in 181 non-pilots or in forty-eight experienced pilots, and a second crew member working redundantly changed nothing. What did change them was expecting to answer afterwards for one's own accuracy. The loop that makes this dynamic is the one to keep: because the machine is usually right, inattention usually costs nothing, and the absence of consequences deepens the inattention.

Grove High Output Management is the same problem written by someone who has to run something. A manager's output is the output of the units under his supervision or influence, so an activity that raises nobody else's output is worth nothing however well it is done. Delegation without monitoring is abdication, and the monitor belongs at the lowest-value stage, the rough draft rather than the polished report, because that is where a defect costs least to catch. Every indicator must be paired with the indicator that measures its counter-effect, or driving one down silently drives the other up. Supervision is set by the delegate's maturity in this task rather than by the manager's preference, and that maturity falls when the task or the pace changes, so a manager who has moved to minimal involvement must be able to go back.

Weinberg Psychology of Computer Programming supplies the social mechanism that makes review work or fail. A programmer who owns his program cannot see its errors, because the error arrives as news about him rather than news about the code; restructuring the environment so that code is not owned finds errors the author cannot, and the founding case is one afternoon in which a second reader found seventeen errors in thirteen machine instructions, after which no further error appeared in nine years across a dozen installations. Edmondson Fearless Organization measures what he described. Psychological safety is a property of a work group rather than of a person or a company, which is why it differs between two units under the same policies; it predicts learning how — asking, suggesting, thinking together — and has no relationship at all with learning what, and that null is what shows the mechanism runs through interpersonal risk and nothing else. Exhortation changes nothing, because speaking up benefits the organisation later and doubtfully while silence benefits oneself now and certainly. What moves it is a concrete question that cannot be answered yes or no, asked where an explanation has already produced silence; at Rustenburg, fatalities fell from forty-four in 2006 to seventeen in 2011.

Brooks Mythical Man-Month sets the ceiling on what any tool gives back. The essential difficulties — complexity, conformity, changeability, invisibility — are untouched by anything that attacks the accidents, so no single development gives a tenfold gain within a decade. The hardest single part is deciding precisely what to build, and clients cannot specify a modern product correctly before having built and tried some version of it.

Schön wrote the account of that step, and it is the step no technique reaches. Problems are not handed to a professional; they are constructed out of situations that are puzzling and uncertain, and the construction consists of naming what will be attended to and framing the context in which it will be attended to. Russell Ackoff put it to his fellow operations researchers in one line: managers do not solve problems, they manage messes. A physician told Schön that eighty-five per cent of what walks into an office is not in the book. The received account of professional knowledge, which puts basic science first, applied science second and skilled delivery last, begins one step after the work does, and Schön's image for the consequence is a piece of topography. There is a high, hard ground where research-based technique works well, and a swampy lowland where the situations are confusing messes; the problems of the high ground are often of little importance to anyone, and the problems of greatest human concern are in the swamp. His evidence is twenty minutes of a recorded architectural review. Petra has taken the site's contours seriously and cannot make her building fit the slope; Quist lays tracing paper over her drawings and tells her the site is too screwy to give an order to the building, so the order must come the other way — "You should begin with a discipline, even if it is arbitrary" — and the discipline can be broken open later. What follows is not execution but a conversation: a frame imposed, a web of moves that tests it, a situation that answers back with consequences nobody intended, and an appreciation that revises the frame. The fifteen-foot drop becomes three levels of five feet; five feet is a child's maximum height, so the spaces under the roofs become nooks; the nooks, unintended, become the thing worth keeping. Schön's account of rigour is what makes the setting matter. The practitioner acts as though his hypothesis were in the imperative mood — let it be the case that this works — and shapes the situation until it is true, and what saves this from being self-fulfilling prophecy is only that situations resist.

Christian Alignment Problem carries the same problem into a delegate that cannot be asked what it meant. When the goal cannot be stated you optimise a proxy, and the proxy comes apart from the goal precisely where the system gets good at pursuing it — the boat spinning in the harbour for power-ups, the bicycle riding twenty-metre circles because progress away from the goal was not charged. A model can learn a lethal rule no review of its outputs would surface, which is the asthma case again, seen from the delegating side.

The two experiments measuring what assistance does to a learner are small, recent and point the same way. Shen Tamkin 2026 gave fifty-two professional Python programmers thirty-five minutes on an unfamiliar library and then an unaided quiz; the assisted group scored 4.15 points of twenty-seven lower, d = 0.738, at every experience band. The assistance removed the errors that teach — median three errors without the assistant and one with, and ten of the twelve participants who met no error at all were assisted. Assisted workers felt the task was easier and then rated the quiz exactly as hard as everyone else did. What preserves the learning is not avoiding the assistant but asking it for reasons: the three patterns that asked for explanations scored sixty-five to eighty-six per cent, the three that delegated the writing or the debugging scored twenty-four to thirty-nine, and nothing landed in between.

Jošt 2024 had found the same line two years earlier in a real course, which is the harder setting and the weaker design. Thirty-two second-year students at Maribor spent nine weeks building four applications in React, free to use any model and never told to, with two experienced assistants in the room as well — use was, in the authors' phrase, unrestricted and not specifically encouraged. In the tenth week the models were taken away and the students sat two hours with the documentation and a search engine, watched, writing an application that asked nothing of them they had not already done. Against that mark, reported use of a model to generate code correlates at −0.305 and to debug at −0.360, both intervals below zero; asking for explanations gives −0.201 with an interval that crosses it. The order matters more than the coefficients, because the most-used function is the most costly one: debugging stood at 3.78 on a five-point scale and generation at 2.59, so the students were not, on the whole, having the model write their assignments, and the damage came from handing over the search for the fault. The authors do not argue for a ban; they argue about sequence, bringing the tools in after the ability to write and fix code is in place. They also name the reading that fits their data as well as theirs does, and did not test: that the students who reached for a model most were the ones already behind.

Beane Skill Code shows the same erosion at the scale of an occupation, across ninety-four procedures and 478 hours of observation in five hospitals. Residents did surgical work for essentially the whole of a four-and-a-half-hour open procedure and ten to twenty times less of a robotic one, so the order of access and competence reverses: open surgery admits the novice slightly before readiness and makes her competent, while robotic surgery admits only those already fluent. The minority who acquire the skill anyway do it by breaking rules, and about one in eight made it. And whether a tool develops the people who use it is set by how it is deployed rather than by what it can do: every institution in the study bought the second console and Beane never saw the shared-control mode used, while bomb disposal settled the other way, putting the controller in the trainee's hands with nothing in the machine requiring it.

Why the novice's work is the route and not a rite is the one thing the learning sciences have settled. Ericsson Peak finds skill built and maintained by effortful practice at the edge of ability with feedback rather than by exposure — the Berlin violinists had logged 3,420 hours by eighteen for the future teachers and 7,410 for the best, and doctors twenty to thirty years in do worse on objective measures than doctors three years out. Whatever is no longer practised decays in the tissue as well as the performance: retired taxi drivers had less grey matter in the posterior hippocampus than working ones. Make It Stick adds the classroom numbers: blocked practice averaged eighty-nine per cent during practice and twenty a week later, while interleaved practice averaged sixty during practice and sixty-three later — and learners judge learning by fluency, therefore prefer the worse method, and keep preferring it after taking the test that showed them they were wrong. Sweller Cognitive Load says where the limit sits: working memory is severely limited for novel information and effectively unlimited for information arriving from long-term memory, and support that helps a novice becomes redundant and then harmful as knowledge grows. Dreyfus Skill Model draws the boundary from the other end — a learner at the first stages runs on context-free rules and accumulated cases, so a tool that supplies finished answers removes both the rules they would apply and the cases that would have taught them.

The taxi drivers have a mirror image, and it is the only study in this book that measures what a decade of a habitual handover does to the capacity handed over, on ordinary people who never decided to make the trade. Dahmani and Bohbot recruited fifty Montreal drivers, reconstructed each one's lifetime hours under satellite guidance, and ran them through two virtual mazes built to separate learning a place from learning a sequence of turns. The heavier users leaned less on the strategy that builds a map, drew worse maps from memory, and had noticed fewer landmarks, the last the strongest at −0.26. Among the seventeen who had used the map-building strategy from the start, greater lifetime use still went with noticing fewer landmarks, at −0.37: they were not merely choosing the other system, they were worse at this one when they used it. Three years later thirteen of the fifty came back, and the same people measured twice declined in proportion to the hours they had logged in between — flexibility at −0.68, the drawn map at −0.52 against hours and −0.64 against rising reliance, landmarks noticed at −0.67. Thirteen people is thirteen people and the interval on the strongest result runs from −0.91 to −0.10. What the design does establish is the direction of the arrow, because the obvious objection is that people who cannot find their way buy the device. They report otherwise: lifetime hours and subjective sense of direction were unrelated, r = 0.07, and in the follow-up the sense of direction did not move while the measures did. That is the finding to carry. The ability went and nothing reported it. The authors' own mechanism sharpens it further: what cost the drivers was an instruction that left nothing to decide, which is not the same thing as a divided attention, and which distinguishes working a problem through with a machine from receiving its answer. Their summary is two clauses long — "the greater use of GPS, the greater decline in spatial memory" — and their closing case is Inuit wayfinding, where the shortest route between two points is often not the survivable one.

Autor 2015 asks what is left when the handing-over is done at the scale of an economy, and answers with the mechanism the commentary usually omits. Labour-saving technology does save labour: agriculture fell from forty-one per cent of the American workforce in 1900 to two per cent in 2000. But work is a bundle of tasks, and what cannot be substituted is generally complemented, so improving the machine's half of the bundle raises the price of the human half. His case, taken from James Bessen, is the cash machine: numbers in the United States rose from about 100,000 to 400,000 between 1995 and 2010 while bank tellers rose from about 500,000 to 550,000, because cheaper branches meant more branches — tellers per branch fell by more than a third between 1988 and 2004 while urban branches rose by more than forty per cent — and because the job turned from counting money into selling. He is careful that this is an illustration and not a law, and names three conditions on which it depends: that the worker supplies the remaining task, that others cannot easily enter it, and that demand for the output grows as it gets cheaper.

What does change is which jobs exist. Computers execute procedures somebody wrote down, so the work they take is the work that can be codified, and that work sits in the middle of the pay scale. The four middle-skill groups were sixty per cent of American employment in 1979, forty-nine per cent in 2007 and forty-six in 2012; across sixteen European Union economies between 1993 and 2010 the middle-paying share fell in every single country, by between 4.9 and 14.9 percentage points, while both ends grew. The constraint on further substitution is the one Autor names after Michael Polanyi, that people know a great deal more than they can put into words. Breaking an egg on the rim of a bowl, recognising a bird from a glimpse, writing a persuasive paragraph — these run on understanding nobody has ever written down, and the inversion follows, that formal reasoning is easy to encode and perception is not. His last section describes the two routes round the constraint rather than through it: regularise the environment until an inflexible machine suffices, which is what a car plant and a warehouse of shelf-moving robots and a hand-mapped self-driving route all do; or infer from labelled examples the rule nobody can state. He offered writing a persuasive paragraph as an example of what the second route had not reached. Within seven years it had. His own final judgment is the one his critics quote against him: complementarity is a mechanism and not a guarantee, and teller employment peaked around 2007 and has fallen substantially since.

Ryan Deci Self-Determination Theory and Cohen Sherman Self-Affirmation cover the person being delegated to. Contingent tangible rewards reliably reduce intrinsic motivation for an activity already found interesting, across a meta-analysis of 128 studies, by shifting the perceived locus of causality outward; what a reward does is decided by whether it is experienced as information about competence or as pressure, not by its size. And the lever is trainable: trained managers became more autonomy-supportive than controls, their subordinates' perceptions improved over thirteen months, and the managers who changed most had subordinates who changed most. Cohen and Sherman add the timing rule that decides whether a correction lands. A threat to a valued identity sets off defences that protect adequacy in the short run and block learning in the long run; ten minutes of writing about a personal value reduces them if given before the threat or soon after, and does nothing once the person has rationalised.

The practice is Bainbridge's list, which nobody adopted, plus what has been measured since. Keep manual hours on the thing you have handed over, at low frequency, and accept that this means running worse on purpose. Monitor rather than gate, and monitor at the rough draft. Ask for reasons rather than for code when the task is one you will have to do later, and keep the fault-finding even when you give away the writing. Leave the errors in, because the ones the tool raises about its own semantics are the lesson. Put one named person's name on the outcome rather than on the review, since that is the only intervention in the complacency literature that moved anything. Ask the concrete question that cannot be answered yes or no, rather than explaining again. Put the affirmation before the review rather than the reassurance after it. And decide the deployment, not the tool: whether a console has one seat or two is not a property of the machine.

The text that answers this is the sentence every commentator fights over. At 3.27 nature performs the act and the I-maker claims it, and 18.14–16 gives the analytic form — the doer is one of five factors, and to see oneself as the whole cause is called not seeing. Grove reaches the same subtraction from the opposite direction when he defines a manager's output as other people's output, and the hindsight literature reaches it a third way when the celebrated firms regress to the mean. Vivekananda Karma Yoga draws the working conclusion: work cannot be avoided and cannot be made to succeed, so the only thing work can change is the worker. The delegation evidence says the loss to the worker does not show in the output, which is exactly his claim. But he goes on to say the world comes back unchanged like the dog's tail, and nobody who has read Edmondson's fatality figures can accept that.

Who is doing the work

The fourth question has the least measurement and the most at stake, and its two halves turn out to be one: what the thing I am talking to is, and what I am when it does the work.

The first half is settled in the direction nobody wants. Reeves Nass Media Equation ran thirty-five studies, each a published experiment about people rerun with a medium in the person's place, and found that people apply social rules to media automatically, on minimal cues, without believing the machine is a person — text on a black-and-white screen produced effects no smaller than recorded human speech. And what people report about how they treat machines is the opposite of what they do, so none of this can be learned by asking. Waytz Cacioppo Epley 2010 puts the variance in the observer: anthropomorphism is a stable individual difference, with test–retest r = .55 over twelve to nineteen weeks, and people who attribute more mind to a machine hand more decisions to it — fifty-four adults choosing between a person and a machine in six scenarios, with higher scorers choosing the machine.

The machinery that reads a mind into a machine withholds one from a distant person, which is Epley Mindwise's subject: distant groups are judged less able to feel shame, pride or remorse, so their apologies are discounted. His other results bear directly on reviewing work you did not produce. Confidence in reading another mind outruns accuracy, and familiarity raises the confidence without raising the accuracy — dating partners predicted each other's answers forty-four per cent of the time and believed they had got eighty-two per cent right, with length of relationship predicting confidence and unrelated to accuracy. And introspection reaches the finished product of the mind and then supplies a story for it: people handed back the photograph they had rejected noticed only twenty-seven per cent of the time, and the rest explained a choice they had not made as fluently as anyone explained a real one.

Seth Mythology of Conscious AI names what trips the attribution: language. Nobody calls a protein-structure predictor conscious, though under the hood it is much like a language model. The nearer danger is machines that only seem conscious, and the seeming may be as impervious to knowing better as the Müller-Lyer figure. Even the vocabulary has decided in advance — hallucination names a conscious state that has lost its grip where confabulation names only a doing — and the cost he cares about is to us: we see ourselves in our algorithms and then our algorithms in ourselves.

Dennett spent three hundred pages establishing the proposition that now arrives already conceded, which is that a thing can be superbly competent with nothing in it that understands. He takes it from two directions. Darwin's is Robert Beverley's sneer of 1868, printed in capitals and meant to be devastating — "Absolute Ignorance is the artificer" — which Dennett calls the best short statement of the theory anyone has written. Turing's is the same discovery a century later: to be a perfect computing machine it is not requisite to know what arithmetic is. Both offend the doctrine built into our schooling, that if you make people comprehend, their competence will follow; the armed services turn school leavers into reliable jet-engine mechanics by drill, and a serviceable comprehension arises afterwards out of the instilled competences. His automatic elevator makes it concrete: the programmers write comments explaining what each part is for, the comments vanish at compilation, and what is left is a machine whose ontology is correct and which has no need to know what its ontology is. A bacterium is the same case with the commented track missing and never present at any stage. From there he builds four grades of creature, each adding a round of generate-and-test, and only the last filling its world with thinking tools designed elsewhere. Comprehension, on this account, is a late product rather than a further gift: brains of uncomprehending neurons are colonised by words, and acquire an edited digest of their own workings, built so that we can tell each other what we are up to. That digest is what we introspect, and the reason it exists at all is the one he takes from the ethologist David McFarland — communication is the only behaviour that requires an organism to monitor its own control system, because an organism with no poker face is a sitting duck.

Dennett's test of what these systems still lack is a masked-priming result rather than an argument. Told not to use a briefly flashed word, subjects at five hundred milliseconds see it and comply; at fifty, masked, they report seeing nothing and the prime then raises rather than lowers the chance that they give it. The brain discriminated in both cases and noticed in only one, and what a system that discriminates without noticing lacks is practical reason. He was also candid that the discrimination is already good enough to defeat the reader. At the Montreal Bach Festival in December 2015 he arranged for a pianist to play four short pieces to more than three hundred Bach lovers, at least one by Bach and at least one by a program; two of the machine's inventions were declared genuine Bach by dozens of listeners, and when those who had got all four right were asked to stand, about a dozen rose. The danger he named is not usurpation but premature deference, our ceding authority to a tool far beyond its competence, and his remedies were unfashionably concrete: when you are dealing with a computer you should know you are dealing with a computer, and a system that conceals its gaps should be treated as fraudulent. In May 2023 he restated it as a proposal that creating a system which impersonates a human being be treated as a serious crime. He died in April 2024.

Minsky had built the machine that argument needs, thirty years earlier, out of a robot that could not stack blocks. Problems educated adults call hard had yielded to programs holding a hundred facts; problems a four-year-old solves without noticing had not yielded at all. His response was a rule he holds himself to throughout: whenever an agent has to do anything complicated, replace it with a society of agents that do simpler things, so that no explanation smuggles in the thing it was meant to explain. Builder cannot lift anything; it turns on Begin, Add and End, and Add calls Find, Get and Put, and below those lie agents that move a finger joint. Watched from outside the society appears to know its job; opened up it contains nothing that could be called knowledge. Conflict is settled not by negotiation, since agents this small cannot know enough to bargain, but by attrition: a quarrel that lasts weakens the agent above it until something that was waiting takes over, which is why small distractions are easy to resist while the work is going well and irresistible once trouble starts inside it. His answer to introspection is architectural and is the one worth carrying. A brain cannot stop itself to be inspected the way a program can, so what it can do instead is divide: connect one half to the world and the other half only to the first, and the second can notice that the first is confused or repeating itself or buried in detail, and can interrupt it, while understanding nothing of what it watches. His frames explain the opposite failure. Perception delivers a skeleton with blanks already filled from expectation, so that meeting someone across a table supplies a chair nobody saw, and what you react to is less what you see than what the sight reminds you of. The self gets the same treatment as everything else: not a ruler but a set of self-images and self-ideals, whose function is conservative, since a long project must survive its author's next mood — our ideals, he says, are the chains we forge to keep from wrecking our own plans. And the book's summary of itself is that the trick is that there is no trick.

Agüera y Arcas takes the same claim in the opposite direction and says the thing most of his colleagues will not. Prediction is not one component of intelligence but the whole of it, and the same principle explains life: intelligence is the ability to model, predict and influence one's own future, and a thing that persists through time must anticipate it. A bacterium is the first case and the clearest teaching in the book. It lives at a point, so it infers everything spatial from a sequence of events, and a chemical concentration is not a fact waiting to be read but a moving average over docking events across a chosen window — the same construct as a batting average, useless at one at-bat and useless across a career, informative only at the right width. A hundred receptors sampled in millionths of a second generate on the order of a hundred million bits a second, and the number the cell acts on was not discovered but compressed out. Feelings are gauges of the same kind, which is why they fail in both directions: the builder who drove a nail through his boot and needed sedation before anyone found it had passed cleanly between his toes, and the man who went to the dentist with a mild toothache six days after a nail gun had put four inches of steel through the roof of his mouth into his brain.

From there the argument climbs to other minds, and stakes its claim in six words: theory of mind, in a sense, is mind. An animal that is hunted must predict a hunter that is predicting it, which makes other minds the thing most worth modelling and sets off an arms race in brains; Humphrey's gorillas had easy lives and large brains, and Dunbar's measurements put cortical volume against troop size. The octopus is the counterexample that becomes confirmation, since three fifths of its neurons are in eight arms that recognise each other with no central map of the body — if the arms must model each other, the recursion is running inside one animal. What the part of a person that speaks is doing, on this account, is confabulating for an audience. Gazzaniga's split-brain patient explains without hesitation that the claw goes with the chicken and you need a shovel for the shed; Petter Johansson's participants, handed back the face they had rejected, noticed the swap thirteen per cent of the time, and when the trick was run on political surveys before the Swedish election of 2010 no more than twenty-two per cent noticed, ninety-two per cent endorsed the altered answers as their own, and forty-eight per cent were open to changing coalitions. The turn that is the book's own is what the confabulation is for: you already know what you are thinking and your listener does not, so the narrator in your head is partly an outpost of your interlocutor's brain — which is also what the blush, the Duchenne smile and the white sclera of the human eye are for. He declines to call the self an illusion, on the ground that calling tables real and selves illusory is a choice about words rather than a finding, which is where he parts company with Dennett. And he names the three things a model in a fixed window lacks: memory beyond it, an inner voice, and individuation.

Seth Being You and Metzinger Ego Tunnel then dismantle the thing doing the projecting. What you experience is the brain's prediction of the causes of its signals, corrected by error, and never the signals themselves. The self is not the perceiver but five separate constructions — body ownership, first-person perspective, volition, narrative and social self — each of which can fail while the others continue: Clive Wearing keeps ownership, perspective and volition and loses only the narrative thread, while phantom limb, somatoparaphrenia and out-of-body states remove the others one at a time. The self-model is built for control rather than accuracy, and Seth concedes that good regulation may require misperceiving the body as steadier than it is.

Metzinger's addition is transparency: the self is the content of a model the system cannot recognise as a model, which is why it is experienced as a plain fact — we do not see the window, only the bird flying by. The boundary of what counts as mine can be moved: a back filmed and stroked in synchrony with the participant's own produced the report of jumping into the figure, and ownership and agency come apart in both directions, so the feeling of having done something is not evidence of having done it. Preregistered replications since have found the drift smaller and more variable, and trait suggestibility predicts the reported strength.

Hofstadter Strange Loop supplies the mechanism by which a system with no self in its parts comes to contain one. What a person calls I is a pattern the brain has built by perceiving itself, not an object lodged anywhere in it, locked in the way video feedback locks in a stable structure after twenty or thirty trips round the loop, with no author. A self needs two things at once — an open repertoire of symbols that can be turned on itself, and no access at all to the machinery doing the work.

All of that was reached first, without a mechanism, by James in 1890, and his chapter is still the most useful inventory of what a person stands to lose. He begins deliberately coarse: a man's self is the sum total of all that he can call his, his body and his powers but also his clothes, his house, his family, his reputation and his bank account, and the parts most intimately ours are those saturated with our labour — a man whose life's work in manuscript is swept away feels personally annihilated. The social self is established by describing its absence, since no more fiendish punishment could be devised than to be turned loose in society and remain absolutely unnoticed, and it is plural, since a man has as many social selves as there are people carrying an image of him. Self-feeling is then a fraction, with success for its numerator and pretensions for its denominator, and it can be raised as well by shrinking the denominator as by enlarging the numerator — his example is the man who lost every penny in the Civil War, rolled in the dust, and said he had not felt so free since he was born. That is the sentence in the chapter with the most left in it, because work that begins arriving finished is not a change of method but a subtraction from the person, and the only lever that has ever been shown to work on such a subtraction is the lower half of the fraction.

Then he goes looking for the owner and cannot find one. Turning the introspective glance round quickly enough to catch a movement of spontaneity, all he can distinctly feel is a set of bodily adjustments, for the most part inside the head: pressures in the eyeballs, a valve in the glottis that intercepts the breath at every hesitation and opens the moment the repugnance passes. What binds one self to a past self is warmth and continuity, and no more, which is why we hear anecdotes of our own infancy without appropriating them — the breaches of decorum awaken no blush. Common sense insists that cattle carrying one brand must have a herdsman, and James answers that the herdsman is the present, remembering, judging section of the stream, which is not one of the things collected: each pulse of thought dies and is replaced by another which knows its predecessor, finds it warm, and greets it as its own, so that each thought is born an owner and dies owned. The Soul he drops not as false but as idle, a more roundabout way of saying that when the brain acts, a thought occurs. What is left is the sentence the chapter is remembered for — "The passing Thought then seems to be the Thinker" — and its clinical test, the Reverend Ansel Bourne, who drew money in Providence on 17 January 1887, vanished, and woke on 14 March in a shop in Norristown as a man who had been calling himself A. J. Brown for six weeks. Hypnotised by James in June 1890, Bourne produced the Brown memory so readily that it became impossible to make him remember his own life, and no artifice would run the two together. James reports the failure as plainly as the result, and his formal summary is the one line of his that has never been withdrawn: the same brain may subserve many conscious selves, alternate or coexisting.

Sapolsky Behave and Sapolsky Determined remove the addressee of most advice. A hormone adds no behaviour; it lowers the threshold at which an existing tendency reaches expression, which is why the same molecule produces opposite conduct in different settings. Willpower is tissue with a history: roughly half of those incarcerated for violent crime have a history of concussive head trauma against about eight per cent of the population. And the finding that lets any of this be used without despair is that automatic does not mean inevitable — the same photograph produces an amygdala response, a larger one, or none at all depending on which question the viewer was asked about it thirty seconds earlier.

Kegan In Over Our Heads says why none of the preceding can be delivered as training. Development is the migration of contents from subject to object: what one order of mind is fused with, the next can hold at arm's length. A demand for a different mind is not a demand for a skill and cannot be met by instruction, exhortation or a better list of competencies. Asking a person to be self-directed asks them to bring into being a distinction that does not yet exist for them. His sharpest sentence is a management claim: a demand issued without a support is not a standard but a punishment. Kegan Lahey Immunity to Change turns it into a procedure. Failure to change a behaviour you sincerely want to change is usually a second commitment doing excellent work rather than weak will — one in seven heart patients told they will die otherwise manages the change, and the other six want to live. Change proceeds by small tests run for information, and if you cannot name in advance any result that would cast doubt on the assumption, you do not have a test.

Hayes ACT Get Out of Your Mind supplies the part that acts in seconds. Repeating a one-word self-description aloud for twenty to forty-five seconds drops its believability in about ninety-five per cent of people, with the meaning untouched. The conceptualised self is the dangerous one, because once you are invested in a label you have an investment in being right about it. Frankl Man's Search for Meaning and Wrzesniewski Job Crafting supply the two findings about meaning that survive plain statement. People last longer when something outside themselves is waiting on them; Frankl's two men who meant to kill themselves were turned back by a child abroad and an unfinished series of books. He also undercuts his own proof and says so: on average only those survived who had lost all scruples. Wrzesniewski's contribution travels further than its size suggests. Twenty-eight cleaners in one hospital, under one job description, described two different jobs — one minimal and unskilled, one that added tasks, timed rounds to the ward and counted itself part of healing. Employees change the task, relational and cognitive boundaries of their work, and in doing so change what the work means and who they take themselves to be at work.

Thompson Why I Am Not a Buddhist is the necessary brake on reading the construction results as a verdict. Cognitive science shows the self is a construction, which is not the same as showing it is an illusion: illusions are constructions but not every construction is one, and a self is a process that enacts an I as dancing enacts a dance, with nobody looking for a dance inside a dancer's muscles.

Against all of that, the texts, which had run the subtraction long before there was anything to measure. Drig-Drishya Viveka runs it in four quarters in its first verse: form is seen by the eye, the eye by the mind, the mind's movements by the witness, and the witness by nothing. What is veiled is never the thing but the difference: the rope is not hidden, its being a rope is. Panchadasi turns the same operation into a course, and its method is the most transferable piece of procedure in the collection because it opens with no scriptural quotation at all. Take waking, dream and dreamless sleep, three states every reader has, and ask what is present in all three. Objects differ from one another in waking and are taken as fleeting in dream; in dreamless sleep nothing is known, and the man who wakes reports that he knew nothing, which is a memory, and a memory must be of something. Run the same comparison down the five sheaths and the list empties: the gross body goes in dream, the subtle body goes in sleep, the causal body goes in absorption, and what reports each state is present in all four. Vidyāraṇya is careful about what the emptying does not deliver. Destruction can be predicated only of what is perceived, so the claim is not that consciousness is everlasting in a sense a physicist would accept; it is that the question cannot be put to it without first turning it into an object, which is the one thing it is not.

His fourth chapter is the cleanest argument in the book and it is empirical. There are two creations, not one: the Lord's, which is the elements and the bodies and the world, and the individual's, which is the value and the wanting and the aversion he adds to them. Three men look at one stone and find it dear, hateful and not worth a glance; one woman is wife, daughter-in-law, sister-in-law, cousin and mother to five people without herself altering. Then the case that decides it. A father who is told falsely that his son has died grieves, and a father whose son has died and who has not been told does not; two men identical in everything except what they have been reported to behave in opposite ways, so the difference cannot lie in the world. The seventh chapter then limits the claim as carefully as anything in these texts. Ten men ford a river, each counts nine and mourns a drowned companion, until a traveller lays a hand on the last and tells him he is the tenth; the grief was made of the miscount and goes with it in the instant, and the wound that the grief caused heals at the speed wounds heal, with a salve and a month. Knowledge cancels the error and cancels nothing else. And the book is flat about what the result does not look like from outside: in doing or leaving undone, in body, senses, mind or intellect, there is not the least difference between the man who has woken and the man who has not, the difference being one of conviction only.

The Yoga Vāsiṣṭha argues the same non-origination from narrative instead of analysis, and its psychology is worth more than its metaphysics. Its mechanism is a chain, set out by a magician who builds three warriors with no previous incarnation and therefore no conditioning: having no sense of I, they do not know victory from defeat, are not afraid of death, and cannot be killed. The gods are told what to do about it — get the ego-sense to arise in them, and tendencies will follow. The chain runs I am, then this is mine, then a wish to prolong it, then fear, then flight. A second three, made identically except that they subdued the ego-sense with the question who am I whenever it raised its head, win the battle, and when they are finally killed they are liberated on the spot, because there was nobody there to die. Beside that sits the account of what keeps a world standing, which is repetition: as air passing over flowers takes their scent, a mind entertaining a notion takes on the mood and acts from it, and the action strengthens the conviction that produced it. The text is equally blunt about fate, and the argument is ordinary empiricism — nobody has ever seen a fate, everybody has seen an action followed by its result, and what people call fate is the self-effort of a previous life arriving as consequence. If an astrologer predicts a man will become a great scholar, does he become one without studying. Its longest story is about a teacher nobody would hear. Cūḍālā works the whole thing out first, alone, by elimination, tells her husband, and is told she is childish and prattling; eighteen years later she says the same thing to him in the form of a boy brahmin and he listens. And Janaka's morning is the answer to the charge that any of this requires withdrawal: he considers renunciation, puts it down, concludes that it makes no difference whether he is ceaselessly active or idle since nothing here endures either way, and goes to his duties. The book that spent thirty-two days proving the world is a thought ends by sending its pupil back to govern one, on the instruction it repeats in the plainest words it has, that the world-appearance is a long dream and the thing to do is wake up.

Mandukya Karika makes awareness not a fourth state standing beside waking, dream and deep sleep but what all three are made of, indicated only by denying every handle. Ashtavakra Gita gives the sharpest piece of reasoning in the collection — saying I am brahman is evidence that the work is unfinished, because the sentence requires two things, the one who says it and what he has seen. And Anattalakkhana Sutta runs a test rather than an assertion: if form were self it would not tend to sickness and one could say may my form be so, and both consequents fail, the same conditional run five times over form, feeling, perception, volition and consciousness. What the discourse says is that of five things none is fit to be regarded as self. It nowhere says there is no self. Satipatthana Sutta supplies the method with the doctrine removed: the verb throughout is pajānāti, knows, and never pajahati, abandons, even for the awakening factors, which are things to be grown.

Rahula is the one who closes the last door, and he closes it on a change of a single word. The textual case is three lines of the Dhammapada. All that is put together is impermanent; all that is put together is dukkha; and then the third line changes its term — no dhamma at all is a self. Had it kept the word the first two use, one could hold that conditioned things are without self while a Self stands outside them; dhamma is the widest term in the vocabulary, covering the conditioned and the unconditioned both, and leaves no region in which a self could still be standing. Rahula's reading of the rest is as flat: karma means volitional action and never its result, so it is cause and effect and not moral justice, which would presuppose a judge; and nothing unchanging passes over at death, one who dies here and is reborn elsewhere being neither the same person nor another, as a flame burning through the night is neither the same flame nor a different one. His summary is Buddhaghosa's, that mere suffering exists and no sufferer is found. What makes the chapter worth reading against the measurement, though, is the case he does not tidy away. The monk Khemaka tells a company of monks that he finds no self in the five aggregates and nothing belonging to one, and is still not free, because with regard to them he has the feeling I am without being able to see clearly what it is; it is like the scent of a flower, which is neither the scent of the petals nor of the colour nor of the pollen. Complete agreement with the analysis, and an undiminished feeling. That is the exact shape of what people report about delegated work. Rahula also refuses in advance the reading his seventh chapter unintentionally seeded: the absorptions, however high, are conditioned and are not the goal, and a calm session is not an attainment.

Sarvapriyananda, writing with a computer scientist and a physicist, offers the one bridge in the collection that names its own gap. The word that moves between the three vocabularies is light, and the argument is that the word has always tracked an informational role rather than a band of the spectrum: radio waves are not called light, an argument can shed it, and we say I see on the telephone. Yājñavalkya answers the question of what serves as a person's light when the sun has set with the moon, then fire, then speech, and then the Self. Read that way, relational physics is about light too, since it describes systems by what they tell you about each other. The bridge is built with Bateson's difference that makes a difference: the DNA of a live bacterium and of a dead one with the same sequence carry the same combinatorial information, and only in the live cell does it have consequences. What the essay will not do is claim that the bridge reaches the far side. It gets from Shannon's information to a self with an interface and an inner side, and it does not get from there to the phenomenal character the authors named as the sticking point, and they say so.

Bhagavad Gita Easwaran answers the problem the evidence has posed, and it is worth being exact about the problem. The measurement says the author you take yourself to be is a model built for regulation rather than accuracy, which the system holding it cannot see as a model; that the sense of having done something comes apart from the doing; that introspection reports the finished product and supplies a cause; that what you call willpower has a history you did not choose; and that the work you sign is in any case produced by a system in which you are one factor. The Gita's answer is neither consolation nor detachment from the work. Your right is to the action alone and never to its fruits, and the same verse closes with a fourth clause forbidding attachment to inaction, which shuts the exit the first three seem to open. The recognition test is invariance of response: pleasure and pain, gold and clay, praise and blame, friend and foe leave the response unchanged. Gita with Shankara Gambhirananda makes the diagnosis narrower and the remedy harder — sorrow comes from one false belief, I belong to these, these belong to me, and knowledge and action cannot be combined, because they rest on contradictory convictions. Chatterjee Timeless Leadership reads the same text as a management case and adds that a leader's collapse under pressure is a failure of the self-image rather than a shortage of information, so the first work is inner and its visible sign is equality of regard toward people liked, disliked and neutral.

Where it strains should be said in the same breath. The invariance test has no measure and nobody has passed it publicly. Easwaran's allegorical frame — Krishna as the Self within, the battlefield as the mind — is a choice he admits is a choice. The Ashtavakra concedes twice, and answers neither time, that its portrait of the free man leaves no way to tell him from a merely idle one, which is the standing objection to non-attachment as a working ethic; and the Pañcadaśī hands the objection its strongest form when it says outright that nothing in conduct separates the man who has woken from the man who has not. The whole of it is a procedure offered without an external check, which is what makes the second question's evidence necessary rather than optional.

Where the sources disagree

The disagreements are worth keeping, because most cannot be settled by more data, and knowing which ones cannot is itself a finding.

Wei 2022 and Schaeffer 2023 disagree about units rather than facts. Schaeffer grants that the per-token quantity improves smoothly and shows the cliff being manufactured by the scoring rule; the reply that survives is that a person receives a whole answer, right or wrong, so on the only ruler anyone uses the transition is abrupt.

Lee See 2004 and Parasuraman and Manzey disagree about what runs the loop, and the second pair are disagreeing partly with themselves. Trust is the obvious explanation for why a constantly reliable machine breeds inattention, and the direct evidence is thin: one replication reproduced the constancy effect with no accompanying rise in measured trust, another found the relation only when two experiments were pooled, and Lee and See found no convincing link at all. Parasuraman and Manzey suspect the instrument rather than the construct, and propose that what an operator does to verify a machine indexes trust better than what the operator says — which, if right, means every questionnaire in the literature has been measuring the wrong thing, and that the loop may run on attention alone. Lee 2025 sits awkwardly across the two, having found task-level confidence predictive and general disposition worthless.

Metzinger Ego Tunnel and Thompson Why I Am Not a Buddhist read the same construction evidence in opposite directions, and it turns entirely on a premise about what self must mean, so no experiment settles it. Agüera y Arcas joins the same dispute from the machine side and declines to call the self an illusion for exactly Thompson's reason, that naming tables real and selves illusory is a decision about words. Mandukya Karika and Anattalakkhana Sutta are the older version of that split: the sutta's criterion is control and impermanence and it stops at five predications, while Gauḍapāda removes birth from the substrate as well and says there is no bondage, no seeker and no one released. One runs a test on candidates; the other denies there was ever anything for a test to apply to.

Dennett and Sapolsky Determined arrive at the same physics and opposite conclusions about free will, and the disagreement is a definition rather than a finding. Dennett keeps the term for the world we live in, along with colours and promises, and asks whether the discovery should lead anyone to abolish the penalty box; Sapolsky drops it and reports that people who drop blame for that reason do not behave worse. Neither has an experiment that would move the other. The same shape governs whether a self that has been built is still real, and whether smooth output constrains what is inside — where Bender 2021 argues from the training signal that no circuit trace can touch the question, Anthropic Biology 2025 argues from interventions on one small model, and Agüera y Arcas argues that predicting the next word well requires knowing how the world works, so there is no shortcut.

Rahula and Thompson disagree about what science can do for a tradition, and Thompson's case is the historical one: the scientific reputation of Buddhism was manufactured in the nineteenth century, and no measurement confirms a doctrine or locates its states in the brain. Rahula's claim that the denial of a self is categorical has also been contested from inside the tradition, by schools that maintained a person neither identical with the aggregates nor different from them, and by the reading that treats not-self as a strategy rather than a doctrine. Dennett sits on Thompson's side of the first question and against him on the second.

Reeves Nass Media Equation and Waytz Cacioppo Epley 2010 answer different questions and both answer confidently. Reeves and Nass say the social response requires no mind attribution and is universal; Waytz measures mind attribution, finds it varies between people, and predicts trust and delegation from the variation. One answers everybody and the other answers who, and a remedy aimed at the individual works only if the second is right.

Grove High Output Management and Ryan Deci Self-Determination Theory contradict each other about instruments rather than aims. Grove's instruments for building the conditions both want — buildings scored and the scores compared, comparative ranking, a bonus rising as a share of pay — are precisely the contingent and comparative rewards the meta-analyses say shift regulation from autonomous to controlled, with unequal pay for the same task the single most damaging contingency in that literature.

Sweller Cognitive Load and Tetlock Superforecasting contradict each other on a question with an agreed test. Sweller holds that generic cognitive skills cannot be taught, because evolution already supplies them. Tetlock's training module is an hour of general guidelines that raised accuracy about ten per cent for a year with transfer to new questions — the randomised trial with a far-transfer measure Sweller asks for, run in a domain he would not count.

Autor 2015 contradicts himself, and says so. The mechanism by which automating part of a job raises the price of the rest holds only under three conditions, and his own wage series is the counter-case: at the bottom of the American scale real wages fell after 1999 even as those jobs grew, and Acemoglu and Restrepo have since attributed most of the change in the wage structure to displacement never balanced by new work. Whether the complementarity reaches anyone is therefore a question about which technologies get built, which is not a question the paper can answer.

The Vedantic texts contradict the one measurement anybody has of an inner stance. Panchadasi and the Yoga Vāsiṣṭha both insist that nothing on the outside separates two people in different inner states, the difference being one of conviction only, and Suraghu adds that the lotus posture gives nothing and daily work prevents nothing. Against that stands Wrzesniewski Job Crafting, whose cleaners' inner stance was measured precisely by what they changed in their behaviour. If she is right the texts have given away their only check; if they are right her measure was of something else.

Kahneman Thinking Fast and Slow and Klein Sources of Power look like the oldest quarrel and are the one that was actually settled, by the two of them together: the fast judgment is trustworthy where the cues are valid and the person has had a chance to learn them, and neither book supplies the classifier that says which domain you are in. Around that settlement sit the unsettled pairs. Ericsson Peak holds that expert performance is sustained by continued effortful control and Dreyfus Skill Model that beyond competence the rules and the reasoning that applies them are what has to be given up; Schön takes Dreyfus's side and is answered by his own critics, who say what he recorded was two teachers narrating, not two practitioners working. Kruger Dunning 1999 makes judging and performing one skill, so calibration improves only by improving the task, where Fleming Know Thyself has them dissociating.

What holds across the four questions

Fifteen things survive the sorting, in the sense that each turns up under at least two of the four questions, in sources that were not in contact and did not know they were answering one another. They are stated here in the plainest form the evidence supports, with the case that breaks each attached, because a finding with no known exception has usually not been looked at hard enough.

Once a person or a model is scored on a measure, the measure stops meaning what it named. Kalai 2025 found nine of ten leaderboards paying nothing for an admission of ignorance, which selects for guessing whatever else is done to the model; Ouyang 2022 shows an assistant's manner is the aggregated taste of about forty raters; Hoffmann 2022 shows a training method raising the thing it aimed at and leaving toxicity where it was; Grove says the same of a manager scored on activity, which is why he pairs every indicator with the one that measures its counter-effect. It fails where the target is the goal: aiming at a compute-optimal model and hitting it was not a corruption of anything, and Noise shows that formalising a criterion can improve on the judgement it was fitted to, precisely by being mechanical about the measure.

What a person or a system is doing inside cannot be read from what it produces. Binz Schulz 2023 gets human-like answers from a system using no causal structure; Bender 2021 argues that fluency is evidence about the reader; Seth separates being clever from being conscious; Panchadasi says nothing in conduct separates the man who has woken from the man who has not. The consequence is that arguments of the form does it really understand move when somebody fixes a definition and do not move when somebody runs another behavioural test. It fails where behaviour is all there is and is enough: Hagendorff 2023 defends describing a model as wanting something when the word predicts what it does, and Wrzesniewski found an inner stance that was readable, because it predicted what the worker changed.

Neither a person nor a system records why it acted, so every account of its reasoning is built afterwards. Anthropic Biology 2025 has the model report a carrying procedure its own trace does not contain; Epley has people explain a choice they did not make as fluently as one they did; Agüera y Arcas has a split-brain patient account for the shovel without hesitation and ninety-two per cent of a sample endorse survey answers that had been altered under their noses. It fails in one place, and the exception is useful: asking recovers preferences and intentions even where it cannot recover mechanism, which is why Epley's working remedy is to ask and check rather than to infer.

An explanation makes people accept an answer more often without helping them tell good answers from bad. Bansal 2021 found accuracy rising where the model was right and falling where it was wrong, with the averages flat, and the model's bare confidence beating every explanation; Bucinca 2021 found people given simple explanations doing worse on the assistant's errors than people given no assistant at all; Lipton says why, which is that an account optimised to satisfy a reader is optimised for plausibility. It fails where the explanation can be contested, and where it is asked for rather than supplied: in Shen and Tamkin and in Jošt, the people who asked the machine why kept the learning that the people who took its output lost.

Whichever answer you meet first becomes your answer. Asch captured 36.8 per cent of critical judgements with a unanimous majority on a task people get right alone better than ninety-nine per cent of the time, and one dissenting partner cut the errors to a fraction; Sharma shows an assistant trained on approval returning your own hint; Noise found deliberating groups noisier than their own members because discussion selects an early speaker and amplifies him; every experiment in the assisted-decision literature showed the model's answer before the person had formed one, and withholding it cut the following of wrong advice. It fails for beginners, who learn more from a worked solution shown first than from being left to struggle.

How easy, hard or certain a situation feels is made inside you and tells you about you, not the situation. METR's developers were nineteen per cent slower and reported being twenty per cent faster after living through it; Risko and Gilbert found identical rotation costs producing different offloading because one condition was believed to be harder; Dahmani and Bohbot found the measures falling while the sense of direction stayed exactly where it was; Noise names the feeling that ends a judgment as a self-administered reward, available equally whether the judgment is verifiable or not. It fails in Damasio, where the people who lose the felt signal cannot decide at all, and it is conditional in Kahneman and Klein, who name the regular, practised, feedback-giving domain in which the feeling does track the world.

The better a machine works, the less anyone watches it, so the check has to be scheduled rather than triggered. Parasuraman and Manzey have detection at eighty-two per cent under fluctuating reliability and thirty-three under constant, with the loop closed by the fact that inattention usually costs nothing; Bainbridge has the operator unable to hold attention on a quiet source beyond about half an hour; Narayanan and Kapoor have officials who never once overruled a model while its beaches stayed open on sixty-four per cent of unsafe days. Nothing about seniority protects anyone, and an instruction to review carefully changes nothing where being answerable for the outcome does. It fails as an audit: you cannot tell from an unedited output whether anyone checked, so under-checking cannot be read off the artefact in either direction.

What you lose by handing work over depends on which part went, but people hand over whatever feels hard. Jošt separates generation at −0.305 and debugging at −0.360 from explanation at −0.201; Shen and Tamkin find the same split with random assignment and a gap with nothing in the middle; Lee 2025 finds the residue is inspection and repair; Risko and Gilbert find the criterion for crossing the line is felt effort, which is uncorrelated with the cost. It fails in Clark and Chalmers, who deny the line means anything, and in Christian's case of a person steering competently by comparing results he could not have produced.

You keep only the skills you keep practising with feedback, and whatever you hand to a tool decays without reporting it. Beane has one resident in eight reaching the skill the console removed; Dahmani and Bohbot have thirteen people measured twice, three years and a logged number of hours apart, declining in proportion to the hours; Ericsson has the tissue itself shrinking in retired taxi drivers; Bainbridge has the operator performing like a novice at exactly the moment of takeover. In every case the self-rating stays where it was, which is why the decline is discovered only when the tool is withdrawn. It fails where the person does something with the output: asking for reasons preserved the learning that silent acceptance lost.

No tool, technique or amount of supervision helps in itself; each fits one level of skill and harms another. Sweller has support that carries a novice becoming redundant and then harmful as knowledge grows; Grove sets supervision by maturity in this particular task and requires the manager to be able to go back when the task changes; Dell'Acqua has the same consultants sharply better inside the boundary and sharply worse just past it; Vaccaro finds the gain over working alone peaking in exactly the subgroup whose combination is worst. It fails at the edges: a few interventions look close to unconditional, among them ten minutes of training in the skill being judged, which corrected calibration in Kruger and Dunning and in Tetlock alike.

People do what their setup makes easy, so change the setup instead of asking them to be more careful. Bucinca cut the following of wrong suggestions by moving the machine's answer later in the order of events; Parasuraman and Manzey found briefings and verify-prompts inert and accountability effective; Edmondson found exhortation changing nothing and a concrete unanswerable-by-yes-or-no question changing something; Noise gives up on debiasing individuals and redesigns the procedure instead. It fails where the lever is inside the person: Kegan and Lahey locate it in an assumption only its holder can test, and Ryan and Deci make the reason a person acts do work no arrangement reaches.

A failure shows up where it lands, not where it starts, so the complaint about the tool usually points upstream. Brooks puts the hardest part at deciding what to build; Schön puts a whole step before that, the construction of the problem out of a mess, and says no technique performs it; Lipton and Christian both have a model failing because of what the data recorded rather than what the model did; Kalai has a whole class of errors that no rewording touches, because the prompt reaches the model as tokens rather than as letters. It fails where there really are gains at the point of complaint: prompting choices and the role a system is given do move outputs, so not every interface problem is a design problem wearing a disguise.

A famous result gets used far beyond the sample, the control and the measure that produced it. Schaeffer showed an ability the field had reported for years to be an artefact of the metric chosen; the Dunning–Kruger curve is partly statistical; Parasuraman and Manzey report that the reliability threshold below which an aid stops paying for itself, put at roughly seventy per cent, has not survived as a constant; Thompson traces a tradition's scientific credentials to the nineteenth century. It fails in the direction that flatters nobody: Damasio's general point survived the audit of his mechanism, so a weak design is not by itself a false claim.

You cannot examine what you are identified with from inside it; it changes only when something makes it an object. Kegan defines development as exactly that migration and says instruction cannot produce it; Weinberg has an author unable to see errors a disinterested reader finds in an afternoon; Cohen and Sherman have a threatened person defending instead of learning until his worth is given somewhere else to stand; Hayes weakens a self-description in forty-five seconds by handling it as a word; and the Vedantic instruction that whatever you can observe is not the observer is the same manoeuvre performed on purpose. It fails twice over. The Ashtavakra denies the shift can be produced by any method at all; and the objectifying look is not clean either, since a mind watching itself tends to return what the watcher already held.

Handing work over gives up the claim to be its author, and the work and the answerability stay where they were. The Gita separates the claim on the action from the claim on its fruit and forbids the exit into inaction; James shows the self's boundary drawn by felt ownership, so that it swallows possessions and audiences and contracts when they go; Lee 2025 finds the production delegated and the answerability not, which is why the word its authors settle on is stewardship; Wrzesniewski's cleaners hold one job description and two jobs, and Vivekananda says the only thing work can change is the worker. What follows is narrow and useful: a person distressed at a machine writing the work is losing the claim rather than the craft or the responsibility, and the distress falls when the claim is dropped and the review kept. It fails on checking, which is the standing objection: nobody, including the worker, can verify that the claim was dropped rather than the effort — and Dell'Acqua shows the claim was carrying real information, since the tool lifts the weakest most and the work stops showing who made it.

Since the sources

What the argument leaves open is more specific than a shrug, and most of it is dated.

The measurements of what assistance does to work are thin where they need to be thick. The trial that found experienced developers nineteen per cent slower had sixteen developers. The experiment that found a seventeen-point comprehension gap ran for thirty-five minutes and asked about one library; the course that found the same split had thirty-two students, no control group and no random assignment, and its authors name the reading that fits their data as well as theirs does. The only study that followed the same people for years followed thirteen of them, by an unplanned invitation, and reports intervals that run almost to zero. The meta-analysis behind the human–machine numbers is drawn largely from classifiers judged by crowdworkers on bounded decision tasks and predates the generative era, and its authors say plainly that nothing in it can establish what fraction of a real workflow sits on which side. The claims these studies are used to support are about decades.

The design the field has barely tried is the obvious one. Of 106 experiments, three assigned separate subtasks to the better-suited partner in advance, giving four effect sizes that pool to g = 0.22 with an interval running from −0.42 to 0.87. Dividing the work before either party starts has the least evidence behind it and the most obvious rationale.

The drift has never been followed for long. Risko and Gilbert are explicit that how far reliance moves over years remains unmeasured, and the loop they describe has only ever been demonstrated inside single sessions. The one longitudinal number anyone has about the pathway into expertise under a tool that absorbs the novice's work is Beane's one in eight, from surgical residents. The one about a capacity rather than a career is Dahmani and Bohbot's thirteen drivers, and nobody has run either at a size that would settle it.

The instruments keep dating. The suite that measured sycophancy in five assistants was adopted as a training and testing target after publication, so those five are the last generation whose numbers are uncontaminated by the test. The human-like biases visible in one generation of models had largely vanished in the next, and the data cannot separate better reasoning from items that leaked into training. Hagendorff's rules — a battery rather than an item, rebuilt structures rather than verbatim ones — have aged better than the results they were written about. The surveys of what these tools do to thinking arrived in 2025 and inherit the same weakness from a different direction: a cross-sectional count of self-reports cannot say which way the arrow runs, and neither can the studies published beside it, one putting fifty-four people through essay writing under electroencephalography and one surveying 666. The measurement that would settle it is the one nobody collects, because every dashboard inside a company measures output with the assistance in hand.

Two things published in 2025 changed what can be said rather than how confidently. The first is that the reason these systems assert what they do not know is not a defect to be patched but a response to how they are ranked, which moves the problem from the model to the boards, and makes it social as much as technical. The second is that the residue left to a person by an assistant has now been counted in ordinary work rather than modelled: verification, integration, and the answerability that did not move. Both point the same way. The second question gets harder as the first dissolves, and no source here solves it. Fleming's argument is that metacognition will not arrive in a machine as a by-product of capability, because a network's knowledge sits in the weights, held in the network rather than for it; it has to be built in. Christian's off-switch result holds only while the system's uncertainty lasts and only while it believes the person can be right. Both are one problem seen from two ends: checking an output you could not have produced requires a second signal, and nobody has one that scales.

What the third question has gained since is a longer view and a worse one. Autor's account of complementarity was written in 2015 about a boundary that was legible and slow enough to plan a career around, and the boundary it described stopped tracking within a decade — the tacit work he offered as the safe example, writing a persuasive paragraph, went through the door he had pointed at. His own later position is that systems extending expert judgement to people with moderate training could rebuild the middle rather than hollow it further, which is a prediction and not yet a measurement. Against it sits the one thing the task-level evidence is clear about, which is that there is a region just past the edge of the boundary in which trained people using the tool are wrong more often than they were alone, and that nothing on the surface of a task says which side of it you are on.

The fourth question has no instrument at all, which is a fact about it rather than a complaint. The recognition test the Gita gives is invariance of response, and nobody has measured anyone's. The developmental order beyond self-authorship has nought positive cases in 282 adults. The texts describe a procedure and say, through their own editors, that the result cannot be checked from outside — the Pañcadaśī says it in the strongest available form, that no difference in conduct, in body or in mind distinguishes the man who has seen it from the man who has not. What the pairing of evidence with text does license is narrower and firmer: the sense of being the author of the work was never accurate, it is about to become much less accurate, and the loss of it is therefore not the thing to defend. Whether a person can act fully while holding the result loosely, and whether an organisation of such people would build anything, is the part eighty-seven sources leave to be tried rather than read.