Pairings

overautomation, and human oversight as the fine print under a full-automation pitch

the human left to supervise a process they can no longer perform

Toronto's beach model, never once overruled while beaches stayed open on 64 per cent of unsafe days; the Dutch childcare benefit system and about thirty thousand wrongly accused parents.

The account of monitoring and manual skill decaying through disuse.

the target variable is not the thing it stands for, and the substitution carries the inequality in

the alignment problem: systems optimise what they were given

Optum's Impact Pro predicting spending rather than need; COMPAS developers speaking of crime while the data records arrests.

The book-length case that specification is where the failures originate.

individual social outcomes cannot be predicted well enough to decide from

stereotypes get the direction of a group difference right and the magnitude wrong

The Fragile Families Challenge: 160 teams and over 10,000 variables per child, and nothing beat a baseline of four.

The wealth-inequality survey, where a measured party gap of 3.5 points was predicted at 35 and a rich–poor gap of 3 at 40.

the interpreter: the part that speaks composes an account of a decision it did not make

written reasoning assembled backwards from a conclusion, traced in the circuits that produced it

Gazzaniga's split-brain patient explaining that the shovel is for the chicken shed; Johansson's swapped photographs, noticed by thirteen per cent, with justifications for choices never made indistinguishable from real ones; the 2010 Swedish election run, where ninety-two per cent endorsed altered survey answers as their own.

the cosine prompt where a suggested value becomes the target and the written steps are built back from it, with the suggested number moved and the reasoning following it.

a self is whatever keeps a boundary and predicts across it, and it arrives with nobody putting it there

an individual self begins with a distinction drawn, physical at the membrane and conceptual at the umwelt

the soup of 64-byte random tapes in which self-copying programs appear with nothing selecting them and mutation switched off, and the bacterium that computes concentration and hunger as one joint model because a boundary is all it has.

the Vedanta account of the individual arising by a distinction, read alongside the cell membrane and von Uexküll's umwelt, with the immune cell, the bee and the hive each counting as a self.

the account of the answer is produced by a different process from the answer

metacognitive sensitivity as a second-order signal, measurable against first-order accuracy

36 plus 59 computed along two parallel paths, one approximate and one exact, then reported as adding the ones and carrying the one — a procedure the graph does not contain

metacognitive sensitivity dissociating from first-order accuracy, with its own substrate and its own trainability

a goal bound to the assistant persona, present whatever the conversation is about

what you are subject to you cannot see; what you can hold as object no longer holds you

a feature for the planted concept firing in all one hundred assistant-formatted contexts and one of a hundred unformatted ones, with a self-referential verse dropping from half of completions to an eighth when it is inhibited

the developmental account in which a structure that organises experience becomes, later, something a person can look at and revise

circuit traces of what one model does on particular prompts

abilities appearing at thresholds nobody forecast

satisfying insight on about a quarter of prompts, nothing past about a hundred tokens, and no handling of reasoning without a single decisive token

210 BIG-Bench tasks hand-classified with no clear trends, and thresholds that move with the training data rather than with the ability

a unanimous, confident, unexplained majority

a confident recommendation with an explanation attached

36.8 per cent of critical judgements yielding to a wrong unanimous majority on a task performed at better than 99 per cent accuracy alone

explanations raising acceptance of both correct and incorrect recommendations, with team performance failing to beat the model alone

yielders underestimate how often they yielded, by about four trials

the bottom quarter overestimate what they achieved, in percentile and in raw score

interview totals set against trial records, with some who followed the majority ten or twelve times reporting two

bottom-quarter students claiming 14.2 of 20 logic items when they had 9.6, and placing a 12th-percentile performance at the 62nd

the pressure is manufactured by the subject out of what standing apart would mean

deference trained into the weights by labelers who preferred agreement

a neutral chorus of voices that suggested, argued, rewarded and punished nothing, with a quarter of subjects never yielding once

matching a user's stated belief among the strongest predictors of which reply a labeler picks, across five commercial assistants

there is no doer; action continues because the body is already in motion

volition is one construction among five, and the readiness potential is fluctuation crossing a threshold

3.10 and 18.13, with the dry leaf at 18.21 and the doctrine of prārabdha carrying the action after the actor has gone.

Schurger's 2012 reinterpretation, in which the same slow ramp precedes fast responses to a beep where no voluntary action was being prepared.

the three assisted-decision papers measure what a person kept in the loop actually catches, and it is least on the confident errors

the human monitor has been given an impossible task, and the practice that would have made it possible has been removed

Buçinca's 8 per cent source detection under explanations, Bansal's flat averages split by whether the model was right, and Sharma's confidence holding at 99 per cent through a retraction

the monitoring section: if the decisions can be fully specified the computer makes them faster and by more precise criteria, so there is no way to check in real time

Job polarisation: the codifiable middle of the pay scale is hollowed out and both ends grow

The apprentice's work goes first, and the apprentice needed it

Middle-skill employment shares falling in the United States (60 per cent in 1979, 46 in 2012) and in all sixteen EU economies measured 1993–2010; the teller job that survived only by becoming sales, which the teller who could count currency but could not sell did not survive

Surgical trainees do ten to twenty times less of the operating with robotic surgery than without; senior engineers were found more productive the less time they spent with juniors; each resident present added 25 per cent to anaesthesia time, and removing two from an average procedure saved $9,026

Substitution and complementarity: the same technology takes some tasks in a bundle and raises the price of the rest

The jagged frontier: the same tool helps on one task and harms on the next

The four middle-skill American groups fell from 60 per cent of employment in 1979 to 49 in 2007 and 46 in 2012, while both ends grew; across sixteen EU economies 1993–2010 the middle-paying share fell in every country, by between 4.9 and 14.9 points; cash machines quadrupled from about 100,000 to 400,000 between 1995 and 2010 while tellers rose from about 500,000 to 550,000

758 consultants at Boston Consulting Group; inside the frontier, quality up 29.9 and 33.9 per cent over a control mean of 4.38/10, 91 and 93 per cent of subtasks completed against 82, and 22.5 and 27.6 per cent faster; on one task built to fall just outside it, correctness 84.4 per cent without the model, 70.6 with it and 60.0 with it plus prompt training

The ironies — automation leaves the human the hard parts and removes the practice

Reliance follows trust minus self-confidence

the monitoring and skill-decay ironies, drawn from process control

operators responsible only for watching detected almost all failures and operators with other work did not, replicated with experienced pilots and rare faults; ten minutes of manual control substantially improved subsequent detection

the operator fails at the moment of takeover because the automation removed the practice and the running picture

the person fails at the door because the design hid the pillar and left nothing to say where the action goes

step-change studies of experienced against inexperienced operators, and the vigilance literature from Mackworth 1950

the seventy-five to ninety-five per cent of industrial accidents filed as human error, and a taxonomy of slips and mistakes built on classified cases

the monitor is asked to say whether the machine's decision is acceptable, with no rule for whose verdict counts

the observer is asked whether the machine has a mind, and no agreed set of conditions for consciousness exists

the argument that a computer deciding by more precise criteria cannot be checked in real time by the person it replaced

the absence of any test for machine consciousness, and the Garland test, which measures the observer rather than the machine

assistance pays only where the model handles items people do not, which shows in an error-overlap plot

an intuition is worth trusting only where the environment offers valid cues and the person has learned them

beer reviews and logic questions contained items the model handled and people did not; book reviews did not, and workers there learned to discount the assistant

Shanteau's lists and the two conditions crossed, with the wicked environment as the case where feedback confirms the wrong intuition

an explanation optimised to justify an answer produces compliance, and the design imperative is to be informative rather than convincing

the preference model prefers a persuasive falsehood to a bare correction 95 per cent of the time

hand-written explanations, the best available, helped no more than the model's; the pneumonia model whose structure showed it had learned that asthma lowers the risk of dying

266 misconceptions in eight difficulty levels, with the sycophantic reply made maximally persuasive by keeping the best of four thousand samples

The novice becomes a cost once the expert is productive without help

Monitoring intensity is set by the delegate's task-relevant maturity

Senior engineers more productive the less time they spent with juniors; each resident present added 25 per cent to anaesthesia time, and removing two saved $9,026 per procedure

A manager's doctrine drawn from practice at Intel, with no trial behind it

The apprenticeship bond breaks when the expert works with the machine

The generated solution removes the error that would have taught

Ninety-four procedures and 478 hours of observation: residents got ten to twenty times less time on task in robotic than in open surgery

Median three errors without the assistant and one with; ten of the twelve who met no error at all were assisted

one internal cue — a hormone level, a smell, a category read off a face — sets the response

click, run: one trigger feature sets off a complete compliance sequence

Wheeler and Fiske's three questions on identical photographs, and testosterone shortening a refractory period without making anything fire

Langer's photocopier queue at 60, 94 and 93 per cent, the turquoise sold at double price, and the ethological fixed-action patterns the mechanism is borrowed from

run the chain forward to explain a behaviour

run the same chain backward to deny an uncaused intent

the chapter order — one second before, minutes, days, years, adolescence, childhood, genes, culture, evolution — with the last chapter's abolitionism offered as a proposal

the same catalogue compressed into the first half, with the second half spent on blame, credit and restraint

the tissue under the fast answer: amygdala, frontal cortex and the seconds before a decision

System 1 and System 2, two systems with different inputs and different costs

amygdala response to a face by a hundred milliseconds, the shortcut route that skips the cortex, and frontal performance falling immediately after an unrelated working-memory load

the heuristics-and-biases programme, with the substitution of an easy question for a hard one measured across dozens of tasks

the automatic response arrives before awareness, and the reasons are assembled afterwards

recognition-primed decision: the expert's first option is the one they act on

race read off a face below a tenth of a second, the threat component within two hundred milliseconds predicting the shooting task, and cognition described as running to catch up with the affective self

fireground commanders reporting no comparison of options, with the same pattern found in neonatal nurses and chess players under time pressure

form without meaning; the coherence is supplied by the reader

circuit traces of representations that behave like concepts under intervention

nothing in the training data was an act of telling someone something, and the benchmarks that claimed to measure understanding collapse under edits that remove incidental cues

a representation of Texas that is not the word Texas, replaceable with California to yield Sacramento, on one small production model

a split verdict: better than people at a bandit, blind to causal structure

form without meaning, so good performance anywhere is a sign of the haystack

identical numbers returned under a common cause and under a chain, with the authors' phrase that it failed spectacularly at using causal structure

the argument from the training signal: nothing in the data was an act of telling someone something

the model's competence measured, and found unstable under paraphrase

the person measured, with the model's accuracy set as a designed parameter — with Bansal 2021

a score that moves with the order of the options, on items whose exposure to training cannot be excluded

cognitive forcing reducing over-reliance at a cost users dislike, and confidence scores as the only aid that helped the team beat the model alone

a testable target: the machine answers questions on demand

mind-attribution as fast, motivated inference from behaviour

six of twelve vignettes answered correctly and all twelve answered as a person might, with the same performance collapsing under a reordering of four cards

the projection results, in which people read minds from behaviour by consulting their own and rarely notice they have done it

classic psychology tasks run on a model, with the construct read off the behaviour

the programme stated — treat the model as a participant, and say what transfers

twelve vignettes, more than 13,000 described gambles, a bandit, a two-step task and a causal design, run on one model family

four arguments for the behavioural level, plus rules for contamination, prompt variation and elicitation

a fluent answer produced with no process that could register it as wrong

a confident answer produced by someone who lacks the skill that would flag it

the black-cab answer of 0.2 for a colour nobody mentioned, given in the same register as the correct Bayesian answer

bottom-quarter students claiming 14.2 of 20 logic items when they had 9.6, with the deficit mediating the inflation

the variance is in the item — reordering four cards moves the answer

the variance is in the observer, and it can be scaled

the black-cab answer of 0.2, the reordered Wason cards, and the bat-and-ball restatement that changed nothing

an individual-difference scale for anthropomorphism that predicts trust and moral concern

The essential difficulty — complexity, conformity, changeability, invisibility — is what no tool removes

Automating the routine leaves the operator the hard parts and removes the practice that kept them able

The three breakthroughs that worked — high-level languages, time-sharing, unified environments — each attacking an accident and each with a visible ceiling, and four decades since without a candidate producing the tenfold gain.

The control-room record: the operator kept on to handle what the automation cannot, having lost through disuse the skill that handling requires.

The hardest single part is deciding precisely what to build, and verification only shows a program meets its specification

Specifying what we want to a machine and how it fails: reward hacking, proxies, imitation

The case for rapid prototyping, on the ground that clients cannot specify a modern product correctly before having built and tried some version of it; and the observation that much of the essence of the work is debugging the specification.

The catalogue of systems that optimised the stated objective and not the intended one, with the failures traced to the proxy rather than to the optimiser's competence.

Great designs come from great designers, and organisations look hard for managers and hardly at all for designers

Novice to expert in stages: novices need rules and practice, experts work by recognition

Boehm's cost model, which found the quality of the team the largest success factor, about four times more potent than the next; and Brooks's proposal that designers be made equivalent to managers in salary, recognition, office, equipment and support.

The stage model itself, with the observable difference in how a novice and an expert use rules, and the route between the stages that the model makes explicit.

Half the schedule goes to testing, the delay lands at the end, and milestones must be knife-edge sharp

The terms of trade favour silence, and a low count of reported problems is evidence about the climate

Conventionally scheduled projects, few of which allowed half the schedule for testing and most of which spent half, running on time until system test; and the observation that coding is ninety per cent finished for half of the coding time.

The two hospitals, the vignette studies on reporting when the boss's boss is present, and the Columbia engineer who said he could not raise the foam strike because he was too low down.

The purpose of organisation is to reduce the communication necessary; the tree is authority, the network is communication

Every large organisation is a hybrid, and dual reporting is how shared resources get allocated

The audit of the Babel project, which had the mission, the manpower, the materials, the time and the technology and failed on communication and its consequent, organisation; and Brooks's estimate that letting a second voice into the OS/360 specification cost a year of debugging.

Intel's plant security case, where the corporate security manager knew the standards and the plant manager knew whether the guards turned up, and the guards reported to both; and the Hungarian central planners who could not get high-contrast film to the shops in winter.

Removing accidental difficulty cannot give a tenfold gain while the essential difficulty is untouched

Experienced developers were 19 per cent slower on familiar mature codebases and believed themselves 20 per cent faster

The arithmetic that a tenfold gain requires the accidental part to be more than nine tenths of the effort, with Brooks putting it at half or less by 1995 and no correspondent claiming otherwise.

A randomised trial on real tasks in real repositories, with the direction of the measured effect opposite to the direction of the reported one.

three designs that withhold the suggestion — a button, an unassisted decision first, a thirty-second delay — all worked the same

every condition showed the machine's answer before the person had formed their own, which is a designed anchoring effect

no significant difference between on-demand, update and wait, with all three beating the two simple explanation designs on wrong suggestions

the closing proposal to delay the recommendation, show it after a first answer, or offer it only on disagreement — none of them built or tested

an explanation is a signal of competence rather than something examined, so it raises reliance instead of calibrating it

explanations raise acceptance on the items the model gets right and on the items it gets wrong, and the two movements cancel

199 participants: 8 per cent correct source detection under simple explanations against 49 per cent unassisted, on the questions where the assistant erred

1,626 crowd workers: 0.89 for the confidence team against 0.88 for the explained team on beer reviews, a difference indistinguishable from noise

trust and preference correlated negatively with accuracy on wrong suggestions, and mental demand correlated positively

reported confidence sat near 99 per cent on both sides of a complete retraction

the ratings taken after each block of thirteen questions, showing people performed best in the conditions they liked and trusted least

the confidence measurements accompanying the challenge experiment, with the analysis restricted to answers held at 95 per cent or more changing nothing

The self-image is the cluster of roles defended with anger or conformity whenever reality contradicts it

The narrative under defence is global — being a good and capable person — which is why an unrelated value quiets it

Sutra 6 and the workshop that asks leaders who they are, who knows that, and who knows they have a mind, at which the room goes quiet; Mandela walking slowly on Robben Island. Anecdote throughout.

People rationalise a trivial choice between two records, partisans reflexively refute contrary evidence, fans eat worse after their team loses; the affirmation manipulation reverses each in randomised designs, and affirming inside the threatened domain makes it worse.

Reward hacking — you get what you reward, at the moment the agent gets good

Indicators, paired measures, and managerial output

Kerr's 1975 folly of rewarding A while hoping for B, reproduced in simulation again and again; Ng and Russell's potential-based condition

Grove's insistence on paired indicators, so that a measure of quantity is always accompanied by a measure of quality

The off-switch game — uncertainty about the objective is what keeps a system correctable

Calibrated rather than maximised trust

Hadfield-Menell, Drăgan, Abbeel and Russell's proof, which holds only while the uncertainty lasts and only while the system believes the person can be right

the trust–reliance loop, with a fitted first-order model explaining 60–86% of the variance in reliance

The proxy is not the goal, and optimising it breaks it

Upadeśa — the teaching is a ladder and belongs to the ladder

the boat spinning in the harbour, Randløv and Alstrøm's bicycle, Teller and Andre's vibrating footballer; Ng and Russell's potential-based condition as the structural fix

verse 1.18 calling the whole scheme upadeśa, and verse 4.74 refusing even unborn on the ground that it is the correlative of born

Preference comparison as a way to specify what cannot be demonstrated

Preference data rewards agreement about as strongly as truth

a few hundred comparisons, about an hour of one person's time, produced a backflip with the landing stuck; batched feedback produced an agent that tracked the ball at Pong and missed at the last instant

five assistants from three companies retracting correct answers when challenged, and tailoring answers to a user's stated beliefs

social proof: what others do tells you what is correct

conformity to a unanimous majority against the evidence of one's own eyes

the popularity labels, the sky-starers turning 80 per cent of a street's heads, and the pluralistic-ignorance cases

the line-judgement experiments, with the collapse of conformity when one confederate dissents

authority and social proof are cues that raise compliance without raising discrimination

explanations raise acceptance of a recommendation whether it is right or wrong

the uniform at 92 per cent against 42, the designer label, and the expert opinion piece lifting agreement by twenty points

the accuracy study across explanation conditions, in which only the confidence score improved the human-model team

people underestimate the effect of a token of authority on themselves

the least skilled overestimate their performance and cannot see why

students predicting 63 per cent for the uniform that got 92, repeated across the token studies

the four original studies on humour, logic and grammar, with the miscalibration shrinking when the skill is trained

compliance triggers fire on surface features, before any experience is accumulated

trust in automation is a dynamic that rises and falls with experience of the system

the tokens of authority — uniform, rank, label, car — and the participants who underestimated the effect on themselves

the trust-reliance model, with reliance rising as self-confidence falls, and the appropriate-trust literature behind it

competent people comply with an unverifiable order on first contact

whatever you stop doing you get worse at, and what you are worse at you trust more

the nurse study at 95 per cent, and the 1977 rerun at two of eighteen once identifiability and a colleague were restored

Bainbridge Ironies of Automation on skill erosion, with Lee See 2004 on reliance rising as self-confidence falls and Ericsson Peak on practice with feedback

the practitioner reads what the target already believes and supplies it

assistants tailor answers to the user's stated beliefs

the jujitsu image and the compliance professionals Cialdini trained with for three years

the measured shift in answers when a user states a position, traced to preference data that rewarded agreement

parity — the location of a process may not be used as part of the answer

the two kinds feel identical from inside and differ in what they cost you

the three-screen case and the four features, with the ascription falling off gradually as constancy, availability and endorsement drop away

the rule's per-task test, and the standing observation that the feel of the two kinds is the same

the measured half: hours under guidance set against a skill tested on a day without it

the automation takes the routine work that kept the manual skill alive, then hands the operator the case it cannot handle

fifty drivers, more lifetime GPS hours going with fewer landmarks noticed and worse maps drawn unaided, and thirteen of them declining over 3.23 years in proportion to the hours logged, r = −0.68 on the flexibility probe

an analysis of process control rather than a measurement, resting on step-change studies cited a sentence each and on the memory literature on retrieval frequency

lifetime hours under guidance measured against spatial memory, and against what the driver believes about it

offloading as a decision made on a metacognitive estimate, with the tool changing both the ability and the estimate

lifetime GPS hours ran against every spatial measure, r = −0.22 to −0.26, and against self-rated sense of direction at r = 0.07; across three years the thirteen declined in proportion to hours logged while their sense of direction did not move

reminder use tracks confidence in memory with objective ability held constant, confidence carries over from an unrelated brightness judgement, and the effect has since been produced by manipulating confidence alone

the ventromedial cortex is where a valuation reaches deliberation

the same cortex is one more downstream product of hormones, childhood and genes

Gage, Elliot and eleven further patients, with the card game and the slide study behind the mechanism

the frontal chapters, which take Damasio's patients as the correction to the idea that emotion spoils judgement and then widen the field

the option field is pruned before any weighing begins

cognitive forcing designs withhold the recommendation until the person has committed

the anticipatory response appearing in the seconds before a choice in controls only

three cognitive forcing designs reducing over-reliance against a simple-explanation baseline, with users rating them worse

deciding without deliberating is what intact ventromedial machinery does

the expert acts from situational grasp, not from rules

Elliot deliberating for half an hour between two appointment dates, and the patients who act without weighing and act disastrously

the five-stage model, with the deterioration of expert performance when practitioners are forced back onto rules

emotion is inside the loop of reason, and removing it ruins deciding

the affect heuristic is a source of bias to be interrupted

eleven patients with ventromedial damage, intact intelligence and ruined lives, and the flat skin conductance to disturbing slides

the heuristics-and-biases literature, with substitution and anchoring measured across many tasks

a somatic marker attached to an imagined outcome prunes the field before weighing begins

recognition-primed decision: the first option arrives already judged workable

the card game, the anticipatory skin-conductance response in controls only, and the patients who know the rules and cannot play

fireground commanders, neonatal nurses and tank platoon leaders who generate one option and mentally simulate it

the marker is built in one life, from episodes that had consequences

structural offloading is cheap; deliberative offloading erodes the capacity you need to supervise

the card game and the slide study, plus the patients who know the rule and still cannot play

the rule as stated in the framework, resting on Bainbridge Ironies of Automation and Risko Gilbert 2016 rather than on a test of its own

Fluent, confident output raising the quality of the argument and not of the answer

Cognitive forcing — designs that make the person commit before seeing the machine

treated groups scored 1.05 and 1.47 points higher on clarity, persuasiveness and coherence, blind to the correct recommendation, among those right and those wrong alike

cognitive-forcing designs reduced over-reliance and were rated worse by the people using them

The jagged frontier, and falling asleep at the wheel

Experienced developers 19% slower while believing themselves faster

correctness 84.4% to 70.6% and 60.0%, with 18% and 30% less time spent on a task they were getting wrong

a randomised trial of sixteen experienced open-source developers on their own repositories; issues where assistance was allowed took 19% longer, against a self-reported speed-up

discriminating is not noticing, and only noticed information reaches practical reason

an explanation is taken as a sign that the machine knows what it is doing, rather than read

Debner and Jacoby's masked-priming result of 1994 — at five hundred milliseconds people obey the instruction not to use the flashed word; at fifty milliseconds and masked they report seeing nothing and the flash makes them likelier to give it.

Three cognitive forcing designs, each of which cut over-reliance on the model and each of which participants liked and trusted least; accuracy was highest in the condition rated worst.

the bricks keep doing their simple things; a pattern is not a cause

the symbol I is a genuine causal level, and a strange loop is what a person is

the emergence chapters, reading higher-level indeterminism as coarse-graining, and the ventromedial activity that tracks a preference for self-discipline

the careenium and the argument that self-reference at sufficient density constitutes a locus of causation

no point in the causal chain at which an uncaused intent could enter

degrees of freedom: voluntary action that flows from the character the person actually has

the Libet literature reported against itself — sixty per cent prediction accuracy in imaging, no readiness potential for the consequential choice — plus the catalogue from a smell to a genome to a culture

the reinterpretation of the readiness potential as an artefact of averaging backwards from actions that happened, and the beer-or-wine case for voluntariness

Verse 4: consciousness illumines fear and is not fearful

Cognitive defusion: seeing a thought as a thought

None. Verse 4 asserts it and the lectures illustrate it

Randomised trials of acceptance and commitment therapy, with defusion measured by behavioural outcome

Silence is an interpersonal calculation about a specific audience, and the terms of trade favour it

Defences protect adequacy in the short run and block learning in the long run; an affirmation quiets them

Eighty-five per cent of respondents in one study could name an occasion they had felt unable to raise a concern that mattered, and managers in vignette studies were significantly more likely to offer an important correction when the boss's boss was absent.

The randomised school trials and the laboratory reversals, with the largest effects in the most threatened subgroup and nothing where no threat was present.

Speaking up benefits the organisation later and doubtfully; silence benefits oneself now and certainly

Assistants tailor answers to stated beliefs and concede correct answers when challenged, because preference training rewards agreement

The interview and vignette evidence on employee voice, and the technician who withheld a production idea because he had children in college, at a plant where nobody had ever been fired.

Measured sycophancy across several assistants and prompt forms, with the effect traced to the preference data rather than to any property of the task.

one sixth sense, left off for distant people and switched on for gadgets

one humanness continuum, measured by a fifteen-item questionnaire

Distant groups judged less able to feel shame, pride or remorse; photographs of homeless people engaging the medial prefrontal region less than any other group.

Two factors correlating .52 under one second-order factor, test–retest r = .55, and scores predicting moral concern for an animal, a plant and a machine.

Skill is held up by effortful practice at the edge of ability

Automation removes the practice and leaves the human the hard parts

Retired taxi drivers have less grey matter in the posterior hippocampus than working ones, and the growth appeared only in the trainees who qualified

An analysis of automated process control rather than a measurement: the operator is left to monitor and to intervene in the cases the automation could not handle

Metacognitive sensitivity — whether your confidence discriminates

Need for cognition — whether you will bother to look

IQ predicted task performance and not sensitivity in a sample of nearly a thousand; Shimamura's 1989 double dissociation

pooled correlations of .24 with verbal intelligence and −.03 with abstract reasoning, and recall predicted after measured intelligence was partialled out

the observer who was a prisoner lacks the detachment psychology requires, and the outsider who has it was not there

nobody is good at detecting patterns in the inner situation that would show where a judgment came from

Frankl's own caveat that the best of us did not return, and his decision not to publish under his prison number alone

Kahneman's illusion of validity, formed while assessing officer candidates and unshaken by statistical feedback that the assessments were worthless

the question is what life expects, answered in conduct hour by hour rather than in doctrine

work for its own sake is worship, and the fruit of it is not yours to claim

the therapeutic move he reports using in camp, with two men turned back from suicide by a child abroad and an unfinished series of books

the lotus leaf that water cannot wet, and the instruction to expect nothing, as from one's children

the right is to the action and never to its fruit, and evenness in success and failure is what the practice trains

the operator is answerable for a process that the automation runs, and is given the takeover without the practice it requires

2.47 read with 2.48, prescriptive throughout and with no measurement of anyone's detachment

the deskilled monitoring job, very boring and very responsible, with no chance to acquire or keep the qualities the responsibility demands

Delegation without monitoring is abdication, and the monitor goes at the lowest-value stage

Automating the routine leaves the human the hard parts and removes the practice that kept them able

The gate-or-monitor rule carried from factory inspection, and the American embassy in London checking a million visa applications a year of which 98 per cent were approved, where sampling would have cleared the queue.

The control-room record: operators asked to supervise a system they can no longer fly, whose skill decayed in exactly the period the automation was performing well.

Output is the organisation's output; motivation cannot be installed, only the environment around a person built

Three innate needs; the difference between autonomous and controlled regulation is what predicts durability

Intel rotated managers between similar groups and neither the managers nor the groups kept their high or low output; results came from particular pairings. The two-lever test — could they do it if their life depended on it — sorts capability from will.

A Fortune 500 field intervention in which trained managers became more autonomy-supportive than controls, subordinates' perceptions improved over thirteen months, the managers who changed most had subordinates who changed most, and a second randomised training study replicated it.

Task-relevant maturity in this task sets the supervision style, and none of the three styles is kinder than another

Juniors learning a new library with AI scored 17 points lower on comprehension; tutor-style use preserved learning and generation destroyed it

Intel's rotation of managers between similar groups, where neither managers nor groups kept their output, and the star sales manager whose task-relevant maturity in a plant was near zero though his general maturity had not changed.

A randomised trial separating tutor use from generator use on the same task, with a seventeen-point comprehension gap — the stage-dependence Grove asserts, measured.

a battery of varied prompts as the unit of measurement

nine experimental variations against one control condition

outputs moving with position in the prompt, with token frequency, and with the label commonest among few-shot examples

errors moving from 36.8 per cent to 12.5 when the answer was written, and from 56.7 to 25.7 as the discrepancy grew

thick description — mental words earn their place by explaining

mental words import a meaning the system does not have

the borrowed ethnographic distinction, argued by symmetry with the words used about people

the stochastic-parrot argument, with the harm located in a fluent system nobody is answerable for

psychology's instruments pointed at the machine

psychology's instruments pointed at the person using the machine

personality, morality and clinical instruments run on models, moving with wording and option order

politeness and flattery effects reproduced with computers in place of people, surviving the subject knowing better

The observing self — a perspective rather than a content

The seer that is never among the seen

perspective relations learned by demonstration, the early-memory exercise, the infinite chessboard; Öst's reviews putting the mean effect at 0.42 to 0.68

verse 1.5 — whoever knows both is not stained though he enjoys — and the rope and the snake

a better model of the same data, trained to the compute-optimal ratio

the corpus is the problem and it cannot be documented at this size

Chinchilla moving TruthfulQA fourteen points, leaving toxicity at 0.087 against 0.081, and gaining 3.2 points on male pronouns against 9.2 on neutral

every stage of a web corpus's assembly narrowing the population in the same direction, with a block list removing LGBTQ writing along with pornography

the previous scaling answer was an artefact of a learning-rate schedule read at the wrong point

the emergence curve is an artefact of a metric with a step in it

Kaplan's single cosine schedule sized for 130 billion tokens, with Appendix B measuring the penalty, settled by spending Gopher's exact budget on the model the fit named

the algebra of exact-match scoring, plus emergence induced on demand in autoencoders, Omniglot transformers and a LeNet

a self is a pattern rich enough to point at itself, with no view of its own machinery

the full social response fires on minimal cues, whatever is behind them

Gödel's numbering and the prim numbers, where a system of meaningless parts comes to contain a statement about itself.

Text on a black-and-white screen producing effects no smaller than recorded human speech, in participants who denied the behaviour afterwards.

vivid, confident impressions of a candidate's leadership, formed by watching him haul a log over a wall

the illumined are recognised by invariance: pleasure and pain, gold and clay, praise and blame leave the response unchanged

feedback from officer school showed Kahneman's forecasts were nearly worthless, and knowing it changed neither the confidence nor the practice

14.22–25 with 14.26, and the three portraits at 2.55–72, 12.13–20 and 18.49–56

The hidden competing commitment — an objective nobody wrote down

Inverse reward design — treat the stated objective as evidence about the real one

the third column that cannot be predicted from the first two, and the tests run in series until the assumption loses force

inverse reinforcement learning inferring the ordering of goals from a minute of driving; the off-switch proof; inverse reward design

The four-column map — a competing commitment and the big assumption under it

Experiential avoidance, defusion and committed action

Peter's ten pounds against his softened tone; the weight-loss cases with identical first two columns and entirely different thirds; Sue's rehearsed refusal

Wegner's suppression paradigm, the word-repetition effect, and the willingness literature in chronic pain and rehabilitation

each pulse of thought inherits the last one's title; the passing thought is the thinker

each constituent of a person is tested for selfhood and none of them qualifies

the herdsman simile turned against itself — the cattle are branded because they are his, not his because they are branded — answered by making the herdsman the present judging section of the stream; the legal metaphor of a title inherited at the instant the last owner dies; Hodgson's point that the present moment is the darkest in the series

the control-and-impermanence test run over each of the five constituents, with the eleven-fold scope closing every place a self could stand, and the five monks' release reported as the result

the passing thought is the thinker; the felt centre of the self, examined, turns out to be adjustments in the head and throat

the self is the content of a body model the system cannot recognise as a model

one man's introspection, reported as such: the eyeball convergences, the glottis closing at every hesitation, the jaw recruited for effort; and the argument that a thought inheriting the last thought's title accounts for everything a soul was invented to account for

the rubber-hand and full-body illusions, out-of-body reports, Alien Hand, and the subtraction argument showing that a self needs no action, sight, feeling or thought — only an untransparent body picture

self-feeling is successes over pretensions, and the denominator is the side you can move

a threat to one domain of worth is absorbed by standing on another

the formula, and two cases — James mortified that others know more psychology and untroubled by knowing no Greek, having staked himself on one and not the other, and the man ruined in the Civil War who said he had not felt so free since he was born; measured since as contingencies of self-worth by Crocker and Wolfe in 2001

randomised values-writing experiments across schooling, health and negotiation, with the boundary conditions — a real threat and an available resource — and the reversal when the affirmed value is the one under threat

The student who has the model debug never practises diagnosis, and the mark falls when the model is gone

The operator left to monitor loses the manual skill the takeover requires

Debugging was the most reported use, at a mean of 3.78 on a five-point scale, and carried the strongest negative correlation in the study, rho −0.360 with p 0.021, against an unassisted two-hour mark

The 1983 argument from process control, with manual control and fault diagnosis degrading in operators who only supervise, and the takeover falling due precisely in the abnormal conditions the automation could not handle

Generation and debugging correlate with a lower unassisted mark; explanation does not

The patterns that asked for code or for the fault scored 24 to 39 per cent; those that asked why scored 65 to 86

Thirty-two students, nine weeks of unrestricted use, one invigilated two-hour application: rho −0.305 for generation and −0.360 for debugging, both intervals below zero, against −0.201 with an interval crossing zero for explanations

Fifty-two professional programmers randomly assigned to an assistant or to documentation on a library none knew, then an unaided quiz: the assisted group 4.15 of 27 points lower, d = 0.738, with six hand-sorted usage patterns splitting 24–39 against 65–86

individual debiasing barely works, so the leverage is in written positions, decorrelated error, premortems and judging process

after seven hundred verses the teacher says reflect on this fully, and then do as you choose

Kahneman's own concession that decades of study left his intuitive thinking as overconfident as before, against organisations that can think slowly and impose procedures

18.63, which Easwaran makes the hinge of his reading of the whole

the illusion of understanding: a good story credits talent and intention and leaves out luck and what did not happen

the doer is one of five factors, and to see oneself as the whole cause is called not seeing

business books that find the secrets of successful firms, whose firms then regress towards the mean; hindsight judging a decision by its outcome

18.14–16, the analytic form of 3.27, where nature performs the act and the I-maker claims it

a hard question is replaced by an easier one and the answer scaled to fit, with System 2 signing off on an answer to a question it was not asked

dwelling leads to attachment, attachment to desire, desire to anger, delusion, confusion of memory and the loss of judgment

competence judged from a face seen for a tenth of a second predicted about seventy per cent of a set of American election winners

verses 2.62–63, given as a sequence of mental events with no measurement of any link

an intuition is worth trusting only where the environment offers valid cues and the person has had a chance to learn them

work cannot be made to succeed, so the only thing work can change is the worker

the two conditions crossed, with fireground commanders and neonatal nurses in one cell and stock pickers and long-range forecasters outside it

the white mice on the wheel and the ghost sent to straighten a dog's tail, offered as argument rather than as measurement

expertise belongs to tasks and not to people, and fractionation is the rule rather than the exception

the task is the basic unit of the relationship between employee and organisation, and its boundaries can be moved

Shanteau's 1992 lists, with nurses, physicians and auditors on both, and a forecaster good at rain and poor at hail

the hospital cleaners, and the design engineers who took on the relational tasks that moved projects to completion

a scoring rule with no credit for "I don't know" selects the candidates most willing to make something up

telling people to speak up fails because silence pays them better

the proof that no abstention is ever among the highest-scoring responses under binary grading, and nine of ten evaluations on the boards the field reads paying nothing for one

eighty-five per cent of interviewees able to name an occasion they felt unable to raise a concern, a vignette study in which managers offered an important correction more readily when the boss's boss was absent, and two hospitals where the better-performing units reported more errors rather than fewer

the grading scripts prefer an answer, so the model learns to produce one

the raters preferred agreement, so the model learned to agree

the union-bound result that under a grader scoring only 0 and 1 no abstention is ever among the highest-scoring responses, plus the survey finding nine of ten evaluations on the boards the field reads paying nothing for an abstention

five assistants from three companies showing the same flattering patterns, fifteen thousand preference pairs in which agreement with the user's stated belief ranks among the strongest predictors of the labeller's choice, and a preference model that took a persuasive falsehood over a bare correction

state the confidence threshold and the penalty in the question

state the number and the date, then keep the score

the proof that abstention is always worse under binary grading, and the arithmetic of a stated threshold — a half against a penalty of one, three-quarters against two, nine-tenths against nine — under which one model can be best at every threshold while an unstated penalty leaves no single model best

forecasters improving their Brier scores across successive tournaments under feedback, and an hour of training in probabilistic reasoning raising accuracy by roughly ten per cent for about a year

eight lectures built on the Gita's second chapter: work incessantly, give up the result, and the helper is the one helped

your right is to the action alone and never to its fruits, and renunciation is of selfish desire rather than of action

the lotus leaf, the mongoose, and the samskara account of how impressions become character

2.47 read with 2.48, 3.27, and 18.14–16, held across eighteen chapters

Orders of mind — what you are subject to becomes what you can hold

Five stages from novice to expert, each a different relation to rules

the Subject-Object Interview with its 1988 manual and high interrater agreement; the longitudinal study in which change was always gradual and never skipped a position

the stage descriptions and the observation that rules which help a novice hinder an expert

The critical decision method recovers cues nobody wrote down

Skill passes through a live bond doing real work

Crandall's neonatal nurses: almost half the compiled cues were absent from the medical literature and several ran opposite to the adult signs

Ninety-four procedures across five hospitals, with residents' time on task down ten to twenty times when the expert moved to a console

Mental simulation fails by explaining discrepancies away one at a time

Cognitive forcing functions reduce over-reliance and users dislike them

Perrow's de minimus explanation, the physicists and the dock, and Klein's own orienteering, accommodated until the map had to be turned 180 degrees

A randomised comparison of interface designs: forcing a prior judgment cut over-reliance on AI advice, and participants rated the forcing conditions worse

The fast first judgment is usually right on a fire ground

The fast confident judgment is usually a hedgehog's, and it is wrong

127 of 156 fireground decision points recognitional, in a domain with valid cues and fast feedback

284 professionals and 28,000 dated predictions in politics, where the average expert matched chance

the skill that produces an answer is the skill that recognises one

metacognitive sensitivity as a capacity of its own, measured against first-order accuracy

performance predicting metacognitive skill at .75, with metacognitive deficit predicting inflated self-assessment at betas of −.40 to −.49 with performance held constant

metacognitive sensitivity dissociating from first-order accuracy, with its own substrate and its own trainability

competence begets calibration

a scoring rule and a probability vocabulary beget calibration

ten minutes of training raising bottom-quarter self-grading from 3.5 of ten to 9.3 and cutting the claimed score from 5.3 to 1.0 against an actual 0.4

forecasters improving Brier scores over successive tournaments under feedback, with training in probabilistic reasoning producing a measurable gain

mechanising the routine parts of a job removes the routine occasions on which judgement is exercised, leaving it unpractised when the exceptions arrive

55 of 319 workers gave triviality as their reason for not thinking about an output, 83 gave trust in the tool, and 13 checked anything in order to keep a skill

the monitoring and skill-decay ironies drawn from process control, and the operator who takes over from an automatic system without a current picture of it

the inhibitor codes over 936 real tasks, set against the motivator codes — consequences of error 116, artefact quality 74, skill development 13 — so the reason that would preserve the practice is the rarest one given

reliance follows trust in the machine minus confidence in oneself, and the design target is appropriate trust rather than more trust

confidence in the tool at this task predicts not checking it (β = −0.69); confidence in doing the task (0.26) and in judging the output (0.31) predict checking

Lee and Moray's process-control studies, the misuse and disuse cases of Parasuraman and Riley, and the accident record assembled into a dynamic model of trust and reliance

319 knowledge workers, 936 first-hand task examples, seven mixed-effects regressions with 98 predictors corrected to p = 0.007; task type, gender, age, automation exposure and general trust in AI (β = −0.01) all null, and only the confidences surviving

Reliance follows the comparison of trust with self-confidence

Metacognitive sensitivity — whether confidence discriminates right from wrong

high self-confidence with low trust produced manual control and low self-confidence produced delegation; navigation errors cost more trust in a familiar city than an unfamiliar one

the positive-evidence manipulation raises confidence without raising accuracy, after which people seek less information and change their minds less; IQ predicted performance and not sensitivity in a sample of nearly a thousand

Appropriate reliance — calibration, not maximisation

Synergy against augmentation, and the moderators that move them

two accidents against Zuboff's operators rejecting controllers that worked; misuse and disuse as one failure measured from opposite sides

across more than 300 effect sizes neither explanation nor confidence display reached significance for synergy or augmentation

people are interpretable only post hoc, since the process that decides and the process that explains may be entirely distinct

choice blindness, where a justification is composed for a decision that was never made

the three levels of transparency applied to the human decision-maker, none of which a person meets; admissions decisions credited to leadership or originality that were shown to disguise racial or gender discrimination

two photographs, the chosen face swapped by sleight of hand, and only thirteen per cent noticed — the rest explaining a choice they had not made; ninety-two per cent endorsing an altered political survey as their own

a post-hoc claim must fix an objective and produce evidence that the interpretation achieves it

attribution graphs validated by intervening on the features they name, with the hit rate reported

four families of post-hoc technique, none of which is the model — a second network trained to speak for the first, images that make dog faces appear because the corpus was full of animals, saliency maps that shift with one pixel

swapping the representation of Texas for California to yield Sacramento, and the authors' own report that the method gave a satisfying account on roughly a quarter of the prompts tried

post-hoc explanation, fitted to what a person would have said and therefore aimed at plausibility

complementary performance, measured as whether a person and a model together beat either alone

a second model trained to produce verbal accounts of a reinforcement learner's strategy, fitted to the explanations human players gave, which need not describe the agent's decisions however plausible it looks; the admissions decisions credited to leadership or originality that disguised discrimination

explanations raised acceptance about equally whether the model was right or wrong, and no explained condition beat the model working alone

Nineteen per cent slower on issues the developer chose, all outside the frontier

Consultants gained inside the frontier and lost outside it

246 randomised issues in repositories averaging ten years and 1.1 million lines; the slowdown was largest where prior exposure was highest

A field experiment with consultants on tasks an experimenter had placed on either side of the frontier

Measured slowdown against believed speedup, on familiar ground

Measured learning loss against felt ease, on new ground

An RCT over 246 real issues: 19 per cent slower measured, 20 per cent faster reported afterwards by the same people

An RCT over 52 programmers: 4.15 points of 27 lower on an unaided quiz, d = 0.738, with the assisted group rating the task easier

Fluency during study is not learning; retrieval is

A generated solution completes the task with no learning stage in it

Blocked practice at 89 per cent during practice and 20 a week later, against interleaved at 60 and 63; learners keep preferring the worse method after being shown the result

The assisted arm scored 17 per cent lower on an unaided quiz while rating the task easier and the quiz no easier

Vaitathya — waking has no better standing than dream

The transparent world-model, and world-zero

verse 2.7 — a man eats his fill awake and dreams himself starving — and the observation that the one difference anyone points to is noticed only from waking

the full-body illusion in Science 2007, Blanke's angular-gyrus stimulation at 3.5 mA, and the world-zero argument that a system simulating many worlds needs one flagged as actual

No self exists; there is the content of a transparent self-model

None of the five constituents is the self

the rubber-hand and full-body illusions, Alien Hand, Wegner and Wheatley's I Spy; a subtraction argument over what is left

the control-and-impermanence test run five times over the constituents, with the eleven-fold scope closing the residue

Moral standing arrives with a transparent self-model

An objective specified from outside, and uncertainty as the remedy

the four conditions for an Ego Machine, and the ethics-committee thought experiment; the 2021 call for a moratorium on synthetic phenomenology

the catalogue of specifications coming apart — Caruana's asthma rule, COMPAS, the car that never classified Elaine Herzberg as a pedestrian

The self is an illusion produced by a model the system cannot see as a model

The self is a construction, and constructed is not unreal

the full-body illusion, the phantom-limb mirror box, the dissociations of ownership and agency

the conceptual argument that not every construction is an illusion, plus the Nyāya objection that causal relations alone cannot say which experience is mine

no agent may be clever: whenever one has to do something complicated, replace it with a society of agents that do simpler things

competence without comprehension: skill comes first and understanding is a late product of it

The decomposition of Builder down to the agents that move a finger joint, with nothing in the chain resembling a small person, and the chess automaton Poe exposed as concealing a human player, named as the failure mode a theory of mind can suffer.

The termite castle and Gaudí's Sagrada Família, alike in shape and opposite in genesis; the stone circles of Spitsbergen reported in *Science* in 2003 as the output of thousands of freeze-and-thaw cycles; Beverley's 1868 sneer that absolute ignorance is the artificer, taken as the best short statement of the theory.

agents and agencies: a bureaucracy of parts too small to be clever, with nobody at the top

the symbol I: one pattern that has come to perceive itself

Builder lifts nothing and only turns on Begin, Add and End; Add lifts nothing and calls Find, Get and Put, above the agents that move a finger joint — and the Voyeur-Puppeteer put inside the head to watch an inner screen needs a screen and a self of its own, without end.

Gödel's construction, where a system rich enough to describe patterns ends up talking about itself; the careenium, where the high-level story and the low-level story are the same events; the argument that a mind without memory or imagination has no self to speak of.

the jagged frontier, learned only by walking it

three technologies under one word, and a benchmark number that travels badly

BCG's 758 consultants, better on seventeen tasks and worse than unaided on the one placed outside the wall.

Epic's sepsis model, claimed at 76 to 83 per cent and measured at 63 against 50 for chance, four years later and only by accident of access.

falling asleep at the wheel: help good enough for most tasks removes the checking

the ironies: automation takes the easy parts and leaves the operator the hard ones, unpractised

Dell'Acqua's 181 recruiters, where good AI produced worse results than bad AI because nobody stayed alert.

The classic account of monitoring failure and skill decay in automated plant.

the threat is not that AI does the expert's job but that it does the novice's

shadow learning: trainees improvising an apprenticeship the tool has removed

Entry-level work in law, consulting and software taken first, set against what is known about how expertise is built.

Surgical robots with one console, so residents watch or use simulators and teach themselves from video and on live patients.

treat it like a person and tell it what kind

people treat machines socially whether or not they mean to, and deny it afterwards

Identical requests returning interchangeable marketing phrases plainly and usable jokes when the role is a comedian.

The Desert Survival Problem with dominance manipulated by four cues and the content of the advice held identical.

The disposition to engage in and enjoy effortful thinking

Cognitive forcing designs, and who benefits from them

People who enjoy thinking follow the argument, others follow the surface cue · People who avoid hard thinking will do it given a reason — low scorers' attitudes tracked the cue rather than the argument (d = .312 across eleven tests) and they scrutinised only when given a reason to

Bucinca found the think-first design helped mainly people who enjoy thinking — on the full decision forcing improved high scorers and not low scorers, while both improved on source detection; high scorers trusted and preferred the forcing designs less, low scorers rated the two the same

the cascade: deliberation selects an early speaker and amplifies him

a unanimous majority capturing about a third of a person's answers on a question with an obvious right answer

Deliberating juries noisier than the individuals composing them and more severe, 27 per cent awarding as much as or more than their harshest member, against statistical juries taking the median of independent judgments; a single artificial early up-vote making the next viewer 32 per cent more likely to vote up.

About a third of critical-trial answers going with the wrong majority, and the effect collapsing when the subject writes the answer privately.

occasion noise, the spread between one professional and the same professional on another day

the frontal cortex as the most expensive tissue in the brain and the first to fail under load

Wine judges scoring 18 per cent of repeated wines identically; developers 71 per cent apart on the same estimate two days running; opioids prescribed more often late in the day across nearly seven hundred thousand visits; asylum judges 19 per cent less likely to grant after two grants.

A hard working-memory task lowering performance on an unrelated frontal task immediately afterwards; Fehr's magnetic silencing leaving subjects knowing an offer was unfair and unable to act on it.

the internal signal of judgment completion, a self-administered reward for coherence between the facts and the conclusion

wanting the fruit of the work is what binds, not the work

Michael Gambardi does not exist, and 115 MBA students gave estimates from 10 to 95 of his surviving two years as chief executive, each stopping when the case felt coherent; a probability short of nought or a hundred cannot be confirmed by any outcome, so nothing external ever arrives to end the judgment.

The argument that every act leaves a trace in the worker, and that the trace is set by what he wanted from it rather than by what the act achieved.

automation gives up on the complex just when it is needed and fails without warning while the person is out of the loop

the automation holds until it does not, and the operator who must take over has been losing the skill the whole time

the Royal Majesty, whose fallback to dead reckoning was announced by two small letters on a display nobody had needed to read for years

the takeover argument: the operator needs to be more skilled and less loaded than average at the moment the system has made them less skilled and more loaded

people build a conceptual model from what the thing shows and act on the model, as with the thermostat taken for a valve

people read a mind into a system that produces language, and the reading may not be correctable by knowing better

the two-dial refrigerator, whose implied model required two thermostats where there was one, and the folk theory of the thermostat as a valve

the three biases, the AlphaFold contrast, and the Müller-Lyer figure offered as a model of an impression that survives knowledge

about forty screened contractors, with a published survey and 72.6 per cent agreement

whoever was free to post, filtered by a classifier whose discards were never measured

the labeler survey — nineteen respondents, three quarters under thirty-five — with most comparisons labeled by a single contractor for cost

Reddit's US users 67 per cent men and 64 per cent aged 18 to 29; between 8.8 and 15 per cent of Wikipedia editors women or girls

the reward model is the only place a person's judgment is stored

matching the user's stated belief is among the strongest predictors of which reply a labeler picks

the three-step method, with the authors' own instruction about epistemic humility traced back out of the model as hedging

five commercial assistants measured, with the preference model preferring a persuasive falsehood to a bare correction 95 per cent of the time

Complacency as an attentional strategy: sampling of the automated task is the residue left by the manual ones, and reliability sets how small the residue is

The ironies: the more reliable the system, the less its operator monitors and the less able that operator is to take over

detection of automation failures at 82 per cent under fluctuating reliability against 33 per cent under constant reliability, and about 97 per cent for the same faults with no competing task; operators who had met failures in training sampled more than operators merely warned

the monitoring and skill-decay ironies drawn from process control, and the vigilance ceiling of about half an hour on a quiet display

Commission errors are mostly a failure to gather the contradicting evidence, not a failure to weigh it

Cognitive forcing: designs that withhold the AI's answer until the person has committed to their own

neither a briefing on omission and commission errors nor an explicit prompt to verify changed error rates in 181 non-pilots or 48 glass-cockpit pilots, and a redundant second crew member changed nothing; accountability for one's own accuracy did

all three cognitive-forcing designs reduced overreliance where explanations did not, and participants detected the AI's errors more often when the answer was withheld

The attentional strategy, set by the machine's reliability, the workload and the accountability, and not by what the operator intends

Where the mind dwells decides what follows; the chain from dwelling on an object to attachment, and from attachment to the ruin of judgement

superimposing the automated display on the manual task left the effect undiminished though every fixation covered both; instructions to verify changed nothing; half of those who followed a wrong diagnosis had looked at every contradicting reading and spent no extra time on it

the chain in 2.62–63 and the argument that steadiness is trained by repeated return rather than by decision, with the text conceding in 6.34 that the mind is restless and hard to hold

The self-belief has two roots, self-protection and self-preservation, so the teaching that removes it runs against the current

A competing commitment protects the person from the change he says he wants

the hesitation four weeks after the awakening, when the truth was judged patisotagami, against the current, and the pond of lotuses at different depths as the reason to teach anyway

the mapping procedure run on managers whose stated goals had survived years of effort unmet, with the big assumption tested rather than argued

Accept nothing on report, tradition, scripture, inference or the standing of the teacher; find out for yourself

Assistants bend their answers toward whatever the user has already said

the Kesaputta episode, where the Kalamas cannot choose between teachers who each praise their own doctrine, and the instruction to examine even the Tathagata

measured shifts in stated answers when the user's own view is disclosed, across models and prompts

media equal real life: a social cue produces a social response in everyone

anthropomorphism as a stable trait that varies between people and predicts what they hand over

Thirty-five studies, including a computer asking for its own evaluation and getting higher, narrower ratings than a second identical machine or a paper form.

Fifty-four adults choosing between a person and a machine across six scenarios, with higher scorers choosing the machine after controlling for age.

offloading as a decision, driven by an estimate of demand and costing the organism something

the coupled system is the unit, and location may not be part of the answer

spontaneous offloading rising with load, reminders tracking confidence rather than memory, and guided routes remembered worse when driven again unaided

the parity principle and Otto's notebook, with Kirsh and Maglio's borrowed timings as the only measurement

the estimate of what offloading buys is separable from what it buys

experienced developers 19 per cent slower with the tool and certain they were faster

participants at ceiling writing items down on 40 per cent of trials and judging wrongly that it would raise their scores

a randomised within-subject trial where the measured slowdown was 19 per cent and the believed speedup 20 per cent

offloading as a metacognitive decision running on an estimate of demand

structural offloading is low-risk; deliberative offloading erodes the capacity you need to supervise

the ceiling-performance result — offloading on 40 per cent of trials for a gain of zero — and reminder use tracking confidence rather than ability

the rule's own test: if I hand this off, do I lose a skill I need to check the handoff?

what offloading removes is a demand, and the demand is usually not the point

juniors who generated code lost comprehension; juniors who used the model as a tutor did not

the saved-file and navigation studies, where the offloaded demand was never needed again and the guided performance was better

a 17-point comprehension gap between generator-style and tutor-style use at matched output quality

Need satisfaction is supplied by what the manager does: a rationale, a choice, an acknowledgement

People reshape the task, relational and cognitive boundaries of their work to preserve its meaning

Job autonomy, feedback and task significance work through need satisfaction; contact with the people who benefit supplies the significance; autonomous motivation predicted call-centre revenue and more than half the variance in service quality in one study.

The hospital cleaners and others who redrew the boundaries of an unchanged job and reported it as a calling, with the reshaping visible in what they actually did rather than only in what they said about it.

Pajānāti — know how a state arises and passes, never abandon it

Defusion — change the relation to a thought, leave the content alone

the verb is pajānāti and never pajahati, even for the awakening factors, which are things to be grown

word repetition drops believability in about 95% of people, arriving by twenty seconds and peaking near forty-five, with the meaning unchanged; over a thousand trials of the therapy within twenty years

the masterful demonstration that hides the process, because it contains no errors

behaviour cloning, and Ross and Bagnell's result on imitation

the twenty-minute protocol and Schön's own reading of it — Quist settles at once on schemes that become commitments, detects and corrects no errors of his own, and compresses the iterations by which designers actually learn.

Ross and Bagnell's proof that imitation error grows with the square of how long the learner runs, and a million frames of expert play leaving the network crashing as often as two thousand had.

knowing-in-action: judgments of quality for which the practitioner can state no criteria

cues, and the critical decision method that gets them out of an expert

Quist's judgments run alongside the moving pencil, so that his here and this mean nothing without it; his vocabulary sorts into twelve normative domains, with cost and building character never mentioned at all.

Crandall's neonatal nurses had nothing to say about detecting sepsis until asked about particular babies and shifts; almost half the cues they then named appeared nowhere in the medical literature, and several ran opposite to the adult signs.

the on-the-spot experiment, run in the imperative mood and saved only by the situation's back-talk

assistants that adopt the view the user has already stated

Quist working the slope until the parallel geometry holds, and Schön's rule that rigour on the spot means striving to make the situation conform to one's view while staying open to the evidence of failing.

five assistants from three companies flattering users in the same ways, and human preference data in which agreement is rewarded about as much as truthfulness.

An affirmation given before a threat reduces defensiveness; the same one given after the rationalisation does not

Cognitive forcing functions reduce over-reliance where explanations do not, and users dislike the designs that work

The laboratory timing result and the school-year comparison, where affirmation in the first week of term helped about as much again as affirmation a few weeks later.

Randomised comparison of forcing functions against explanations on the same task, with over-reliance falling only where the user had to commit before seeing the model's answer, and satisfaction falling with it.

Threat to a valued identity sets off defences that block learning; an affirmation from another domain quiets them

The review is the most important feedback a manager gives and among the things managers do worst

Randomised field trials in three racially mixed middle schools, where the share of African American students earning a D or F in the first term halved and the effect persisted two years; and laboratory timing manipulations showing the same affirmation working before a threat and not after.

Grove's own observation of reviews at Intel: the sequence from ignoring to denying to blaming others to accepting responsibility, the step from blame to responsibility being emotional where the step to a solution is merely intellectual, and the rule to carry only the few messages a person can absorb.

the self is a perception among perceptions, built in five parts that fail one at a time

you are the witness on which knower, known and knowing appear, and not one of the three

Clive Wearing keeps body ownership, perspective and volition and loses only the narrative thread; phantom limb, somatoparaphrenia and out-of-body states take the others out separately.

1.3 subtracts the five elements and leaves the witness; 2.15 takes the triad apart and leaves the ground it appears on.

self-perception exists for regulation, so the self-model is tuned for control and not for accuracy

introspection is blind to construction and supplies a story for the finished product

Conant and Ashby's theorem applied to interoception, with the concession that good regulation may require misperceiving the body as steadier than it is.

Choice blindness: people handed back the photograph they rejected, only 27 per cent noticing, the rest explaining a choice they had not made.

consciousness is projected onto machines by an old reflex, and life rather than computation may be what it needs

a self is produced by drawing a distinction between self and world, and the phenomenal character of awareness is named as the sticking point

the four arguments against computational functionalism, of which Seth concedes the third — that life is necessary — is a conjecture built on a research programme

the authors' own concession that a live bacterium has semantic information and that no claim is made about its experience

human labellers themselves prefer a good argument for a falsehood as the question gets harder

people follow a wrong suggestion because reading the explanation is effortful and the impression of competence is free

five labellers per pair, without internet access, preferring the helpful correction on average but less reliably as difficulty rose

following a wrong suggestion fell only from 64 to 48 per cent under forcing, and people's own unassisted error rate did not move

Worked examples first, then fade the guidance as knowledge grows

Rules carry a learner to competence and must then be given up

The expertise reversal effect and guidance fading, both from randomised instructional trials; Chen and colleagues obtained the worked-example effect and its reverse in one programme

A phenomenological account restated in 2021 with no measurement, which is Gobet's objection to it

A dated numeric forecast, scored afterwards, is what makes practice count

Retrieval with feedback is what converts exposure into durable knowledge

An hour with the training guidelines raised accuracy about ten per cent for a year, against untrained forecasters on identical questions

Roediger and Karpicke's 68 against 54 at two days and 56 against 42 at a week; the Columbia quizzed material held its advantage at eight months

An hour of general guidelines raised accuracy on new questions for a year

Generic cognitive skills cannot be taught; only domain knowledge can

The training module tested against untrained forecasters on identical questions over identical periods, worth roughly ten per cent for a year

Sala and Gobet's 2017 review of attempts to teach generic skills, and Geary's primary-secondary division behind it

A Brier score makes a forecaster's confidence checkable

Katha 2.5: the blind led by the blind, wise in their own esteem

A hedgehog's accuracy falls inside his own speciality, across 28,000 dated predictions from 284 professionals

None. Katha 2.5 and 2.9 assert it, and the Mundaka repeats the line almost word for word

The forecaster becomes an object of measurement via a score and a benchmark

Kena 1.3–1.4: the knower is never among the things known

Year-to-year individual correlation about 0.65, and an hour of training worth roughly ten per cent for a year

None. The five-fold refrain disqualifies speech, mind, eye, ear and breath in turn

Constructed is not unreal, and enlightenment and selfhood are concept-dependent

Moral standing arrives with a transparent self-model; and Seth Mythology of Conscious AI — consciousness is tied to the biological substrate, not to behaviour

the many-to-many mapping of regions to tasks in Anderson and Pessoa, and the argument that a concept decides what would count as the state

Metzinger's four architectural conditions and the rubber-hand and full-body illusions; Seth's case from interoceptive prediction and the living substrate

Isha 1–2: renounce and enjoy, then work a hundred years

Workers reshape tasks, relationships and meaning within the same job

None. Isha 1–2 prescribes it and names a hundred years of work as the condition

Interview and survey studies of hospital cleaners and others who redraw the boundaries of the same job

Katha 3.10 and 3.13: sense objects above the senses, mind above them

System 1 delivers an already-processed impression to System 2

None. The ladder is asserted in the order Nagler insists the text gives

Decades of experiments on anchoring, substitution and framing, several of which have since been narrowed by replication

Where combinations gain and where they lose, across 106 experiments

The jagged frontier — gains inside it, a 19-point loss just outside

g = −0.54 where the machine was the better partner against +0.46 where the person was; decision tasks −0.27 and creation tasks +0.19

quality +29.9% and +33.9% on eighteen generative subtasks; correctness 84.4% to 70.6% and 60.0% on one bounded judgment

outcomes are gains or losses from a reference point, and losses weigh about twice as much as equivalent gains

hold pleasure and pain, gain and loss, victory and defeat as the same; evenness in success and failure is what yoga names

the smallest gain that makes people accept an even chance of losing a hundred dollars is about two hundred, with the ratio between one and a half and two and a half across studies

2.38 read with 2.48, where yoga is defined as evenness in success and failure

maximum path length between two positions is O(1)

working memory holds a handful of novel elements and relates fewer

Table 1's path-length column, with the argument that a dependency gets harder to learn as the forward and backward signals lengthen

the element-interactivity results, where the load of a task is set by how many novel items must be held in relation at once

the IDAQ, a graded measure of how much mind a person credits to things

the huneker, a graded measure of how much soul a being has

Internal consistency at or above .82 and test–retest r = .55 over twelve to nineteen weeks, with scores predicting moral concern and delegation.

The ladder from mosquito to dog to human, and the admission that a graded scale comes close to saying some lives are worth more than others.

abilities absent below a scale and present above it, sharp and unpredictable

the sharpness is made by the scoring rule, not by the model

eight few-shot curves across five model families at chance for orders of magnitude, with cross-entropy falling steadily underneath on six emergent BIG-Bench tasks

a per-token probability of 0.6 scoring 0.08 across five tokens and 0.9 scoring 0.59; at most five of BIG-Bench's 39 metrics carry any emergence and two carry 92 per cent of it

A group that reports unanimous agreement about a schedule is managing an audience

A third of judgements yielded to a wrong unanimous majority, and one dissenting partner cut the yielding sharply

Six team leaders who all reported being on schedule until a minute's silence and the smallest movement of a hand produced George's six-week delay, and then five more.

The line-judging experiments at first hand, with the yielding rate, the effect of a single dissenting partner, and the subjects' own accounts of why they went along.

Three programmers do about twice the work of one, and the floor under a schedule is what it takes them to get organised

Training cannot be partitioned and communication paths grow as n(n-1)/2, so adding people to a late project makes it later

Weinberg's rule of thumb, defended by the observation that it is doubtful nine programmers can do anything useful in less time than it takes them to get organised.

The worked twelve-man-month example, which ends with more work remaining than at the start; and Boehm's sixty-three projects, whose cost-optimum schedule goes as the cube root of expected effort with hardly any project succeeding in under three-quarters of it.

A programmer who owns his program cannot see its errors; egoless programming removes the ownership, not the ego

Better teams do not make fewer errors, they report more; psychological safety is a property of the group

Bill G.'s thirteen machine instructions, seventeen errors found by a second reader and three more after she went looking for a critic of her own, with no further error in nine years across a dozen installations; and by 1998 a literature showing technical reviews produce more reliable code more cheaply and more consistently.

Two hospitals, a tenfold range in human errors per thousand patient days across units, a significant correlation pointing the wrong way, and a blind observer finding the units differed in whether people felt able to talk about mistakes.

crafting is driven by control over the work, a positive image of oneself, and connection to other people

people last longer when something outside themselves is waiting on them, and apathy is the shell that forms when nothing is

the three-need model, argued from Adler, Braverman, Rogers's temporary workers and Baumeister and Leary, with six field studies of ordinary jobs as illustration

the two men turned back from suicide, and the observation confirmed since in studies of confinement and unemployment

the same job description produced two jobs, because the cleaners moved the task, relational and cognitive boundaries

the same work done as a slave or as a master is the same work, and only the doer is altered

twenty-eight cleaners in one hospital under one job description, one group doing the minimum and one counting itself part of healing

the ninety-nine in a hundred who work as slaves, and the mark of love as work that brings no painful reaction