40The Skill Code

Matt Beane, The Skill Code: How to Save Human Ability in an Age of Intelligent Machines, HarperBusiness, 2024; with "Shadow Learning: Building Robotic Surgical Skill When Approved Means Fail", Administrative Science Quarterly 64(1), 2019

Linked to 27 chapters: supports 22 · contradicts 6 · mechanism 10 · same claim 4 · pairing 5

Beane is an ethnographer who spends one to two years at a time standing next to people who work with robots, and what he found in an operating room he then went looking for across thirty-one occupations. His claim is that the skills a person can use under pressure are not acquired from courses or manuals but through a working bond between someone who knows how and someone trying to learn, a relationship he dates back at least 160,000 years and which surgery still calls see one, do one, teach one. That bond delivers three things, and he argues they are the whole of what learning on the job requires: challenge, work near the edge of a person's capacity; complexity, exposure to the context in which the task sits; and connection, trust and respect with the people who decide what a learner may touch. Intelligent machines rarely threaten this by doing the expert's job. They threaten it by making the expert productive without help, which turns the novice from a necessity into a cost. Then the three fail together, and a small minority of learners get their skill anyway by bending rules, out of sight, at risk to themselves and sometimes to other people. Beane calls that shadow learning, treats it as evidence rather than a recommendation, and builds the last third of the book on it: what individuals, organisations and technologists would have to do differently, and what the same technologies could do for learning if anyone designed them for it.

The 160,000-year-old school

The book opens on a field trip. At nine, with his chin on a fence post at Old Sturbridge Village in Massachusetts, Beane watched a master tinsmith work a sheet of tin into a candleholder with an apprentice, and what held him was not the candleholder but the way the expert shaped the apprentice: keeping him challenged, pushing him into more of the process, depending on him, and taking the piece back for the difficult move while the apprentice watched. He sets this beside Athens in 507 BC and a twelve-year-old who fetches lumber and brass, hauls marble, cleans up after the senior boys and watches, until six years later he carves on his own — a marble Orestes and Electra survives, signed as the work of Menelaos, pupil of Stephanos. Formal instruction, Beane argues, has never replaced this. Courses and books buy a person the right to start playing; the capacity to deliver under pressure still comes from working alongside someone who already can.

Then he moves to Rochester in 2020 and a resident he calls Kristen. In an open prostatectomy she wheels the patient in at half past six, leads the eight-inch incision, and calls the attending only when the field is set. For the rest of the two hours four hands are inside the patient, hers leading, and she is allowed a little of the nerve-sparing dissection that preserves erectile function. The attending breaks scrub and goes to his paperwork; she closes by a quarter past eight with a junior resident over her shoulder, and lets him place the final sutures. Six months later the same operation is done with a robot. The four-armed, thousand-pound machine is attached to the patient, both surgeons leave the table for consoles fifteen feet away, and the attending does the procedure. Kristen watches. She will be fortunate to operate for fifteen minutes of four hours, on fat rather than nerves, and when she does, the critique is shouted across the room for everyone to hear, or a tap on a touchscreen takes the controls away. In her first independent job the patient loses about ten times the blood, and an operation that should take three hours takes seven. Beane put the question to a chief of surgery, who said of the surgeons already doing this work that they had watched it happen, and that "Watching a movie doesn't make you an actor."

The surgical study behind this is a pair of studies. The first is a two-year comparative ethnography, from 2013 to 2015, at five hospitals in the north-eastern United States, three of them world-renowned teaching institutions and two of them places that took their rotations; it rests on ninety-four procedures, 478 hours of observation, 4,772 single-spaced pages of notes, and sixty-two formal interviews with eighteen surgeons, sixteen residents, twelve nurses and ten scrub technicians. The second is a blinded snowball interview study at thirteen further top-tier teaching hospitals, in which each surgeon he interviewed nominated two residents, one they judged exceptionally fast at learning robotic surgery and one average or below, without telling Beane which was which until every interview was done; that produced thirty-three surgeons and thirty-three residents, twelve of them rated exceptional and nine average or below. The measurement that matters is simple. In the open procedures he watched, which averaged four and a half hours, residents did surgical work for essentially the whole of them. In the robotic procedures they had ten to twenty times less time on task.

Surgery is only where he started. He wanted early adopters because they show the rest of the economy its future, and having found the pattern in one operating room he spent eighteen months reading field studies of work with nothing in common with it. Corporate law firms automated document review, stopped billing junior time, and produced what a trade review called a generation of lawyers missing their training. Police recruits were moved off the beat with their mentors and onto the paperwork that feeds the predictive systems. Senior bankers with FactSet and CapIQ could value a company without junior analysts, who were redeployed onto rote reports; 2021 brought record profits and a complaint from those same bankers that juniors no longer understood the business. His running list of affected occupations includes bomb disposal, chip design, journalism, data science, neonatal intensive care, radiology, long-haul trucking, ridesharing, food service, construction, wealth management and call-centre work. He estimates the cost in trillions, and points out that the half-trillion dollars spent worldwide on skills development in 2020 went almost entirely to formal training, and almost none of it to the relationship that actually produces the skill.

Challenge, complexity, connection

Beane's method for the middle of the book is to take one occupation per component, tell the story of someone who has it in good health, then assemble the research behind it. Andre, a deputy chef in a three-star steakhouse, has challenge; Sita, a shift supervisor in a Baltimore warehouse, has complexity; Emily, a director of enterprise sales, has connection. In each case the thing arrived through a person, not a course of instruction, and could have failed to arrive.

Challenge

Andre came up through food prep under a deputy chef called Jane, who gave him the easy peripheral tasks first, debriefed him over break, told him the truth, and then handed him work slightly beyond what he thought he could do, having made clear she would not rescue him. Robert Bjork's work on desirable difficulties showed that rereading notes feels like learning and is not, while the strain of self-testing costs time and confidence and returns retention. Mark Guadagnoli and Timothy Lee added that how much difficulty helps depends on what the learner can already do, and named the optimal challenge point. Kazunori Akizuki and Yukari Ohashi measured it: on a posture task, with difficulty indexed by salivary alpha amylase and the NASA task-load scale, learning improved with stress up to readings 81 per cent above normal and 51.5 on the scale, and fell away after that.

What an expert adds is range. Lev Vygotsky named the band of tasks a learner can do with guidance and not alone; Benjamin Bloom found that an average student tutored one to one reached the ninety-fifth percentile, and then spent his career on the fact that such tutoring cannot be supplied at scale. From the scaffolding literature Beane takes two obligations: frustration management, naming the payoff before the struggle and the distance covered after the failure, and fading, withdrawing as the learner takes the load. But scaffolding assumes attention the working day does not supply, and his correction is that challenge is a contact sport. Across eighty years of occupational field studies, novices compete for hard problems because problems are scarce, and everyone else has an interest in the work being done by whoever is best at it. Experts already had reason to keep novices away; the technology adds another.

Complexity

Sita began putting perfume samples into bags on a line and rose because a supervisor named Philip rotated her through the work with two-word instructions and no explanations. Watch the taping: she learned that the machine applied pressure, went back to her station, and began shaking the bags level. Watch the product: she saw where items were broken, lost or stolen on the way to the bins, went back, and told her colleagues to take from the top rather than dig. Only then did he ask her what was wrong with the line, and when she found that replenishment stalled because small items lodged in the bins, he promoted her on the spot. Complexity, in Beane's sense, is not the task but the context of the task, and it is learned by exposure rather than instruction. Antoinette Gentile found that stroke patients trained on perfect movements in simplified rooms did worse than patients dropped into ordinary conditions after minimal practice. Instruction should point at where to look and stop there; Richard Magill's rule is to ask a trainee where they were looking to read the pitch, not how the ball was released.

Two failure modes follow. One is reinvestment, applying rules consciously to movements that should be automatic: given three thousand putts, the golfers who accumulated the longest lists of rules putted worse, and other players called them chokers. The other is what Beane calls snapshotting, taking one visit for the whole situation, which he illustrates with Kayla Jordan's finding that people who watched a trivially informative video of an airliner landing became confident they could land one. Experts have the opposite problem: they stop looking. Trafton Drew's radiologists, searching lung scans for nodules, missed an image of a gorilla forty-eight times the size of a nodule in 83 per cent of cases, with eye-tracking showing that most had looked straight at it. Hence the argument's least expected turn. Novices' naive questions and clumsy work are what break the expert's trance, so the expert needs the novice for reasons that have nothing to do with charity.

Connection

Emily runs a sales team and builds her day out of relationships: a quarterly review in which she tells a salesman she is disappointed because he is not pushing his prospects hard enough, asks whether he wants pointers, and when he says he would rather try alone, tells him she trusts him. Beane grounds this in self-determination theory. Competence is one of three needs; the other two are autonomy and relatedness, and none of them is met alone. Where the goals come from is socialisation, and his case is John Van Maanen's year as a ride operator at Disneyland in 1972, where forty hours of formal training mattered less than copying senior staff until customers had become guests and rides had become attractions. What the novice must earn is access, and Howard Becker's Boys in White showed medical residents earning it through comportment rather than technique: whether senior doctors thought they handled patients, nurses and staff properly, and whether they knew their mentor's preferences.

The three tactics Beane extracts are attunement, feedback about the relationship rather than only the task, and joint adjustment of goals and assignments so that both parties keep some authorship. The relationship also has an ending built into it. Ed Hutchins's study of ship navigation describes a chain in which each person coached the one below while doing their own job, and advancement depended on both; the same pattern runs from film production to midwifery, where the expert eventually advertises the novice and the bond converts into a peer one. Where organisations prevent that, Beane calls it talent capture, and his case is an ethnography of the big six accounting firms in which specialists were held inside managerial fiefdoms, denied outside connections, and left feeling like corporate clones.

The threat

The mechanism of the threat is not malice or job loss. Beane grants the estimate of three to six jobs lost per robot purchased, then argues that the more consequential number is job change, which runs to hundreds of millions of people who must learn a new way of working using a bond that is being dismantled at the same time. His clearest example is chip design. An engineer can now run an overnight search across chip layouts that would take a good human five or six days, then judge whether the result is creative enough. Before the tool, senior engineers relied on juniors for the simpler analyses, prepared them for it and reviewed the output. Afterwards the junior is optional, for two reasons that compound: novices are slower and make mistakes, and attending to novices pulls the expert away from what the expert is best at. So the junior is redeployed to simulations of one subcomponent, and never sees a whole design.

The numbers he assembles all point the same way. Natalia Emanuel and colleagues found that engineers in the same building as their teammates gave 23 per cent more feedback than distant ones, that this fell to 6 per cent when the offices closed, and that senior engineers became more productive the less time they spent with juniors. Beane adds a finding he says he has never published: across six years of anonymised records at one top teaching hospital, each resident present in a robotic operating room added 25 per cent to the patient's time under anaesthesia, a risk factor for stroke. Removing the two residents of an average four-hour procedure cuts anaesthesia time by 56 per cent and saves the hospital $9,026. That is the case for cutting novices out, stated as strongly as it can be, and Beane does not dispute it. The scale comes from the 2023 exposure studies of Rob Seamans and colleagues and of Daniel Rock with colleagues at OpenAI, which scored the 19,265 tasks of 1,016 occupations in the American O*NET database. The headline was the handful of occupations near total exposure; the number Beane takes is the quiet one. Eighty per cent of working adults hold jobs at least 10 per cent exposed, which is 108 million people in the United States and 2.7 billion worldwide, and that counts only work done with words and ideas, not the physical tasks robots are now reaching.

What each component's failure looks like on the ground he gives through three people. Together they make his point that the three components fail at once rather than in turn.

When approved means fail

The paper explains the failure precisely, and the explanation is not that the robot is hard. Four features of the work reverse at once. Residents' errors, formerly private and often invisible to anyone but the attending, are magnified up to ten times and shown on screens around the room. Recovery, formerly a matter of the attending putting a finger on a vein and throwing a stitch, becomes coordinated work involving the scrub, the camera and the suction. Involvement, formerly necessary, becomes optional: the attending can operate alone. And delegation, formerly continuous and tacit, a hand opening a plane of tissue so the resident can clamp the artery, becomes a binary public act performed with a tap on a touchscreen. The riskiest phases move from the middle of the procedure to its beginning and end, which is exactly where residents used to work unsupervised. The result Beane calls helicopter teaching, and his field notes record three minutes in which a resident is talked through a dissection, mocked, and removed from the console.

The consequence is that residents finish legally entitled to do work they cannot do. As of 2015 the average American urologic surgeon who performed any robotic surgery performed one robotic prostatectomy a year, neither enough practice to build the skill nor enough to keep it. The paper's claim sits in a single reversal. In open surgery access comes first and competence follows: the attending cannot work without the resident, so the resident is admitted slightly ahead of being ready and learns by doing. In robotic surgery the order flips. Only a resident who looks fluent within moments of sitting down is allowed to continue, so competence must be acquired before access is granted, and the approved routes cannot supply it.

The minority who managed it did three things, in rough sequence; all twelve residents rated exceptional reported doing all three extensively, against four of the nine rated average who mentioned any of them at all. Premature specialisation meant finding robotic surgery years before the training allowed, through a research assistantship in a robotics lab, a mentor who was a robotic specialist, or time on the system after a case was finished, at the cost of the generalist education everyone insisted on. Abstract rehearsal meant simulator practice and video analysis in a culture that held both too far from reality to count: programmes required two to four hours of simulator time a year, and successful learners put in many tens and sometimes more than a hundred, while one resident described watching an hour-long video of a procedure about two hundred times. The book renames this digital rehearsal; the paper's term is the one to keep, because the point is the distance from the real work, not the medium. Undersupervised struggle meant reaching the edge of their capacity with the expert out of the room: by working for one of the six to fifteen superstar robotic urologists on the planet, who run concurrent operations and complete six or seven prostatectomies where an excellent programme completes three; by rotating into departments that took up the robot later than urology, whose attendings could not tell good work from bad; or by rotating to hospitals under less cost and liability pressure. Help was always nearby and was called for; the residents were emphatic that they did not go rogue.

Beane is careful about what this is worth. The practices were neither punished nor endorsed but tolerated in silence, which is why the profession never learned from them. They cost the shadow learners the generalist education they would need in the jobs most of them would take. And they were rivalrous: the console time one resident won was time another lost, a Matthew effect running alongside a shrinking supply of competent surgeons and rising demand. The book rounds the proportion who found their way through to about one in eight, a figure to treat gently, since the interview study was built to over-sample rare successes; in a year at the five hospitals he was living in, he found three.

The book shows the same behaviour outside surgery. Inés, who fled Honduras as a teenager, had an eighth-grade education and no English, and worked a robotic kitting line whose error codes appeared in English. She taught herself what every one of them meant by watching, letting the robot sit a few seconds before she called for help, and coming back early from breaks to watch the engineers; her proposal to replace the codes with pictures was implemented within a week and credited to her line lead. An operations manager summed the phenomenon up for Beane: "Talent flows through this building like water." His historical case is Joseph Lister, who had no approved route to antisepsis either, applied carbolic acid to a patient whose leg the resulting gangrene cost him, and cut hospital mortality from 46 per cent in 1866 to 5 per cent by 1877.

Experts shadow-learn too, and Beane and Callen Anthony call their version the inverted apprenticeship: reassigning novices so that the expert can learn from them without losing face. Darren, an investment-banking director facing tools that turned two hours of analysis into ten seconds, borrowed two junior bankers from other groups, set them quality standards, met them daily and asked naive questions until he could run the reports himself.

Fewer than one in ten of the inverted apprenticeships in his data left both parties better off. That is the proportion that governs Beane's view of the whole phenomenon: worth studying, not worth institutionalising, and no one should have to risk a job or a patient to learn their trade.

Reworking the skill code

The counter-example is bomb disposal. Before robots, a technician walked to the device in a seventy-five-pound suit while the trainee sat a hundred metres back in an armoured truck, learning nothing: the technician was not talking, and there was one attempt at the problem. The PackBot, a forty-pound tracked robot built by the team that produced the Roomba, entered service in 2002, and the pattern that settled afterwards puts expert and trainee side by side in the truck, with the trainee holding the controller and the expert free to watch the situation and ask questions. Nothing in the machine required this; the controller could as easily have ended in the expert's hands. Somebody decided, and that decision is where the leverage sits.

From there he proposes three activities. Discover: find out how the three components actually appear and fail in one particular place, because the answer is local and nobody outside can supply it; and find the shadow learning, before surveillance drives it somewhere darker. Develop: change roles, metrics and structures. Deploy: insist, in the moment, that a technology deliver results and capability at once.

The tactics are specific. Realign roles: one investment-banking group kept junior analysts on the analysis rather than only the data pulls, which slowed the work and introduced errors, and which produced analysts who understood a whole valuation and the discovery that two standard software packages valued the same company billions of dollars apart. Rework metrics: when predictive policing sent officers to quiet places their arrest counts fell and they looked bad, until a few captains began praising a shift that produced nothing, on the grounds that nothing was the intended result. Tap frontline know-how: the scrub technicians who watch a procedure in its entirety know things the surgeon at the console cannot see, and residents already bypass the attending to ask them, so make the role a title, a rotation and a pay rise. Exercise surveillance restraint: turn off the operating room's screens while a resident works, so that only the mentor sees the mistakes, and accept that measurement stops paying where it destroys the unobserved spaces in which people try things. And deploy: the da Vinci's dual console can hand control back and forth, and Intuitive built a mode for sharing control simultaneously, the surgical equivalent of a driving instructor's pedals. Every institution in the study bought the second console. Beane never saw the shared mode used, his informants had never seen it, and the manufacturer's salespeople told him the practice was exceptionally rare; the residents called the second console the napping console. Against that he sets the chair of urology who told him he could not order his best surgeons to let residents operate, because they had patients and malpractice exposure and he had no right.

Skill's chimeric future

The last chapter is set in 2037 and follows a welder in Erie, Pennsylvania who takes a job requiring ultrasonic plastic welding she has never done. A platform recommends it as reachable, runs her through ten hours of graded practice in an augmented-reality rig, and matches her with an expert in Indonesia whose avatar and tools appear on her bench while novices, a school student and a retiree watch. Beane insists this needs no invention, only cost reductions in things that shipped before he wrote. His name for such arrangements is chimeric: neither human nor machine, and better than either alone. His example is chess, where Deep Blue beat Kasparov in 1997 and was then beaten for a while by a human working with an engine. His complaint against the chimeras that already exist, in turbine maintenance, chip layout, tumour resection and crop harvesting, is that they raise output and develop nobody. He is building one that does not: with Juho Kim he has made a system that indexes surgical video by procedure phase, trained on labels its own users produce, so that a resident who would have spent twenty hours hunting for a usable clip of one difficult step finds it in under a minute, and so that an attending can set video assignments and read what the resident noticed.

Against this he sets the prediction he says his discipline actually supports: that we keep taking the productivity and keep cutting the bond, that the few who find a way through race ahead, and that the result is skill inequality. He has seen its lower end. Warehouse work is engineered to reduce what the industry calls skilled touches, because every human contact with an item is a chance of error and more skill per contact means more error; the longer a worker stays, the less skill they can expect to have.

On generative AI he is direct and restrained. It has been adopted faster than any automating technology in history, and it is why the reskilling number runs to billions. He expects prompting to stop being a skill worth having, because systems will soon coach the user toward a better request with no human in the loop, and he points at Khan Academy's tutor as coaching that already comes from the system. That does not solve his problem: such a system supplies challenge and some complexity, and no human connection at all. His closing position is that the technologies causing the damage are also the only things that could carry the bond at the scale required, and that whether they do is a matter of choices made by the people who design, buy and deploy them, not of what the machines can do.

Since 2024

The argument has aged into a live question rather than a settled one. The clearest support arrived from labour markets: in 2025 Erik Brynjolfsson, Bharat Chandar and Ruyu Chen reported that employment for workers aged twenty-two to twenty-five had fallen relative to older workers in precisely the occupations most exposed to generative AI, which is what the mechanism predicts, since the novice's work is the first thing a capable machine absorbs. The strongest challenge came from a study published before the book and not addressed in it: Brynjolfsson, Danielle Li and Lindsey Raymond, tracking five thousand customer-support agents given a generative assistant, found the gains concentrated among the least experienced, because the system diffused the practice of the best agents to everyone else. That is a machine doing some of what Beane says only the bond can do. The open question is the one his account makes sharp: such studies measure output while the assistant is present, not what the novice can do when it is taken away.

Two of his other supports have moved. The estimate of three to six jobs lost per robot, which comes from Daron Acemoglu and Pascual Restrepo, has been contested by work finding that automating firms grow and hire, a result Beane himself reports elsewhere in the book. And the surgical case has changed under him: professional bodies and residency programmes have built the robotic training tracks, case minimums and simulator requirements whose absence his paper documented, which addresses the credential and leaves the console-time problem where he found it, since the constraint was never the training requirement but the attending's judgment about the patient in front of him. What has held is the mechanism and the diagnosis. Productivity and learning are made to trade against each other by the way tools are deployed, not by what the tools are; the trade is usually made in the short run's favour without anyone deciding to make it; and the people who learn anyway do it by breaking rules that nobody is willing to change.