Skill Capture

Chapter 1: The Skill Premium and Why It's Disappearing

In the fall of 2022, a radiologist in Portland, Oregon—let's call her Dr. Sarah Chen—sat in a hospital conference room and watched her profession disappear on a screen.

She wasn't watching a film. She was watching an AI system read a chest X-ray faster than she could think. The scan appeared. A fraction of a second later, colored boxes materialized over regions of concern—tumors, nodules, areas that warranted attention. The system's confidence scores appeared alongside: 97% certainty for the mass in the left upper lobe. 89% for the pleural effusion. The speed wasn't just impressive. It was humbling.

"I spent fifteen years learning to see what takes this thing milliseconds," she said later, not bitterly—she was too measured for bitterness—but with a kind of forensic precision that made the observation more unsettling than any complaint could have been.

Dr. Chen's moment is not unusual anymore. It is, increasingly, the defining experience of skilled work in the early 2020s. A security researcher watches an AI system find vulnerabilities in code he'd spent a decade learning to recognize by hand. A commercial illustrator sees a text prompt generate in seconds what would have taken him three days of work. A tax attorney realizes that legal research—the bottleneck that justified her $400-per-hour rate—can now be crowdsourced to a language model that costs less per query than a cup of coffee.

What's happening to these practitioners is not new in the broad strokes. Automation has been displacing labor for two centuries. But there is something categorically different about what's occurring now, and it lives in that single word: skilled.

For roughly 150 years, from the mid-1800s through the early 2020s, there existed a relatively stable economic bargain: specialization was protected by natural scarcity. If you wanted to be good at something—genuinely expert, the kind of practitioner people paid premium rates to access—you had to spend years learning it. There were no shortcuts. A radiologist couldn't become competent in months. A commercial illustrator couldn't skip the decade of figure drawing and color theory and client management. A cybersecurity researcher couldn't compress the trial-and-error of learning to think like an attacker.

This wasn't a flaw in the system. It was the system's entire logic.

You invested time—years, often a decade or more. You accepted lower income during apprenticeship or training. You endured the frustration of incompetence, the slow accumulation of tacit knowledge that no textbook could transmit directly to your brain. And then, once you emerged from that crucible as a genuine expert, you were protected by an invisible moat: nobody else could easily do what you did. That scarcity justified premium income. It justified credentials. It justified the entire infrastructure of professional licensing, union apprenticeships, and educational gatekeeping that characterized the 20th-century knowledge economy.

An experienced radiologist could command $250,000 to $400,000 annually because there were only so many people willing to spend the time to become radiologists, and hospitals needed radiologists. A master carpenter could charge $150 per hour because mastering carpentry required years of accumulated tacit knowledge that couldn't be compressed. A security researcher who discovered a zero-day vulnerability could command consulting fees in the six figures because that capability—finding novel attack surfaces in complex systems—required a rare combination of technical depth, creativity, and pattern-recognition that money alone couldn't quickly purchase.

The scarcity was real. It was the economic foundation upon which millions of practitioners built their livelihoods, their professional identities, and their life plans.

What AI is doing—what it is currently doing, right now, in real time—is systematically demolishing that scarcity by treating human mastery as a training data problem.

Consider the basic mechanics: A company like Tempus AI gathers hundreds of thousands of pathology slides and medical imaging scans—the accumulated clinical work of thousands of radiologists and pathologists, collected without compensation or often even explicit consent—and uses those images to train a neural network. The network doesn't understand radiology the way a human does. It doesn't develop intuition through years of mentorship or learn the subtle tells that experienced practitioners develop. Instead, it learns statistical patterns: the pixel distributions associated with malignancy, the texture signatures of disease, the geometric markers that correlate with poor outcomes.

And here's the disorienting part: it works. Often it works better than the humans it was trained on.

A 2019 study in JAMA found that an AI system trained on breast cancer imaging data outperformed eleven radiologists in both sensitivity and specificity for detecting cancer in mammograms. Another system, developed by DeepMind and Google Health and tested on chest X-rays, outperformed four radiologists in detecting tuberculosis. The pattern repeated across domain after domain: radiology, pathology, dermatology, ophthalmology. In each case, AI systems trained on human expertise achieved or exceeded human-level performance in a fraction of the time it took to develop that expertise in the first place.

But the math goes beyond matching performance. The AI system, once trained, costs essentially nothing to run. A single API call might cost $0.001. Scale that across millions of images, millions of consultations, and the cost per expert decision collapses toward zero. Meanwhile, a radiologist—who generates, over a 30-year career, something like 8 million individual diagnostic decisions—commands a six-figure salary, benefits, malpractice insurance, continuing education costs. The economic comparison isn't close.

This is where "skill capture" differs fundamentally from traditional automation.

When factories automated manufacturing in the Industrial Revolution, they replaced routine labor—the repetitive work of assembly-line workers, typists, switchboard operators. These were jobs that required little training and offered little premium to the most talented workers. The best assembly-line worker earned perhaps 20% more than the worst. When machines took over, they displaced workers at the bottom of the skill hierarchy, but they didn't collapse the economic value of expertise because expertise wasn't what was being automated. A skilled machinist who could diagnose and repair complex equipment remained valuable. The routine work went away; the specialized knowledge remained protected by scarcity.

AI doesn't work that way. It targets expertise directly. It doesn't just automate the routine parts of radiology; it automates the diagnostic capability that took a radiologist 15 years to develop. It doesn't just handle the boilerplate legal research; it handles the strategic analysis and novel interpretation that justified $400-per-hour rates. It doesn't automate the mechanical parts of illustration; it automates the stylistic mastery that made a commercial artist's work recognizable and worth paying premium rates for.

The mechanism is straightforward and almost banal in its simplicity: treat the accumulated expertise of thousands of human practitioners as training data, train a model on that data, then use that model to eliminate the economic scarcity that made those practitioners valuable in the first place.

And crucially—and this is what makes it a form of extraction rather than simple disruption—the humans whose expertise provided the training data are typically not compensated. They don't own their expertise. They can't prevent its replication. A radiologist spends fifteen years learning her craft, builds a reputation, achieves mastery. A health tech company scrapes her hospital's DICOM servers (or licenses them, or buys an imaging dataset that includes thousands of her interpretations), uses her work to train a model, then markets that model directly back to her employer, undercutting the economic rationale for employing her at all.

The radiologist has been skill-captured: her expertise has been extracted, packaged, democratized, and turned into a commodity. And she has been paid nothing for the contribution.

This pattern is now repeating across knowledge work with remarkable consistency. Consider cybersecurity research. A security researcher might spend five years learning to think like an attacker, studying the literature of offensive techniques, developing the intuition to spot unusual code patterns that signal vulnerabilities. That expertise—the ability to find zero-day vulnerabilities—is economically rare and therefore economically valuable. It justifies six-figure consulting fees and attracts talented people to the field.

Now, train an AI system on hundreds of thousands of published vulnerabilities, exploit code, security research papers, and code repositories. The system learns to recognize the patterns that correlate with exploitability. Researchers at companies like Anthropic and DeepMind have already demonstrated that large language models fine-tuned on security data can find previously unknown vulnerabilities in code. The economic implication is clear: the scarcity that made security research valuable is being artificially demolished by AI systems trained on the collective work of the security research community.

Or consider commercial illustration. An illustrator spends years developing a distinctive visual style, learning composition, studying the work of masters, building a portfolio that demonstrates consistent quality and aesthetic coherence. That style—the particular way she renders light, or how she approaches proportion, or her characteristic color palette—becomes recognizable. Clients commission work from her because they want her style. That premium on style is the core of a commercial illustrator's pricing power.

Now take millions of illustration portfolios, art books, design work, and the vast archive of digital imagery available online. Train a diffusion model on that data. The model learns the statistical patterns that define "illustration in the style of [famous illustrator]" or "commercial art rendered as [aesthetic]." A client can now generate in seconds—for the cost of a few API calls—something that approximates or sometimes exceeds what it would take the human illustrator days to produce. The scarcity of the illustrator's style has been demolished.

The timeline of skill capture is what makes it particularly destabilizing. A radiologist can retrain into a different field. A security researcher can pivot. An illustrator can shift to fine art or teaching. But retraining takes time—years, typically—and assumes markets exist for the retrained skills. Skill capture happens in months. An AI system trained on radiology data doesn't take fifteen years to develop competence. It doesn't apprentice or study or gradually accumulate tacit knowledge. It learns the statistical patterns in the time it takes to run a training job, then immediately scales to process millions of images per day.

The speed is disorienting to those experiencing it.

"I trained for a decade to do something that now seems like it was obvious," an ophthalmologist said in an interview about AI systems that now diagnose diabetic retinopathy from retinal photographs. He didn't mean it was obvious to him—he was being ironic. He meant that once the AI system was built, the diagnosis became obvious, codified into weights and biases that could be copied infinitely. The decade he spent learning wasn't obvious while he was learning it. But in retrospect, after the AI system was deployed, it looked inevitable. It looked like something that didn't actually require a human being to spend a decade learning.

This is one of skill capture's most unsettling properties: it makes the original investment in expertise look retroactively absurd.

The economic logic of this transformation is worth sketching clearly, because it underpins everything that follows in this book. For 150 years, the economic system operated on a principle of protected scarcity: expertise was valuable because it was rare, and it was rare because it required time to develop. The value proposition for specialization looked like this:

Invest 10 years → Become one of 10,000 practitioners with this skill in your country → Command premium income for 30+ years → Retire with security and status.

That equation depended absolutely on step two: the fact that there could only ever be so many radiologists or security researchers or master illustrators because there were only so many people willing to invest a decade in developing that skill. The scarcity was a feature, not a bug. It was what made the whole system work.

Skill capture inverts this equation. Now the logic is:

Extract expertise from 10,000 existing practitioners → Train an AI system on that extracted expertise → Deploy the AI system at scale → Collapse the economic value of the original expertise → Force practitioners into a commodity market competing with an AI system that costs $20/month.

The first equation incentivized specialization. People would spend a decade learning radiology because they knew that decade would create a protective moat around their earning potential. The second equation doesn't just fail to incentivize specialization; it actively punishes it. Why spend a decade learning radiology if the economic premium for that decade of learning evaporates overnight?

This is not hyperbole or speculation. This is, again, happening in real time.

Hospitals are already restructuring radiology departments. They're not eliminating radiologists entirely—not yet—but they're using AI as a tool to increase the number of scans a single radiologist can oversee, or to replace the highest-paid radiologists with lower-paid technicians who review AI flagged cases. A radiologist in California reported a 30% pay cut when her hospital switched to an AI-primary workflow where she reviewed AI findings rather than generating primary interpretations. The job title remained the same. The skill level required decreased. The compensation reflected that decrease.

Security firms are restructuring research divisions. They're not eliminating human researchers, but they're using AI to democratize vulnerability research, which means they can hire less experienced (and thus cheaper) security researchers and augment them with AI, rather than paying six figures for veteran researchers with deep domain expertise. The skill premium for security research is already compressing.

Commercial illustration rates have been visibly falling for two years as generative AI systems produce increasingly competent illustrations. Illustrators report not just lower rates but lower volume—clients are simply generating illustrations themselves. The economic middle has collapsed: you're either a premium "human-created art" specialist commanding high rates, or you're competing directly with AI-generated imagery and losing on cost.

These aren't predictions. These are observable facts, documented by the practitioners experiencing them, visible in market data, measurable in real-time labor markets.

What's happening is an unprecedented form of economic extraction. It combines two mechanisms that were previously separate:

First, it treats human expertise as a commons. Radiologists spend a decade developing diagnostic skill by working at hospitals, studying imaging archives, learning from senior radiologists. That knowledge is accumulated through systems funded by the public, by hospitals, by the medical establishment. Yet that accumulated expertise—stored in the radiologists' brains, in their decision-making patterns, in the probability distributions of disease they've internalized—becomes the training data for AI systems built by venture-capital-backed startups who claim to "augment human expertise" while systematically undercutting the economic value of that expertise.

Second, it monetizes the externality. The value of that extracted expertise isn't returned to the practitioners. Instead, it flows to the companies that built the AI systems, to the venture capitalists who funded them, and to the corporations that deploy them. A radiologist loses income. An AI startup becomes a billion-dollar company by selling access to the radiologist's accumulated expertise.

This isn't a malfunction of the market. It's an extremely efficient exploitation of property rights gaps: expertise doesn't have the same legal ownership status as physical property, so there's no mechanism for radiologists to be compensated when their expertise is extracted and replicated at scale.

The consequences ripple outward in ways that aren't yet fully visible but are already becoming measurable. If expertise no longer generates premium income—if a decade of learning radiology doesn't create a protective moat around your earning potential—then what exactly is the rational incentive for someone to spend a decade learning radiology? What's the economic case for investment in specialization?

This question is not academic. It goes to the foundation of how advanced economies have developed for the past 150 years. The bargain that motivated millions of people to pursue specialized expertise was never purely intrinsic. Yes, some people love radiology or illustration or security research for its own sake. But many people choose specialization because specialization pays. The premium income justifies the years of study, the opportunity cost, the deferred gratification of early career years spent in training.

When that premium disappears, the entire bargain collapses.

Dr. Chen, the radiologist from Portland, had advised her sister—a recent undergraduate considering graduate school—to pursue something other than medicine. "The economics don't make sense anymore," she'd said. "If you're going to spend your twenties and thirties in training, it should be for something that still values expertise when you finish training. Medicine doesn't anymore." Her sister took the advice. She went to business school instead.

That conversation is being repeated, in various forms, across knowledge work. If expertise no longer generates premium income, if a decade of learning can be replicated by an AI system trained in months, if a master's degree in radiology doesn't create the protective scarcity it once did, then rational actors will optimize toward fields that still reward specialization—or they'll pursue entirely different life plans.

The consequence, over time, will be straightforward: the supply of newly trained specialists will decline as the economic incentive to specialize erodes. We may not immediately notice this because universities will continue admitting radiology residents and security researchers and illustration students—institutional inertia is powerful. But over a 10-to-15-year window, as the first cohorts of newly trained specialists emerge into a market where their specialty has been commodified, the realization will spread: spending a decade becoming excellent at something no longer protects you economically.

What we're witnessing is not the replacement of workers by machines. It's the wholesale extraction and commodification of the economic value of expertise itself. And it's happening fast enough that the practitioners experiencing it are still alive, still watching their field transform in real time, still grappling with the realization that the bargain they made—invest now, reap the reward later—has been unilaterally revoked.