Taylor Had a Stopwatch. They Have Your Knowledge Work.

The April unemployment rate is 4.3%. That number is no longer telling you what you think it's telling you.

Headline unemployment measures people actively looking for work who can't find it. It doesn't measure people who have stopped looking. It doesn't reach the information sector worker who left after their third layoff in four years and is now doing contract work while they wait to see what stabilizes. The Bureau of Labor Statistics publishes a broader measure, U-6, which tries to capture some of this. It's running at 8.2%. Labor force participation is at 61.8%, meaning more than a third of working-age Americans are outside the measured economy entirely. The headline number looks stable because the denominator is shrinking.

The information sector has shed 342,000 jobs and is running 11% below its 2022 peak. That contraction has held for 42 months. The duration is what matters. Cyclical downturns correct. Structural shifts don't, not quickly, and not back to where they started. A separate contraction is running through the federal workforce simultaneously, down 348,000 and 11.5% off peak, driven by a political project whose downstream effects on regional wages and labor markets will take years to fully surface.

Where the jobs are growing: eldercare, home health, last-mile delivery, warehouse and retail, family services. The work that resists capture is the work that remains. The work that can be abstracted, compressed, and served back as a model is the work that is leaving.

This is happening during a period when AI is supposed to be lifting all boats. Productivity gains flowing broadly, a larger pie for everyone. So the natural question is: what is actually going on?

The Talent Redistribution Nobody's Naming

The Linux Foundation's 2026 hiring report shows net hiring up 31% across surveyed firms, and AI-specific roles up 60%. Some commentators have used this to argue that the displacement fears are overblown. That reading requires not looking too closely at the distribution.

Only firms over 20,000 employees are showing a negative net effect. The mid-market is hiring. What that describes is not a healthy labor market. It's a transfer. The engineers and technical workers being cut from hyperscalers and large enterprises are being reabsorbed, some of them, into smaller firms with less compensation, less stability, and less leverage. The headline number goes up. The conditions inside it get worse.

The capability gaps tell the same story from a different angle. Security, operations, FinOps, platform engineering, shortages running at 35 to 57% across surveyed organizations. These are not entry-level roles. They require years of accumulated expertise. The prescription the report lands on is upskilling, invest in the people you have, retrain toward the gaps. Organizations are more than three times more likely to upskill existing staff than to hire new people.

Read that again. The pipeline that historically grew new practitioners is being compressed at exactly the moment the official prescription is to develop the people already inside the organization. The workers being displaced from higher-paying roles are being told the path forward is retraining. The workers who remain are being asked to absorb the gaps left behind. Neither group had much say in how the system was designed.

This is redistribution. It looks like resilience from a distance. Up close it looks like a transfer of risk and cost onto the people with the least power to refuse it.

What You're Actually Buying When You Buy AI

Which raises the question the hiring data doesn't ask. What is actually being bought and sold when a company adopts AI? The answer is older and stranger than most people realize.

Every AI system is trained on patterns extracted from people who did the work.

Not abstractions of work. The actual work. Code written by developers solving real problems under deadline. Writing produced by journalists, essayists, and analysts working out ideas in public. Design decisions made by people who understood the constraints. Diagnostic reasoning developed by clinicians over careers. Customer service conversations where someone with patience and pattern recognition talked a frustrated person through a problem.

The model is a compressed, recomposed version of that cooperative human practice. When your company licenses access to it, you are renting back a version of expertise that workers, somewhere, originally produced. The knowledge was collective. The ownership is entirely private.

Three examples make this concrete.

GitHub Copilot was trained on public repositories, millions of them, representing decades of accumulated developer problem-solving. The autocomplete suggestions it offers are not invented. They are distilled from the actual thinking of actual engineers who pushed their code to GitHub, most of them without any expectation that it would become training material for a commercial product.

Image generation models were trained on artists' portfolios, stock photo libraries, and creative work scraped from across the internet. The aesthetic decisions, the compositional knowledge, the stylistic range these models can reproduce, came from people who spent years developing it. Most of them were not asked.

Customer service models were trained on transcripts of human support workers, the patience required to stay present through a difficult call, the pattern recognition that knows when someone is actually frustrated versus when they are just moving through a script, the accumulated institutional knowledge of how problems actually resolve. That expertise is now the product.

This pattern is not new. Charles Babbage walked the factory floors of industrial England in the 1830s, studying how skilled craftsmen worked, identifying which movements could be broken down, systematized, and eventually mechanized. Frederick Taylor did the same thing seventy years later with a stopwatch and a clipboard, decomposing skilled labor into measurable units that could be optimized, supervised, and eventually replaced. What changes with AI is not the structure of the operation. It is the comprehensiveness of what can be captured and the speed at which it can be recomposed into something ownable.

Participating in AI adoption is participating in a large-scale knowledge extraction operation. That is not necessarily an argument against it. It is an argument against pretending otherwise. The knowledge that powers these systems came from somewhere. It came from people. Those people are, in most cases, not compensated for its use, not credited for its origin, and not consulted about how it gets deployed.

A small number of products are beginning to experiment with attribution models that name the source of knowledge and route value back toward it. That is not the norm. But it is worth noting that the loop does not have to stay closed in the direction it currently runs.

The Valuations Don't Add Up

The valuation question is where the extraction logic becomes hardest to ignore.

Ricardo Hausmann and Andrés Velasco published an analysis in April 2026 working through what current AI firm valuations actually imply. Their calculation: the combined market capitalizations of the leading AI firms price in roughly $2.4 trillion in additional annual foreign revenue by 2036. For context, total current US goods exports run around $2 trillion annually. The valuations are pricing in a second US export economy, built entirely on AI services, on top of the one that already exists.

Their calculation prices in specific conditions. It requires that AI services generate export revenue at a scale comparable to everything the United States currently sells to the rest of the world combined. It requires that this happen inside a decade. And it requires that it happen during a period of intensifying trade friction, rising protectionism, and active political pressure to deglobalize the supply chains that AI infrastructure depends on.

These two things are not compatible. You cannot simultaneously pursue the protectionist trade posture currently dominant in US politics and realize the globally integrated AI revenue base those valuations assume. Either the valuations are wrong, which is the bubble case, or they are pricing in a sustained ability to extract rents along the AI value chain regardless of trade friction, which requires a degree of geopolitical leverage that has not yet been demonstrated and may not be achievable.

Martin Sandbu's value chain framing is useful here. The rents from AI don't accumulate evenly. They concentrate at specific points: chip design and fabrication, cloud infrastructure, foundation model development, application layer distribution. Different firms are positioned differently along that chain, and the political economy of each position is different. The valuation question is not academic. It determines who gets capital, who gets political protection, and who gets disciplined when the numbers don't work out.

What it points toward is a system under considerable internal strain. The economics assume global reach. The politics are pulling toward fragmentation. Something has to give, and the people least likely to be consulted about which thing are the ones whose labor built the systems in the first place.

The Trap Everyone's Pension Is In

There is a reason nobody in a position to act on any of this seems able to.

Matt Stoller has been tracking this for years, and the mechanism he describes is straightforward and brutal. Sixty percent of Americans own stocks, directly or through retirement accounts. Every major institution with long-term obligations, pension funds, university endowments, state budgets, hospital systems, large nonprofits, is tethered to equity prices. The Magnificent Seven alone represent a structurally significant share of the S&P 500. When those stocks go up, retirements get funded, endowments stay solvent, and state budgets have room to breathe. When they go down, the pain is immediate and widely distributed.

This creates a specific political trap. Regulating the extraction meaningfully, taxing AI profits aggressively, breaking up platform concentration, requiring attribution and compensation for training data, any of these interventions risks crashing the assets that middle-class retirement security depends on. The people most positioned to push back are the people whose security depends on the extraction continuing. That is not a coincidence. It is the architecture.

The evidence is already visible. The Full Stack AI Export Promotion Act passed a Senate committee 37 to 7. A Democratic governor vetoed a data center moratorium. In Virginia, public support for new data center approvals collapsed from 69% to 37% over eighteen months, while the political class continued approving them. The gap between what voters say they want and what legislators do is not mysterious. It reflects who absorbs the cost of the alternative.

This is not a story about bad actors. The pension fund manager is doing their job. The governor is weighing real tradeoffs. The senator is representing constituents who need their retirements to hold. The system has structured itself so that the most available levers for accountability are also the most politically costly to pull.

Progressive tech leaders have been waiting for Washington to act on this. That wait is going to be long. The structure of asset ownership has made meaningful regulation a political near-impossibility, not because politicians lack courage, but because the system has distributed the cost of inaction in a way that makes action look like the dangerous choice. You are not going to be rescued from outside your own organization. The question is what you do inside it.

The Full Circuit

The last five sections describe separate phenomena: a labor market moving in ways the headline number doesn't capture, a hiring report that looks like growth but describes redistribution, a knowledge extraction operation embedded in every AI product, a valuation structure that strains credibility, a political system that cannot act on any of it.

They are not separate.

The extraction produces the products. The products produce the revenue. The revenue produces the valuations. The valuations anchor the financial system. The financial system's dependence on those valuations makes regulating the extraction politically incoherent. And the same workers whose practice was converted into products are, in many cases, holding those valuations through their retirement accounts. The thing that was done to them is also the thing they now depend on.

You do not need to name this to recognize it. Most people reading this have been inside some version of it, either building the extraction mechanism or absorbing its effects or both.

This pattern did not originate with AI. The specific mechanism (converting what people know through practice into something ownable, scalable, and tradeable) is older than the internet, older than computing, older than the industrial concerns the earlier sections gestured toward. What changes with each new vehicle is the scope of what can be captured and the speed at which it can be productized. AI is the most efficient vehicle so far. It will not be the last one.

Betting that this resolves when AI matures, or when the regulatory environment changes, or when the political balance shifts, requires believing that the pattern ends with its current vehicle. There is no particular reason to believe that.

What You Can Actually Do

The structural trap does not dissolve at the organizational level. But the decisions made inside your organization are not neutral, and some of them are within your control.

Start with your extraction profile. Every product built on AI is built on training data that came from somewhere. Most organizations have not mapped where. Which labor community produced the knowledge your product depends on? Are you building anything that routes value back toward them? This does not have to be philanthropy. There are commercial models that compensate contributors directly, name sources, and build the attribution into the business rather than treating it as optional. Some of the more defensible positions in the current market belong to companies that made that choice before it was required of them.

The capability gaps in the Linux Foundation data point toward something more immediate. The shortage in security, platform engineering, and operations is not going to be resolved by the next hiring cycle. The firms that close those gaps through deliberate development of the people they already have will outcompete the firms that treat their workforce as a variable cost and buy API credits instead. The phrase "do more with fewer people" usually describes a transfer rather than an efficiency gain. The efficiency is real. But the surplus goes somewhere, and it is rarely the remaining workers or the customers. Being honest about that, inside your own organization, changes what decisions you make next.

Any forecast that assumes frictionless global AI adoption is exposed to the Hausmann-Velasco contradiction: the valuations price in conditions that the current political environment actively undermines. Any plan that assumes stable US-centered infrastructure is carrying political risk that most financial models have not accounted for. When did you last model a scenario where your primary AI vendor's pricing or availability changes significantly? Or where the regulatory environment in a key market shifts? Customer concentration, supply chain exposure, regulatory scenarios: these are not hypothetical. They surface on balance sheets eventually, usually when there is least time to respond.

None of these requires waiting for Washington. None requires that the system change before you do.

The discomfort you feel looking at this is not aesthetic. It is not a political opinion or a brand position or a personal values statement. It is the correct response to a structural situation that most of the frameworks available to you were not designed to describe clearly.

You do not need a 19th century vocabulary to act on what you are seeing. You need to know where the value is going, who is producing it, and who is capturing it. Those are not abstract questions. They have answers inside your own organization, answers that most leaders have not gone looking for because looking has not felt urgent enough.

The decisions you make inside your organization are not neutral. They were not neutral before AI and they are not neutral now. The thing that makes the loop stronger is not malice. It is the accumulated weight of choices made by people who had the information to see what they were doing and declined to look at it directly. Pretending your decisions are neutral is participation.

The choice in front of you is not between AI being good and AI being bad. It is whether you participate in shaping what gets built, who benefits from it, and what gets left behind, or whether you let that get decided without you while you focus on the parts that feel manageable.

The people doing the work inside your organization are watching how you handle this. They have their own views on where the value is going and who is capturing it. They are probably more right than wrong. The question is whether they work for someone who can see what they see and act with some honesty about it.

—

PS: Tried Humanize.us yet? You’ll be glad you did.