Formation, Not Training: Educating for the Judgment Machines Can't Supply

In 2025, a manager handed a new hire named David a real problem: "Figure out what we should optimize for" in a new AI customer-service system. David had a 3.9 GPA, five AP classes, and sixteen years of practice identifying correct answers. He waited for a clearer specification. None came. So he defaulted to what he could measure: response time.

Three weeks later, response time dropped sharply and everything looked beautiful. Then customers started describing the system as "cold," "dismissive," "like talking to a machine that doesn't care." Brand trust fell. Churn accelerated. David was baffled — the system did exactly what he'd specified. His education had taught him to execute instructions flawlessly. It had never once taught him to question whether those instructions captured what mattered.

David isn't a failure. The system that certified his competence as success failed him. And it's about to fail millions more.

The tension: an economy that runs on judgment, trained out of people

Here is the contradiction at the center of the AI transition. MIT researchers studying more than 300 enterprise AI deployments found roughly 95% of enterprise GenAI initiatives deliver no measurable P&L impact. The agents access humanity's knowledge instantly and execute procedures flawlessly. What the organizations deploying them can't do is tell the agents what to pursue.

Most educated adults were never taught how to define a worthy objective. Not because someone forgot to add it to the syllabus, but because the system was never designed for it. The classroom most of us passed through is, in its bones, the 1890 Prussian model — fifty-minute periods, age-based cohorts, recall-based tests. It was engineered to produce a population that could follow instructions, absorb standardized information, and execute defined tasks in an industrial economy. It didn't fail at that mission. It succeeded brilliantly. The mission is now obsolete.

Derek Thompson, writing in 2025 about what he calls a "quiet apocalypse" of cognitive de-skilling, argues that outsourcing writing to AI is outsourcing thinking — the two are inseparable acts. We are deploying more capable AI into a workforce measurably less capable of directing it, because we trained the directing capacity out of people.

The reframe: develop who someone becomes, not just what they do

The distinction that matters is between training and formation.

Figure: The four pillars of formation — systems thinking, intent design, taste formation, and AI fluency — form a single system with intent design as the keystone, cultivating the judgment machines can't supply.

Training equips skills for defined tasks — what you do. Learning Python becomes obsolete in five years. Formation develops judgment alongside intellect — who you become. A carpenter's apprentice in 1850 spent seven years learning not how to plane wood but how to recognize grain that will split, whether a joint sits true, whether a request produces furniture that endures. The apprentice learned judgment, not procedures — and learned it by working alongside someone whose judgment they could study, not by reading about judgment and then demonstrating recall on a test.

The mechanism is precise: a trained engineer implements a specification efficiently. A formed engineer looks at the same specification and asks whether it should exist — then has the judgment to argue for a better one. That gap doesn't surface in performance reviews. It surfaces when organizations deploy AI against the wrong objectives and can't diagnose why the metrics look beautiful while the underlying value erodes.

Contrast David with Sofia, who graduated from a formation-focused school. Joining a healthcare AI startup, she faced a diagnostic-uncertainty problem — flag only high-confidence findings, or surface tentative ones? She didn't have an answer. She had a process. She interviewed oncologists, emergency physicians, and patients, surfaced their conflicting risk tolerances, and proposed making confidence thresholds explicit and adjustable by context. Don't hide the trade-off — make it governable. David optimized a system. Sofia questioned whether it should exist. One capacity was trained out of him over sixteen years; the other was trained into her over four.

How it works: four pillars, one system

Formation isn't a vague humanistic aspiration. It has structure. Four capacities, which are not independent competencies but a single system:

• Systems thinking — the cognitive architecture to see what you're operating inside, where optimizing one metric degrades another and today's solution seeds tomorrow's crisis.
• Intent design — the keystone. The practice of defining a worthy objective and noticing when a proxy goal diverges from the purpose it was meant to serve. "Maximize engagement" collapsing into outrage amplification is intent design failure at civilizational scale.
• Taste formation — the evaluative judgment that distinguishes good intent from performative intent, design that serves real needs from design that optimizes a proxy. When AI collapses execution costs and floods markets with competent mediocrity, taste is the moat.
• AI fluency — ensuring that judgment can be applied in AI-mediated environments, including the discipline to recognize confident hallucination and the refusal to offload difficult thinking to the model.

Remove intent design and the other three have no north star. These four cultivate the Architect, Guardian, and Force Multiplier roles of the The New Triumvirate — and they require sustained practice, failure, and refinement, which is exactly what fifty-minute recall-based periods cannot provide. The employer evidence converges here: in an August 2025 survey of 1,030 executives and hiring managers, more than 9 in 10 rated human judgment, critical thinking, and ethical reasoning as important or very important — figures that held or rose as AI adoption accelerated. Formation isn't a luxury. It's increasingly the thing employers cannot find.

What to do

The barrier to formation education isn't curriculum or technology. It's teacher capacity — and teacher capacity is a funding-allocation choice, which means it's actionable now.

1. Redirect existing spend toward teacher development. U.S. teachers average 24 hours of professional development annually. Singapore entitles teachers to 100; Finland requires a master's with 1,600+ hours of pedagogical preparation. The U.S. already spends above the OECD average per student — the money exists, but it flows into administration and facilities, not formation. This is the single highest-leverage intervention available, and it requires a school-board budget decision, not new federal legislation.
2. Separate content delivery from formation. Khan Academy's Khanmigo is designed never to give the answer — it asks the next question. Alpha School runs AI-driven content acquisition for two hours so humans can own the other four to five hours of life skills. Let AI handle what standardized testing rewards; preserve the friction of inquiry for humans.
3. Measure what's becoming, not just what's scored. Formation assessment means portfolios over years, process documentation, authentic performance tasks. More expensive than filling in bubbles, and the point.
4. Guard against the equity trap. Alpha School is private and serves a high-income demographic. If AI-enabled formation becomes a premium for the already-advantaged, it widens the gap it claims to close. Execution through public systems is the unsolved challenge that matters most.

A word of honesty about timing. Formation properly implemented takes a decade-plus to reach scale; the Prussian system itself took 30 years. The displacement window for vanguard sectors is 1–3 years. These timelines don't overlap, which is why formation built now serves children in school today — not the 47-year-old whose synthesis work agents have already commoditized. For those workers, community infrastructure becomes the adult formation tier: a newly displaced person joining a cooperative governance meeting develops intent design through practice, not coursework. Three tiers, three timelines — Bridge stabilizes in months, community forms in years, schools transform the baseline in decades.

What education encodes about human purpose — execute, or discern — shapes the consciousness of every person who then specifies the next generation of autonomous systems. Get the encoding wrong, and the error propagates faster than it can be corrected. That's why formation is the most consequential values-in-architecture decision a society makes.