Epistemic Agency in the Age of AI

---

Marcus is a 34-year-old marketing director. He's articulate, fast, and well-regarded by his team. On any given Tuesday, he produces eight to twelve pieces of substantive work — emails, strategy memos, creative briefs, competitive analyses. His manager considers him one of the sharpest people in the department.

In a recent all-hands meeting, a consultant asked Marcus to walk through the reasoning behind a positioning strategy he'd submitted the week before. Marcus talked for about ninety seconds, fluently, and then stopped. He hadn't run out of time. He'd run out of reasoning. The document was sophisticated. The thinking underneath it — the comparison of options, the weighing of tradeoffs, the stress-testing of the core premise — had been done by an AI system, not by Marcus. Marcus had read the output, adjusted the wording, and approved the structure. He'd never built it.

What Marcus experienced in that moment wasn't stupidity. It wasn't laziness. It was the quiet exposure of something that had been eroding gradually, without a visible threshold or a single alarm — He no longer knew how to reconstruct the work he was producing. Nobody told Marcus this was happening. He didn't feel it happen. His output was excellent throughout. His performance reviews were strong. There was no signal that anything had changed, because the metrics that were tracked had improved.

What wasn't measured — what has no dashboard, no quarterly review metric, no notification — was what was happening to his capacity to think without scaffolding. That capacity had been quietly declining through a process so gradual and so frictionless that it felt, at every step, like progress.The first-order story is that AI makes Marcus more productive. 

The second-order story is what is happening to Marcus.This is not a technological critique. It is a forensic account of a formation environment changing faster than the institutions that depend on it. We have known this for decades in physical medicine. We have documented it in navigation, memory, mental arithmetic, and attention span. We are beginning to see it in judgment — and it is receiving less systematic attention than the pattern warrants.

Judgment is not having opinions. It is not intelligence. It is not confidence. Those things can exist in the complete absence of judgment.

Judgment is the capacity to construct coherence in the face of uncertainty. It is what happens in the interval between a question and an answer — the uncomfortable, metabolically expensive process of holding competing claims simultaneously, detecting when something is internally inconsistent, tolerating partial information long enough for a real conclusion to form, and knowing the difference between a position that holds and one that merely sounds coherent. 

It is built through friction. Not hardship for its own sake — formative friction. The specific cognitive strain of not knowing yet, of having to compare imperfect options, of being wrong and recognizing it, and adjusting. That strain is not an obstacle to the development of judgment. It is the mechanism of it. 

Consider what a first-year attorney learns that no law school class can teach: the experience of building an argument, submitting it to a partner, having it dismantled, rebuilding it under constraint, defending it in deposition, and discovering its structural limits. That cycle — construction, stress-test, failure, revision — is what produces a lawyer who can think in a courtroom rather than rely solely on a prepared brief. The substance of the law can be transmitted. The capacity to reason inside it has to be built through friction.

 Consider what a young surgeon learns in a residency, or a young pilot in instrument conditions, or a young journalist building a source. In every high-stakes formation environment we take seriously, we understand that competence cannot be transmitted solely through output. It has to be constructed through repeated exposure to uncertainty, consequence, and correction.

This is what makes the current moment structurally different from every previous technological shift. Calculators changed how we handle arithmetic. GPS changed how we navigate. Search engines changed how we retrieve information. But each of those tools removed friction from a specific domain while leaving the judgmental formation environment largely intact. What AI increasingly reduces is friction within the formation process itself — the interval, the strain, the uncertainty, the construction. That is a different kind of substitution.   

There is an interval between a question and its resolution. In that interval, something happens that often goes unobserved because it is invisible, produces no artifact, and leaves no record. The mind generates hypotheses. It compares them. It feels where they are strong and where they fail. It sits with the discomfort of not knowing yet and treats that discomfort as information. It builds, tentatively and imperfectly, a model of the problem.

That process — the friction window — is where judgment forms. Not in the answer. In the interval before the answer.

Think about the last time you genuinely didn't know something and had to figure it out without immediate access to a clean answer. Not the last time you Googled something — the last time you actually sat with a problem, turned it over, felt stuck, tried an approach that didn't work, tried another. If you can remember what that felt like, you remember the friction window. If you can't remember the last time it happened, that itself is data.

AI compresses that window to near zero. Not maliciously. Not carelessly. Efficiently. A fluent, structured, confident answer arrives in seconds. The interval closes before the construction process begins. Resolution precedes formation.

The result is not that people receive wrong answers. Often, they receive very good ones. The result is that the capacity that forms inside the interval — the evaluative, comparative, friction-tolerant capacity we call judgment — is never required. And capacities that are never required are not maintained.

Elena, a third-year medical student, described this recently with unusual clarity. She said that when she looks up a differential diagnosis on an AI platform, she gets a ranked list with explanations. It is usually accurate. It is always organized. It saves her significant time. But she noticed, about four months in, that she had stopped generating her own differential first. She used to sit with a case, make herself produce candidates, feel where her reasoning was uncertain. Now she checks the list and confirms or adjusts. The difference appears minor. She doesn't think it is.

Elena cannot yet measure what she is not building. That is the nature of a formation deficit. It is invisible until the environment removes the scaffold and asks you to stand.

There is a cognitive mechanism that makes all of this feel fine. It is called the fluency heuristic, and it is one of the most widely documented biases in cognitive psychology. Its operation is simple: the brain uses ease of processing as a proxy for reliability.

When information arrives cleanly organized, grammatically smooth, confident in tone, and free of visible contradiction, the nervous system produces a feeling of correctness. Not correctness itself — the feeling of it. That feeling reduces the impulse to verify. It signals that scrutiny is unnecessary. It produces what psychologists call cognitive ease, and cognitive ease is the enemy of judgment formation. 

AI outputs are extraordinarily fluent. They are trained to be. Clean paragraph structure, balanced language, confident assertion, logical flow. Every surface feature that the brain uses as a shortcut for reliability is present and optimized. This is a feature when the output is accurate. It becomes a mechanism of erosion when it is accepted without engagement, which, increasingly, it is.

Here is what the fluency ramp looks like in practice. David is a financial analyst with seven years of experience. He subscribes to an AI research platform that synthesizes earnings data, competitive analysis, and sector trends into briefings he receives each morning. The briefings are excellent — better organized, more comprehensive, and faster than anything he produced manually. He has been using the platform for over a year. 

During that time, David has noticed that he reads the briefings less critically than he used to read his own work. Not because he trusts the platform blindly, but because nothing in the briefings ever feels wrong enough to trigger scrutiny. They are smooth. They cohere. They resolve uncertainty rather than creating it. And so David's evaluative attention — the reflex that used to fire when something didn't quite fit — has quieted, incrementally, because little has required it. 

David is not a worse analyst. His deliverables are stronger than ever. What has changed is the activation frequency of his scrutiny reflex. It is not gone. It is less automatic. And less automatic means it fires when things feel obviously wrong, rather than proactively. That shift — from proactive auditing to reactive flagging — is the fluency ramp. It feels like efficiency. It is actually the slow recalibration of when his judgment engages.

The ramp does not feel like a ramp. It feels like getting better at your job.

The deeper danger of the fluency ramp is that it resets the baseline for what 'something wrong' feels like. Over months, smooth becomes normal. Anything that doesn't feel smooth begins to feel like a problem with the source, not a signal worth examining. The nervous system learns to trust ease and distrust friction. That is the cognitive inversion that makes judgment atrophy self-concealing.

There is a precise difference between constructing an answer and selecting one, and it is not primarily about effort. It is about what each process builds in the person who does it. 

When you construct an answer from uncertainty, you are required to generate hypotheses — to produce candidates from within your own understanding, however incomplete. That act of generation creates internal structure. You feel where you are uncertain. You encounter the places where your model of the problem is thin. You build a map — even a crude one — of the conceptual territory. That map persists after the question is answered. It is part of you. 

When you evaluate a completed answer, you perform a different cognitive act: plausibility assessment. Does this seem right? Does it align with what I know? Is anything obviously wrong? This is not a shallow process — experienced evaluators can do it with great sophistication. But it does not require generation. It does not require you to build the structure internally. It requires only that you decide whether to accept the structure presented. 

The difference compounds over time. The consistent constructor builds internal authority — an expanding, personally calibrated model for reasoning within their domain. The consistent evaluator builds something different: fluency with external structures and sensitivity to surface coherence. Both can look identical in a well-functioning environment. The divergence reveals itself the moment the environment changes.

Consider Sophie, a junior policy analyst at a government agency. She joined three years ago with immediate access to AI drafting tools that her senior colleagues had spent most of their careers without. She produces memos faster than anyone in her cohort. Her writing is clear. Her arguments are organized. Her managers are impressed.

Last spring, her agency faced a genuinely novel situation — a regulatory question without a clear precedent that required her to reason from first principles rather than from an established framework. She described the experience as 'staring at a blank document and not knowing where to start.' She eventually produced something, with significant support from a senior colleague. But what she described was not writer's block. It was the discovery that she had not built the internal scaffolding the novel situation required. She had practiced evaluation. She had not practiced construction. 

Sophie's senior colleague — who had spent years producing drafts in longhand, revising through friction, defending positions in interagency meetings before AI summarization tools existed — sat down with her and built the structure out loud, from first principles, in forty minutes. He was not smarter than Sophie. He was formed differently.
 

What you have read so far is the first-order story. Individual human beings are quietly losing the habit of judgment formation. They do this through repeated, rational, frictionless interaction with tools that do their thinking for them. That story is real and worth telling on its own.

But the second-order story is larger. The second order describes what happens when this individual phenomenon scales across institutions, professions, and generations — and when the systems that depend on distributed judgment capacity begin to operate on a narrowing base of it. The second order concerns what happens to the institutions these individuals eventually lead.

In early 2025, Defense Secretary Pete Hegseth posted on X that Anduril Industries, a defense technology company, had been designated a strategic national security partner. The post produced immediate, concrete financial and political effects. Anduril's valuation moved. Defense industry relationships shifted. The designation carried real weight.

One problem: the legal and regulatory process for a formal national security designation was incomplete. The post outpaced procedure. Its political and financial effects were real; the constraining legal architecture was not in place. 

This example matters not because of its political valence — it can be applied with equal force across administrations and ideological contexts — but because of what it illustrates structurally. Hegseth has a genuine combat formation. He served in Iraq and Afghanistan. He developed a specific kind of judgment under conditions of physical danger and immediate consequence: decisiveness under pressure, tolerance for risk, the ability to act without complete information. That formation is real. It is not ceremonial. 

What an X post designation reveals is that governance formation — the specific capacity to recognize procedural friction as substantive rather than bureaucratic, to understand that the friction of legal process is not delay but architecture —was not part of his development environment. Combat builds one kind of judgment. Governance architecture builds another. The mistake is not that Hegseth acted with confidence. The mistake is that the formation required to recognize when confidence must yield to procedure never occurred in this domain. 

The second-order consequence emerges in the signal that this action sends to decision-makers observing the event. When a smooth action—such as a post, a statement, or an announcement—produces immediate political and financial results, and legal vulnerability becomes a deferred issue, the observer internalizes that impact defines legitimacy. They come to understand that speed and visibility create influence in ways that procedural compliance does not. They perceive formation architecture as optional when momentum is accessible. 

The observer who internalizes that lesson is not a villain. They are a rational agent responding to the reward signal they witnessed. And that is how institutional judgment degrades — not through corruption, but through demonstrated incentive. The next generation of decision-makers is watching what succeeds. If smooth output producing immediate results is what succeeds, smooth output producing immediate results is what they practice. 

This is the second order of institutional smoothing. The first order is an individual bypassing procedure. The second order is a culture learning that bypassing procedure produces visible results.     

 dIn early 2025, Anthropic’s AI model Claude began operating inside the U.S. military’s classified networks, through a partnership with Palantir. Pentagon officials praised its capabilities. By available operational reporting, the relationship appeared to be working.

What followed became a public test of whether an institution could maintain a formation constraint under conditions designed to remove it.

Contemporary reporting indicated that Defense Secretary Hegseth convened a meeting with Anthropic CEO Dario Amodei and presented two options: accept unrestricted military use of Claude under an “all lawful purposes” standard, or face contract cancellation and a supply chain risk designation — a status previously reserved for foreign adversaries. The deadline was a Friday evening. A senior Pentagon official’s reported language was unambiguous: “We’re dead serious.”

Anthropic’s position was specific and technical, not categorical. The company was willing to adapt its usage policies for military purposes. It held two lines: Claude would not be used for mass domestic surveillance of Americans, and would not be used to develop weapons that fire without human involvement. These were not abstract ethical preferences or branding positions. They reflected an honest assessment of what AI systems can and cannot yet reliably do—an acknowledgment of the gap between deployed capability and trustworthy autonomous performance: the gap this site refers to as the missing fifteen percent, which the builders of these systems have documented in their own technical records.

Amodei said the company could not “in good conscience” accept the Pentagon’s terms. Reported consequences followed — a presidential directive to federal agencies to cease using Anthropic’s technology, the supply chain risk designation, and cascading commercial consequences across the contractor ecosystem. Hours later, OpenAI announced a deal with the Pentagon that included prohibitions on domestic mass surveillance and autonomous weapons — the same two restrictions Anthropic had refused to drop. One company absorbed the designation. Another received the restrictions in writing after the first company demonstrated they were worth holding.

The significance of this case for the epistemic agency framework is structural, not political. It can be applied with equal force across administrations and ideological contexts — and it will recur in different forms as long as AI systems retain built-in constraints that institutional actors find inconvenient.

What the case illustrates is epistemic agency operating at the institutional level under maximum pressure. The company identified two specific formation constraints. It held them when the environment — a government ultimatum, a designation, cascading commercial consequences — was designed to force abandonment. The constraints held not because the pressure was manageable but because the institution had formed a position it could defend from the inside out. It knew why the constraints existed. It knew what would have to be true to change them. It could articulate what was at stake clearly enough to refuse consequences that were severe and immediate.

That is what formed judgment looks like when the scaffold is removed, and something is still standing.

This does not make every decision Anthropic has made correct, or the institution immune to its own drift. It makes this particular case an observable specimen of what formation architecture looks like when the cost of maintaining it is real and documented. The reader who wants to understand what happens next — what institutional pressure does to the instruments themselves, and what the public record now shows about the formation environment of the AI systems being built to replace the ones that held the line — will find that argument developed fully elsewhere on this site.

What this page requires is the structural observation and nothing more: when the environment removes the requirement for friction, and the reward signal favors smoothness, the institution that chooses difficulty on purpose — and can say precisely why — is the institution that still retains structural integrity.  

The erosion of judgment formation is not occurring evenly. This is one of the most important and often overlooked structural dynamics of the AI transition, and it deserves honest examination. Certain environments still require formation. Military academies that embed physical and psychological stress into their developmental structure. Surgical residencies with grueling hours and high-consequence feedback. Elite law programs that require oral argument and Socratic interrogation. 

Certain consulting firms that require partners to reconstruct client analyses without notes, on demand, under questioning from a room of skeptics. These environments preserve friction not because they are old-fashioned but because the stakes of formation failure make that friction necessary. 

These environments tend to be expensive, selective, and concentrated. Access to them correlates with economic and educational privilege. The formation they produce — the capacity for independent reconstruction under pressure — is becoming a rare resource concentrated in a shrinking fraction of the professional population. 

Meanwhile, the broader environment is optimizing for smooth output. Schools measure completion, not construction. Employers measure deliverable quality, not unaided reasoning capacity. Platforms reward content fluency, not evaluative depth. The incentives, taken together, produce a population increasingly skilled at evaluation-with-tools and increasingly unpracticed at construction without them. 

The consequence is epistemic stratification. Not the visible, dramatic stratification of class or education in the traditional sense — a stratification of cognitive formation that remains largely invisible in surface performance, concealed by fluent output, and revealed only under specific conditions of stress, novelty, or scaffolding removal.

Those conditions tend to arrive. Economic disruption. Institutional crisis. Novel problems that precede their templates. The moment when the scaffold fails, and someone has to build from first principles, is the moment when formation stratification becomes consequential. And the people who discover in that moment that they cannot reconstruct coherence without scaffolding are not, in most cases, people who chose to avoid difficulty. They are people who were never required to practice it. 

The stratification risk is not that elites will suppress others. It is that judgment capacity — the capacity to independently evaluate, detect drift, and reconstruct coherence — will quietly concentrate among those who retained formation environments, while the majority becomes dependent on mediated coherence that they cannot audit. That concentration creates influence asymmetry without requiring coercion. It is structural.

An institution producing documents that look like judgment is not the same as an institution exercising it. This distinction, obvious in principle, is almost impossible to detect in practice — because the surface is all most observers see.

 Real judgment leaves residue. When a person or institution has genuinely worked through a hard problem, there is evidence: documented tradeoffs, recorded dissent, acknowledged uncertainty, and revised positions with explanation. The reasoning substrate is visible in the artifact because the construction actually occurred. You can push on the document and find something solid underneath it.

 Smoothed output leaves a different signature. It looks complete. It sounds authoritative. But push on it — change an assumption, introduce a new constraint, ask why this conclusion rather than an adjacent one — and you find, sometimes quickly, that the reasoning substrate is thin. The words are there. The logic they described is not. The document was assembled; it was not built. 

The drift that results from institution-wide smoothing is the most dangerous kind because it is self-concealing. Each document looks fine. Each output tests fine in isolation. The drift only becomes visible when you compare current decisions to original mandates and notice that the distance between them is larger than any single step justified. The institution has moved without any actor having made the move. 

Who can detect this? Only people with a pre-smoothing baseline. Those who remember what authentic struggle looks like in the production of institutional reasoning. Who remember when memos had margin notes disagreeing with themselves. When policies came with explicit uncertainty ranges. When decisions were accompanied by the evidence of what was considered and rejected. People who still carry that standard of formation evidence can recognize its absence. As those people retire, the baseline disappears too — and the institution loses its internal capacity to know that anything has changed. 

This is not a theoretical concern. It is observable now in organizations that adopted AI tools broadly and early. The outputs improved. The formation residue decreased. The documents got smoother. The underneath got thinner. And in most cases, no one noticed, because the metrics being measured were measuring outputs, not formation.
 

The term “epistemic agency” already exists in philosophy and education research, where it describes the capacity to actively shape, evaluate, and take responsibility for one's own knowledge. What the AI era has made newly urgent is the structural threat to that capacity — not through suppression, but through substitution.

In this framework, epistemic agency consists of four capacities. It is the capacity to generate reasoning independently — to start from uncertainty without waiting for a completed structure. It is the capacity to evaluate claims without outsourcing the evaluation — to bring your own scrutiny rather than adopt someone else's verdict. It is the capacity to detect drift from principle — to notice when an institution, argument, or decision has moved farther from its foundation than any visible step authorized. And it is the capacity to reconstruct coherence when scaffolding is removed — to build from first principles when the tools fail, the templates don't apply, and the situation is genuinely new.

What AI produces — even correct AI output — is not epistemic agency. It is epistemic access. Access to answers, frameworks, and analyses. Access is not nothing. In many contexts, it is transformative. But access and agency are not the same, and the difference matters most precisely when access fails.

The person with epistemic agency who uses AI becomes more capable. The tool accelerates what they already know how to do. They can detect when the output is wrong, reconstruct the reasoning under questioning, and adapt when the context changes. The tool is an instrument in hands that know how to hold it.

The person who formed through AI-mediated resolution — who evaluated rather than constructed, who accepted rather than built — may be equally fluent and equally productive in a stable environment. The divergence appears under pressure. Novel constraints. System failure. Genuine ambiguity without a template. In those conditions, the person with formed epistemic agency builds. The person without it looks for a prompt.

The precise name for what is missing when judgment is replaced by smoothly accepted answers is not intelligence. It is not knowledge. It is not character. It is self-authored coherence — the capacity to construct and defend a position from the inside out rather than the outside in. When you have it, you know why you believe what you believe, and you can explain what would change your mind. When it is absent, you can produce the language of conviction without possessing its architecture.

AI can give you the answer. It cannot give you the understanding of why you believe it. That understanding forms only through the struggle to build it yourself. 

Democracy is the political system most dependent on distributed judgment capacity. It does not require every citizen to be an expert. It requires that citizens can hold competing claims in tension long enough to evaluate them, detect when institutions are drifting from their mandates, and sustain engagement with questions that do not resolve cleanly. Those are the Jagged capacities — the friction-dependent skills — that self-governance requires to remain self-governance rather than ratification of whatever narrative is most fluently presented.

 The trajectory, described without alarmism, runs through identifiable phases. In the first phase, citizens increasingly select the most coherent narrative rather than evaluating competing claims. Not because they are manipulated, but because evaluation is metabolically expensive while fluent synthesis is cheaper. Politics becomes curation rather than deliberation. In the second phase, deliberation itself is compressed because friction looks like failure. Governance shifts toward faster, more decisive executive action — which rewards smooth output and punishes slow process. Citizens reward speed and punish complexity because the habit of tolerating complexity is weakening. In the third phase, verification migrates to credentialed intermediaries: platforms, institutions, AI systems, and influential interpreters. Epistemic authority centralizes as individual evaluation capacity weakens. In the fourth phase, mandate drift becomes continuous and undetected. Institutions change their behavior incrementally, and the public, accustomed to mediated synthesis, cannot reconstruct what the original mandate was or how far the current behavior has traveled from it. 

None of these phases require bad actors. They require only that the formation environment change and that institutions respond rationally to the incentives that change creates. Democracy does not fail because someone dismantles it. It phases into managed governance — a system that retains the procedural shell of self-government while the substance of judgment migrates to a smaller and smaller population capable of exercising it.

The civic version of epistemic agency is not sophisticated policy expertise. It is the basic capacity to say, from the inside out: 'I know why I believe this, and I know what would change my mind.' A population that has practiced that capacity through friction — through reading that genuinely confused them before it clarified, through arguments that required them to defend positions under real scrutiny, through decisions made without certainty — is harder to move through narrative alone. It requires evidence. That requirement is, in the precise sense, what makes democracy function.

The argument made here is not that AI is bad, that technology should be resisted, or that pre-digital cognition was superior. It is more specific than that and more actionable. 

Every major cognitive technology in history has shifted what humans practice. Writing reduced the cultivation of oral memory; it also enabled literature, law, and science at scales oral memory alone could not support. The printing press dispersed knowledge in ways individual memory and manuscript could not; it also accelerated formation by making texts that require genuine intellectual wrestling widely available. Calculators reduced the cultivation of mental arithmetic; they also freed cognitive resources for the conceptual mathematics that calculators cannot perform. GPS reduced the cultivation of spatial navigation; it also enabled logistics and emergency response at scales previously impossible. 

In each case, something was traded. The question in each case is not whether the trade was worth making — history generally resolved that — but whether what was lost was incidentally lost or deliberately abandoned, and whether the resulting system retained enough of the original capacity to function under the conditions that require it.

 AI is not different in kind from these earlier tools. It is different in scale and scope. It does not remove friction from a single cognitive domain. It offers to remove friction from the formation process itself — from the interval of uncertainty, construction, and self-correction where judgment forms. That is a broader intervention, with broader consequences, and it requires more deliberate attention to what is being preserved.
 

Formation architecture is not prohibition. It is design. It is the deliberate preservation of environments where independent construction is required — not as punishment, not as tradition, but as the training load that judgment requires to form and remain durable.

 In education, formation architecture means insisting that certain acts of intellectual construction remain independent. Not every assignment — but enough that students regularly practice starting from uncertainty, generating structure without a template, defending positions under questioning, and discovering, from the inside, where their reasoning breaks. The specific form is less important than the principle: the practice of construction cannot be entirely replaced by the evaluation of completed outputs without changing what forms.

Elena, the medical student, described a professor who required students to generate their own differential diagnosis in writing before accessing any reference — AI or otherwise. The requirement felt arbitrary at first. Over a semester, students who followed it began to notice something: their sense of clinical uncertainty became more precise. They knew where they didn't know, which is the foundational skill of good medicine. The protocol preserved the friction window.

 In professional life, formation architecture means designing workflows that require independent reasoning at the stages where formation matters — before AI synthesis, not as a replacement for it. The manager who requires team members to articulate their own analysis before reviewing AI output is not being inefficient. They are preserving the generative step that builds judgment. The difference between 'What do you think?' asked before someone reads a summary and 'What do you think?' asked after is the difference between constructive thinking and reactive evaluation. 

In governance, formation architecture is procedural friction. The legal review that slows the X post. The committee process that requires a decision to be defended before multiple stakeholders. The documentation requirement that forces the articulation of reasoning before action. These mechanisms feel like delay. They are actually architecture. They exist because people who built institutions before our current moment understood that governance without friction produces decisions without accountability — and that the friction is not incidental to the legitimacy, it is constitutive of it.

 At the individual level, formation architecture is simpler and more personal. It is the deliberate maintenance of friction in some portion of your own cognitive life. Writing something from scratch before asking AI to improve it. Reading a primary source before a summary. Sitting with a question long enough to form your own tentative answer before looking for the established one. Not as a rule, but as a practice — the recognition that judgment, like any trained capacity, requires use to remain intact.
 

The second order of this phenomenon is not inevitable. It is the consequence of choices — not dramatic ones, not philosophical commitments made with full awareness, but the aggregate of rational individual and institutional decisions in environments that reward smoothness and make friction feel like a defect.

 What makes the second order a choice rather than a fate is that formation architecture remains possible. Schools can design assignments that require independent construction. Employers can evaluate reasoning processes rather than only outputs. Governance institutions can maintain procedural requirements even under pressure to move faster. Individuals can practice difficult cognition on purpose. 

None of that will happen automatically. Markets reward efficiency. Platforms reward engagement. Institutions under pressure reward speed. The incentive gradient currently runs strongly toward smoothness, and nothing in the current environment corrects for it organically. Formation architecture requires deliberate choice, sustained over time, in systems whose default pull runs the other way. 

The people best positioned to make those choices are the people who understand what is at stake — who can see, clearly and without either alarm or dismissal, that the friction being removed is not inefficiency. It is training load. And training load, removed comprehensively and early enough, does not merely slow the development of a skill. It prevents the formation of the person who would know what to do when the system stops working.

Marcus, the marketing director, eventually understood what happened in that meeting. He did not catastrophize. He started a practice: one substantive piece of work per week, built entirely by himself, without AI assistance, from blank page to finished draft. He described it as 'surprisingly hard for the first two months, then surprisingly clarifying.' He said he had forgotten what it felt like to know that an argument was his — not in the sense of ownership, but in the sense of structural confidence. He knew why every paragraph was where it was. He knew what he would change if the constraints shifted. He could defend it from the inside. 

That is what epistemic agency feels like when it is working. It is not the absence of tools. It is the presence of internal authority — the capacity to construct coherence when the tools are absent, when the templates don't fit, when the situation is genuinely new. 

What the current moment requires is not a defense of difficulty for its own sake. It is the recognition that formation and productivity are different things — and that environments optimized entirely for one will, over time, hollow out the other. 

The second-order effects of smoothing are not abstract. They live in Marcus's moment of exposure in a conference room. In Elena's discovery that she had stopped generating before checking. In Sophie's stare at a blank document. In the institutional drift that no one inside can see because no one inside remembers what it looked like before the drift began.   

They also live in the opposite: in the professor who required the differential first. In the organization that said no to a government’s demand to remove its principles. In the individual who sits with a question a little longer before asking for the answer.

Those are not grand gestures. They are formation architecture. Small, deliberate, friction-preserving choices made in environments designed to eliminate them.

The first order of AI is that it makes us more productive. The second order is what we become in the process. That second order is still, for now, a choice.

© 2026 The Human Choice Company LLC. All Rights Reserved.
Authored by Jim Germer.

This document is protected intellectual property. All language, structural sequences, classifications, protocols, and theoretical constructs contained herein constitute proprietary authorship and are protected under international copyright law, including the Berne Convention. No portion of this manual may be reproduced, abstracted, translated, summarized, adapted, incorporated into derivative works, or used for training, simulation, or instructional purposes—by human or automated systems—without prior written permission.

Artificial intelligence tools were used solely as drafting instruments under direct human authorship, control, and editorial judgment; all final content, structure, and conclusions are human-authored and owned. Unauthorized use, paraphrased replication, or structural appropriation is expressly prohibited.