Why AI in Education Feels Helpful—and Still Feels Wrong

---

Orientation/Thesis

AI in education is most often evaluated through efficiency, access, and outcome metrics—how quickly work is completed, how polished outputs appear, or how widely support can be distributed. What these evaluations routinely omit is developmental cost accounting: the loss of human capacities that only form through sustained effort over time. Developmental cost accounting refers to tracking the capacities that fail to form when a tool substitutes for the learner’s effort—costs that short-term output metrics do not capture.

AI does not merely accelerate execution; in many educational contexts, it substitutes for formative processes: grappling with uncertainty, organizing thought, correcting one’s own errors. Because the benefits of this substitution are immediate and visible while the costs are delayed, diffuse, and difficult to measure, current debates systematically undercount what is being traded away.

This essay examines that trade not to argue for bans, moral restraint, or technological retreat, but to explain why benefit-only arguments, even when accurate, feel intuitively incomplete to many educators and observers. The claim here is not that AI is bad, but that benefit-only accounting is incomplete when the environment’s goal is formation.

Why AI-in-Education Debates Keep Missing Each Other

Discussions about artificial intelligence in education have settled into a familiar stalemate. The participants are often serious, informed, and acting in good faith, yet the conversation rarely advances. Each side presents arguments that appear internally coherent, and each leaves the other unsatisfied. What emerges is not productive disagreement so much as mutual incomprehension: a sense that the other side is answering a different question.

On one side, advocates emphasize measurable benefits. AI systems can accelerate routine tasks, expand access to tutoring and feedback, personalize instruction at scale, and reduce the workload on teachers already operating under strain. These claims are not speculative. They are visible in pilot programs, classroom tools, and policy proposals. In many contexts, they are demonstrably true. Students complete assignments faster. Teachers offload grading and administrative labor. Institutions see efficiency gains they can quantify and report.

On the other side, skepticism surfaces in less precise language. Concerns are often expressed as unease rather than argument: that students are “not really learning,” that something about thinking or writing feels thinner, that reliance on AI may hollow out skills that education is meant to cultivate. These objections are frequently dismissed as moral panic, nostalgia for older methods, or resistance to technological change. When skeptics struggle to specify mechanisms or metrics, their concerns can appear vague or even reactionary.

The result is a debate in which each camp talks past the other. Proponents respond to concerns about learning by citing improved outputs. Skeptics respond to demonstrations of efficiency by insisting that outputs are not the point. Both are correct within their own frames, and neither is fully answering the other. The disagreement persists not because one side lacks evidence or intelligence, but because they are measuring different things.

Most benefit arguments implicitly treat education as a production system. Inputs are time, tools, and instruction; outputs are completed assignments, test scores, or demonstrable competencies. Within that frame, AI is easy to evaluate. If a tool improves outcomes faster or at lower cost, it is working. The metrics are legible, the gains are immediate, and the comparison is straightforward.

Many skeptical intuitions, by contrast, are oriented toward development rather than production. They are less concerned with what students can produce today than with what capacities they are forming for tomorrow. This perspective is harder to formalize because it deals in counterfactuals: skills that might have developed under different conditions, habits of mind that take years to reveal their absence, forms of independence that only become visible when support is withdrawn. These are not easily captured by short-term metrics, and they do not announce themselves when lost.

Because these two frames rarely meet explicitly, the debate becomes distorted. Efficiency arguments sound dismissive of human development. Developmental concerns sound like opposition to progress. Each side experiences the other as missing something obvious. In reality, what is missing is not goodwill or rigor, but a shared variable that would allow the tradeoffs to be named.

What’s missing is not intent or intelligence—it’s a variable.

The Missing Variable: Developmental Substitution

To understand why AI-in-education debates so often misfire, it is necessary to name the variable that is rarely accounted for directly: developmental substitution. Without this concept, discussions default to arguments about efficiency and access on one side, and vague concerns about “learning” on the other. With it, the tradeoffs become clearer and more tractable.

Education serves at least two distinct functions. One is execution: producing correct answers, coherent essays, or completed problem sets. The other is formation: the slow development of internal capacities that make those outputs possible in the first place. Formation includes skills such as organizing thought, building mathematical intuition, decomposing problems, and maintaining internal feedback loops that allow a learner to notice and correct their own errors. These capacities cannot be acquired through shortcuts; they form only through sustained effort over time.

AI systems are exceptionally good at execution. They can generate text, outline arguments, solve problems, and correct errors with speed and consistency. In many educational contexts, however, this strength creates a subtle shift. Rather than supporting the learner’s formation, AI often substitutes for the very processes through which formation occurs.

The work still gets done, but it no longer happens inside the learner.

This distinction matters because learning systems rely on specific mechanisms that cannot be bypassed without consequence. Friction forces a learner to grapple with uncertainty. Feedback, especially when delayed or imperfect, encourages internal calibration. Iteration builds endurance and pattern recognition. Error provides the signal that something needs to change. Together, these elements form the conditions under which durable capacities develop. When AI compresses or removes them, the immediate experience improves—less frustration, faster completion—but the underlying developmental work may never take place.

The substitution is easy to miss because the outputs often improve. An essay generated or heavily reworked by an AI system can appear more polished than one produced independently by a novice writer. A step-by-step solution provided by an AI tutor can lead to higher homework accuracy. From an execution standpoint, the system is functioning as designed. From a developmental standpoint, however, the learner may have skipped the struggle that would have built internal structure. The cost is not a lower grade, but a diminished capacity that emerges only later.

Crucially, this substitution is not uniformly harmful. For individuals who already possess well-formed capacities, the developmental cost is minimal or nonexistent. An experienced writer using AI to explore alternative phrasings is not outsourcing their ability to structure thought; they are leveraging an existing skill set. In such cases, AI functions as an amplifier rather than a replacement. The same tool that substitutes for formation in a novice can augment judgment in an expert.

This asymmetry is why framing the issue as “cheating” or “laziness” misses the point. The concern is not about intent or effort, but about where the effort occurs. When effort is displaced from the learner to the tool before formation has taken place, the system produces competent-looking outputs without cultivating the capacities that education is meant to build. When displacement occurs after formation, the effect is largely benign and often beneficial.

Developmental substitution, then, is not a moral failure or a technological flaw. It is a structural effect of introducing powerful execution tools into environments designed for formation. Ignoring it leads to incomplete analyses that overstate benefits and underestimate long-term costs. Naming it allows the conversation to move beyond generalized unease and toward a clearer understanding of what is being traded when AI systems enter educational settings.

Timing Asymmetry: Why Early Use Costs More Than Late Use

The effects of AI in education cannot be understood without accounting for timing asymmetry. The same tool, introduced at different stages of development, can produce opposite outcomes. This asymmetry helps explain why some uses of AI appear benign or even beneficial, while others quietly undermine the very capacities education is meant to cultivate.

For learners early in a skill’s development, AI often replaces struggle rather than supporting it. Tasks that would normally require sustained effort—drafting an argument, working through an unfamiliar problem, organizing ideas into a coherent structure—are partially or fully offloaded to the tool. The immediate experience improves. Confusion is reduced, frustration is minimized, and progress appears smoother. Yet the internal model that would have formed through grappling with the task may never fully develop. The learner arrives at the destination without having traversed the terrain.

This loss is difficult to detect at the moment it occurs. Early substitution does not usually result in obvious failure. Assignments are completed. Grades may even improve. The cost reveals itself later, when the learner is expected to operate without scaffolding or to transfer skills to novel contexts. At that point, the absence of an internal framework becomes apparent—not as a single missing fact, but as a general fragility in reasoning, expression, or problem framing.

By contrast, when AI is introduced after formation has largely taken place, its effects are markedly different. A developed practitioner already possesses the internal structures that guide judgment and sense-making. For them, AI compresses routine labor rather than replacing formative effort. Drafting variants, exploring alternative approaches, or checking assumptions can free attention for higher-order decisions. The capacity is not displaced; it is amplified.

Everyday analogies clarify this timing dependence. Training wheels support balance while a child learns to ride a bicycle, but replacing the bicycle with a motorized scooter short-circuits the learning process entirely. Calculators introduced after number sense has developed extend mathematical reach; calculators introduced before it can prevent number sense from forming at all. GPS systems assist navigation for experienced drivers, yet habitual reliance before spatial mapping skills develop can leave users disoriented when the tool fails. In each case, the issue is not the tool itself, but when it enters the learning sequence.

Development is path-dependent. Certain stages are not simply hurdles to be bypassed; they are the mechanisms by which internal structure is built. Once skipped, they are difficult to recreate efficiently. Later remediation often requires more time and effort than initial formation would have, and it can be difficult to recover the same depth under compressed conditions. This is why early substitution carries disproportionate cost: it forecloses possibilities before they are fully visible.

Recognizing timing asymmetry does not require alarmism. It requires acknowledging that educational tools interact with human development differently depending on when they are deployed. Treating AI as uniformly beneficial or harmful obscures this reality. The more accurate view is that AI’s developmental impact is contingent: early use tends toward substitution, late use toward augmentation. Any serious accounting of AI in education must therefore ask not only what a tool can do, but when its use reshapes the trajectory of learning in ways that cannot easily be reversed.

Concrete Educational Scenarios (Tradeoffs, Not Condemnation)

Abstract discussions of development and timing can feel speculative until they are grounded in ordinary classroom activity. The developmental effects of AI become clearer when viewed through specific scenarios, not as moral failures or technological missteps, but as tradeoffs between immediate performance and long-term capacity.

Scenario A — Writing Instruction

Consider a student tasked with writing an analytical essay. With access to an AI system, the student can generate a draft, rephrase unclear passages, and reorganize the structure with minimal effort. The immediate benefits are obvious. The essay reads smoothly, arguments appear coherent, and the work is completed quickly. From the perspective of execution, the outcome is superior to what the student might have produced independently.

What is less visible is what did not happen. The student bypasses the process of clarifying their own thoughts, how one idea leads to another, or where an argument falters. The internal work of organizing thought, sustaining a line of reasoning, and revising based on one’s own judgment is largely displaced. Over time, this can result in a shallow sense of authorship. When the scaffolding is removed—during an in-class exam, a timed assessment, or a professional setting that demands original composition—the student struggles to reproduce similar quality unaided. The cost is not that the essay was “too easy,” but that the formative practice of writing never fully occurred.

Scenario B — Problem-Solving in STEM

A similar pattern appears in quantitative disciplines. An AI tutor can walk a student through a complex problem step by step, correctly decomposing it and explaining each move. Homework accuracy increases. Students feel more confident completing assignments, and instructors see fewer errors. These are genuine improvements.

Yet problem-solving involves more than executing known procedures. It requires learning how to frame an unfamiliar problem, decide which tools are relevant, and recognize when an approach is failing. When AI consistently supplies the correct decomposition, the student may never practice this framing work. The result is a learner who performs well on problems that resemble worked examples but falters when confronted with novel or open-ended tasks. The developmental cost emerges not as lower grades, but as poor transfer: difficulty applying knowledge outside the narrow contexts in which it was acquired.

Scenario C — When Timing Changes the Outcome

Not all AI use in education carries the same tradeoff. Consider a graduate student or experienced professional who already possesses strong writing or analytical skills. When such an individual uses AI to generate alternative phrasings, explore counterarguments, or stress-test an idea, the tool functions differently. The user evaluates outputs against an existing internal standard. Judgment remains internal; the AI serves as a prompt rather than a substitute.

In this case, formation has already taken place. The capacities that allow the user to assess quality, coherence, and relevance are intact. AI augments these capacities by expanding the space of possibilities and reducing routine labor. The same mechanisms that undermine formation in novices can increase leverage for late-stage learners.

Across these scenarios, the pattern is consistent. AI use that replaces early-stage formative effort tends to produce competent outputs alongside thinner internal capacities. AI use that follows formation tends to amplify judgment without displacing it. The distinction is not between good and bad technology, but between different points in the developmental arc. Seeing these tradeoffs clearly allows the discussion to move beyond condemnation or enthusiasm and toward a more precise understanding of what is gained—and what may be quietly lost—when AI enters educational practice.

The Silent Cost Problem

One reason developmental substitution remains under-examined is that its costs are structurally quiet. The benefits of AI use in education are immediate, legible, and easy to count. Assignments are completed faster. Output quality improves. Teachers save time. These effects appear directly in the metrics institutions already track, making them visible to administrators, policymakers, and parents alike.

The costs operate on a different timeline. When formative effort is substituted early, the loss does not register as a failure. Nothing visibly breaks. Students advance through courses, grades remain acceptable, and systems appear to function smoothly. What is lost is not an outcome, but a counterfactual: a capacity that would have developed had the learner been required to do more of the work internally. Because this capacity never fully forms, there is no clear moment at which it can be observed leaving.

Educational measurement systems are poorly equipped to detect this kind of loss because they are designed to track outputs—test scores, completion rates, time saved—not the durability or independence of the underlying skill. This is not a critique of measurement or administrators; it’s a recognition that standard outcome metrics don’t capture developmental dependence until the scaffolding is removed.

A student who completes an assignment accurately with AI assistance is indistinguishable, in most metrics, from one who could complete the same work unaided. Dependency looks identical to competence until the scaffolding is removed.

This creates a systematic bias. Benefits accrue early and are attributed directly to the tool. Costs are delayed, distributed across future contexts, and often externalized to later stages of education or professional life. When a graduate struggles to write independently, frame problems without guidance, or adapt knowledge to unfamiliar situations, the cause is rarely traced back to earlier substitutions. The system registers a skill gap, not the developmental history that produced it.

The absence of immediate evidence is therefore misleading. In developmental systems, absence of evidence is not evidence of absence. Some losses only become visible under stress—when time is limited, support is withdrawn, or novelty demands independent judgment. By the time these conditions arise, the opportunity for straightforward remediation has often passed.

Asymmetric Returns: Who Gains and Who Loses

The same measurement failure that obscures developmental costs also masks how unevenly AI’s benefits are distributed. While AI tools are often framed as democratizing—providing everyone with access to high-level assistance—their effects tend to increase variance rather than reduce it.

For individuals with well-formed capacities, AI functions as leverage. An experienced writer can use AI to explore alternative structures, test arguments, or accelerate drafting without relinquishing control. A skilled problem-solver can use AI to check assumptions or survey solution spaces more efficiently. In these cases, the tool compounds existing advantage. Judgment remains internal; execution is optimized.

For novices, the dynamic is different. Early reliance on AI replaces the scaffolding that would normally be built through effort and error. Rather than accelerating development, the tool can truncate it. The learner appears to keep pace in terms of output, but their internal capacities lag behind. Over time, this can produce earlier and deeper tool dependence, making it harder to operate without assistance.

The result is a subtle stratification. On the surface, everyone has access to the same tools. Beneath the surface, those who enter with stronger foundations pull further ahead, while those still forming those foundations lose opportunities to build them. What looks like equalization at the level of output can become divergence at the level of capability.

This divergence is unintuitive because the tools look identical. Two students may submit essays of comparable quality, both aided by AI. One uses the tool to refine an argument they already understand; the other uses it to supply structure they have not yet learned to create. The immediate products converge, but the long-term trajectories do not. Without attention to formation, systems risk mistaking short-term parity for long-term equity.

Taken together, the silent cost problem and asymmetric returns point to the same underlying issue. When developmental effects are unmeasured, tools are evaluated solely by their visible gains. Those gains are real—but they are not evenly distributed, and they do not come without tradeoffs. Understanding who benefits, who loses, and why requires looking beyond outputs to the capacities education is meant to cultivate over time.

Why Benefit Arguments Feel Incomplete (Even When True)

Many of the strongest arguments in favor of AI in education are not wrong. They are incomplete. This is why the debate produces such persistent dissatisfaction: one side points to real gains, the other senses real loss, and neither can fully account for the other’s experience within their chosen frame.

Most benefit arguments rest on implicit assumptions about learning. Capacity is treated as relatively static: a student either has a skill or does not, and tools merely help express it more efficiently. Learning, in this view, is largely interchangeable with output. If a student can produce a competent essay or solve a problem correctly, the underlying capacity is assumed to be present or at least developing appropriately. Within this framework, AI appears as a clear improvement—an accelerant that reduces friction without altering fundamentals.

Developmental systems violate these assumptions. Capacity is not static; it is constructed. Output is not a reliable proxy for formation, especially when external scaffolding is involved. Two learners may produce identical work while undergoing very different developmental trajectories. One may be consolidating internal structure through effort and feedback; the other may be bypassing that process entirely. Short-term metrics struggle to distinguish between these cases, yet their long-term consequences diverge sharply.

This creates a counterfactual gap that benefit arguments cannot easily address. To fully account for developmental cost, one would need to compare a learner’s current performance not only to their past outputs, but to the capacities they might have developed under different conditions. That comparison is inherently difficult. The alternative path was never taken. The capacity that did not form leaves no direct trace. As a result, losses remain invisible until a later context demands independent competence—and even then, the cause is often misattributed.

Institutional incentives reinforce this blind spot. Educational systems are under pressure to demonstrate improvement through measurable outcomes. Efficiency gains and performance metrics are politically and administratively legible. Developmental deficits, by contrast, surface slowly and diffusely, often outside the scope of the institution that introduced the tool. When a problem appears years later, it registers as an individual shortcoming or a curriculum issue, not as a delayed cost of earlier substitution.

This is why benefit arguments can feel simultaneously convincing and unsatisfying. They answer the question they are designed to answer—does the tool improve observable outcomes?—while leaving a deeper question unresolved. What happens to the learner’s capacity over time? Which forms of effort were displaced, and what would they have produced if allowed to run their course?

Reframing the discussion around these questions does not negate the benefits of AI. It clarifies their limits. The relevant inquiry is not simply whether AI helps, but who it helps, when it helps, and at the cost of what formation. Until benefit arguments incorporate this developmental accounting, they will continue to feel incomplete—not because they are false, but because they leave something essential uncounted.

A Practical Description of Where AI Fits Best (Descriptive, Not Prescriptive)

If the preceding analysis is correct, then the question facing educators and institutions is not whether AI belongs in education, but how it fits within environments whose primary function is capacity formation. Answering that question does not require bans, moral appeals, or urgency. It requires a clearer sense of what AI does well—and where its strengths intersect awkwardly with developmental goals.

This section describes how the tool tends to behave across different stages of development; it is not a policy recommendation.

One way to describe a sane approach is to treat AI primarily as a late-stage amplifier. In this role, the tool operates after foundational capacities are in place, accelerating execution without displacing the work that built those capacities. For learners who already possess internal standards of quality and judgment, AI can compress routine labor, expand the space of options, and surface alternatives that invite evaluation rather than replace it.

AI can also function as a reflective mirror. Used this way, it externalizes possibilities—drafts, solutions, explanations—against which a learner compares their own thinking. The value comes not from adopting the output wholesale, but from testing it: identifying what fits, what fails, and why. When judgment remains internal, the tool supports calibration rather than substitution.

A related role is as a stress-test tool. By quickly generating variations or counterexamples, AI can expose the limits of a learner’s understanding and prompt deeper inquiry. Here again, the benefit depends on timing. Stress-testing presupposes something to stress. Without prior formation, variation becomes noise rather than insight.

What these uses share is a common boundary: they avoid early replacement of core formation tasks. When AI fully automates sense-making stages—deciding what to say, how to structure an argument, or how to frame a problem—before learners have practiced those acts themselves, it shifts effort out of the learner at precisely the moment it is most developmentally productive. The result is not faster formation, but thinner formation.

Seen this way, educational systems are not merely pipelines for producing outputs. They are environments designed to cultivate durable capacities over time. Tools that optimize for production alone risk misalignment with that purpose unless their developmental effects are explicitly considered. This posture does not deny AI’s benefits; it situates them. It recognizes that execution gains and formation costs can coexist, and that both must be accounted for.

The central implication is simple. Any educational use of AI that substitutes for formative effort carries a cost that does not appear on efficiency dashboards or outcome metrics. Naming that cost does not require rejecting the tool. It requires acknowledging what education is for, and evaluating AI not only by what it produces today, but by what it helps learners become tomorrow.

Closing — Naming the Missing Cost

The benefits of AI in education are real, and they are not going away. So are the tradeoffs. What has been missing from most discussions is not evidence of harm or proof of misuse, but a way to name a cost that does not announce itself. When AI substitutes for formative effort, the loss is not immediately visible in grades or outputs; it appears later, as thinner independence and reduced capacity to operate without support. The unease many register around AI in education is therefore not nostalgia for older methods or fear of new tools. It is a recognition that something essential is being spent without being counted. Until educational debates account for development as well as efficiency, they will continue to feel unsettled—not because the benefits are illusory, but because the costs remain unnamed.

Written by W.E. Mercer

February 2026