The rapid emergence of sophisticated AI tools has ignited a fierce debate across classrooms, lecture halls, and online learning platforms about what these technologies mean for the future of education. On one side, critics warn that when students simply hand off their assignments to a chatbot, they short‑circuit the very cognitive processes that deepen understanding and build lasting skill. On the other, enthusiasts point to AI’s capacity to act as a tireless tutor, offering instant explanations, generating practice problems, and even helping learners navigate dense research literature. This tension is not merely academic; it shapes how institutions design curricula, how teachers assess progress, and how learners allocate their limited study time. Recognizing that AI is neither a panacea nor a menace, but a powerful amplifier of existing learning dynamics, is the first step toward developing nuanced guidelines that harness its benefits while mitigating its pitfalls. The conversation must move beyond binary proclamations of “AI good” or “AI bad” and instead focus on how to integrate these tools into deliberate, reflective learning cycles that promote genuine mastery.
One of the most tangible drawbacks of unfettered AI access in education is the temptation to let the technology complete assignments wholesale, a practice that proves self‑defeating for genuine learning. When a learner delegates the entire problem‑solving process to a model, they bypass the struggle, the false starts, and the incremental insights that constitute the heart of skill acquisition. Cognitive science tells us that learning is forged in the effortful retrieval and reconstruction of knowledge; without that effort, the neural pathways that support long‑term retention and transfer remain weakly formed. Moreover, relying on AI for finished work erodes metacognitive awareness—the ability to judge one’s own understanding, identify gaps, and regulate study strategies. Over time, this can create a fragile competence that collapses when the AI safety net is removed, leaving learners unable to perform independently in real‑world settings where such assistance may not be available or permissible.
Yet the same technology that enables shortcuts also opens doors to supportive interventions that were previously costly or logistically challenging. Imagine a student grappling with a complex calculus concept at midnight; an AI tutor can patiently walk through the derivation step by step, adapt its explanations based on the learner’s responses, and generate an unlimited supply of similar problems for practice. In language learning, AI can provide instant, nuanced feedback on pronunciation or grammar, exposing learners to a breadth of idiomatic usage that textbooks often miss. Researchers benefit from AI’s ability to surface relevant papers, suggest methodological alternatives, and even help draft literature reviews, thereby accelerating the early stages of inquiry. When used as a supplement rather than a replacement, these capabilities can democratize access to high‑quality instructional support, particularly for learners in under‑resourced environments or those who lack access to expert human tutors.
Carl Hendrick’s incisive essay offers a useful analogy that clarifies why the problem is not AI itself but our relationship with it. He likens the rise of AI in education to the introduction of autopilot in aviation: while automation has undeniably saved lives by reducing pilot workload, it has also created scenarios where pilots, confronted with emergencies, instinctively reach for more automation instead of reverting to manual control. The danger lies in over‑dependency—when novices are given unrestricted access to powerful aids, they may never develop the foundational competencies needed to cope when the aid fails or is inappropriate. In the classroom, this manifests as students who can produce polished AI‑generated essays but falter when asked to articulate a reasoned argument without technological crutches. Hendrick’s insight shifts the blame from learners to educational systems that have yet to redesign assessments, feedback loops, and instructional scaffolding to accommodate these new tools responsibly.
Institutional responses have so far been marked by a troubling vacillation, swinging between enthusiastic endorsement of AI use—provided students “cite” the output—and a reflexive return to traditional academic honesty policies that treat any AI assistance as cheating. This back‑and‑forth fails to address the underlying pedagogical challenge: how to preserve the integrity of learning while embracing tools that can genuinely enhance it. Policies that merely police output without reshaping the learning process are likely to be ineffective, as students will find ever more sophisticated ways to conceal AI reliance. Sustainable progress requires a rethinking of learning objectives, assessment methods, and classroom activities to explicitly incorporate AI as a part of the learning workflow, rather than as an external contaminant or an illicit shortcut. Only then can schools move from reactive rule‑making to proactive design of learning experiences that are both rigorous and technologically enriched.
To navigate this complex terrain, it is helpful to view AI through the lens of familiar educational technologies such as calculators and Wikipedia. Calculators did not eliminate the need to understand arithmetic; instead, they shifted the focus from rote computation to higher‑order problem solving and conceptual reasoning. Likewise, Wikipedia democratized access to information but did not replace the critical skill of evaluating sources, synthesizing ideas, and constructing original arguments. AI follows a similar pattern: it can offload certain mechanical or informational tasks, thereby freeing cognitive resources for deeper engagement, analysis, and creativity. However, if learners allow AI to supplant the very cognitive work that builds expertise, they risk becoming adept at prompting the tool while remaining novices in the underlying discipline. The challenge, therefore, lies in delineating which aspects of a task can be safely delegated to AI and which must remain the learner’s own responsibility to preserve developmental integrity.
The phenomenon of “vibe coding”—using AI coding assistants to generate code snippets based on informal descriptions—illustrates precisely how the timing and extent of AI assistance can shape learning outcomes, particularly in technical domains. For beginners who have yet to internalize the mental models of syntax, control flow, and abstraction, leaning heavily on AI can rob them of the essential practice of writing, debugging, and refining code from scratch. This deprives them of the iterative feedback loops that cement understanding and develop problem‑solving intuition. Conversely, experienced programmers who possess a robust internal representation of code can use AI to accelerate boilerplate generation, explore alternative libraries, or prototype ideas quickly, thereby allocating more mental bandwidth to architectural design and innovation. The key variable is the learner’s stage of knowledge acquisition: AI assistance is most beneficial when it augments, rather than replaces, the cognitive work appropriate to that stage.
A more subtle but equally important question concerns the potential downsides of too little AI help. Traditional learning theories, such as those championed in the author’s earlier work Ultralearning, emphasized the value of grappling with high‑cognitive‑load problems as a pathway to deep mastery. However, subsequent research presented in Get Better at Anything suggests that when the cognitive load exceeds a learner’s current capacity, persistent struggle can become counterproductive, leading to frustration, disengagement, and the entrenchment of misconceptions. In such scenarios, a timely hint, a worked example, or a brief explanation—potentially delivered by an AI—can reduce extraneous load, restore a sense of progress, and enable the learner to return to independent practice with renewed clarity. Thus, the optimal use of AI appears to lie in a calibrated balance: enough support to prevent debilitating overload, but not so much that it erodes the necessity of personal effort.
Reflecting on the evolution of his own thinking, the author notes a shift from an absolutist stance on effortful problem solving to a more nuanced view that acknowledges the role of strategic assistance. In Ultralearning, the advocacy for tackling difficult problems head‑on was rooted in the belief that mental effort is the primary driver of skill acquisition. Yet the insights gathered for Get Better at Anything revealed that learning is not a monolithic process of sheer grit; it is also about managing cognitive load, leveraging feedback, and knowing when to seek guidance. This refined perspective does not diminish the importance of practice but rather situates it within a broader framework where external aids—including AI—can serve as catalysts that make practice more efficient and effective, provided they are deployed at the right moments and for the right purposes.
Based on this evolving understanding, a provisional set of guidelines for AI‑assisted learning can be sketched, pending more robust empirical evidence. First, learners should make a genuine attempt to solve a problem independently before consulting AI; this preserves the opportunity to engage with the material, confront misconceptions, and develop personal heuristics. Only after this initial effort should AI be invoked, and then specifically for targeted support—such as a hint, a worked example, or clarification of a specific stumbling block. Second, if AI assistance is used to overcome an impasse, the learner must subsequently tackle a comparable problem without AI to consolidate the gains and ensure that the understanding has been internalized. This cycle of attempt, support, and independent re‑attempt mirrors the principles of deliberate practice and helps prevent the formation of a crutch mentality.
Beyond basic problem solving, AI can serve as a powerful catalyst for expansive thinking and metacognitive reflection. When a learner’s initial approach yields a solution, querying AI for alternative methods, related concepts, or contrasting viewpoints can uncover blind spots and stimulate deeper exploration. Similarly, after completing an attempt, asking an AI to analyze the solution, point out inefficiencies, or suggest refinements turns the model into a feedback partner that highlights subtle errors and reinforces correct reasoning. These uses leverage AI’s vast knowledge base and pattern‑recognition abilities while keeping the learner firmly in the driver’s seat of sense‑making and judgment. By framing AI as a conversational partner rather than a surrogate executor, learners can enrich their educational experience without surrendering agency over their own intellectual development.
Effective learning, as articulated in Get Better at Anything, operates through a recurrent loop of seeing, doing, and receiving feedback. The “seeing” phase involves acquiring knowledge via instruction, worked examples, or demonstrations that prime the learner’s cognitive search for solutions. The “doing” phase entails active practice, where the learner applies that knowledge to generate responses, solve problems, or create artifacts. The “feedback” phase supplies information about the efficacy of that effort—whether through self‑assessment, peer review, instructor comments, or environmental outcomes. AI is uniquely positioned to enhance the seeing and feedback stages: it can generate tailored explanations, adaptive examples, and instant analytical feedback on submissions. What it cannot—and should not—replace is the indispensable doing phase, where the learner wrestles with the material, makes mistakes, and refines understanding through personal effort. Maintaining this balance ensures that AI serves as an accelerant rather than a substitute for the core cognitive work of learning.
In practice, learners and educators can translate these principles into concrete actions. Students should adopt a habit of attempting every assignment or practice problem on their own first, noting where they get stuck and why. Only then should they consult an AI for a hint or a worked example, using the response as a springboard for further independent work rather than a final answer. Educators, meanwhile, can design assignments that explicitly require stages of independent effort followed by optional AI‑mediated feedback, and they can assess not just the final product but the process of iteration and reflection. Institutions ought to invest in professional development that helps teachers recognize productive AI use patterns and craft rubrics that reward effort, adaptivity, and metacognitive awareness. By embedding AI within a structured learning loop that honors the necessity of personal doing, we can harness its power to make education more effective, equitable, and future‑ready.