The False-Mastery Trap
This chapter examines the mechanism by which generative AI can improve a student’s immediate task performance while weakening the learning that assessment is supposed to certify. The focus is not whether a chatbot is “smart” or whether a student is “lazy,” but what changes in the sequence of attempting, monitoring, and revising a problem when an answer arrives too early. Understanding that sequence helps educators distinguish a productive scaffold from a shortcut and choose assessments that still reveal capability after the tool is gone.
When the answer arrives before the thinking
Learning is not identical to producing a correct response. A learner has to build or retrieve a mental model, test it against a problem, notice what does not fit, and revise the model when necessary. That work can be effortful. It can also feel inefficient next to a chatbot that supplies a fluent explanation or working solution in seconds. But the effort is often where the learner discovers what they know, what they do not know, and which decision caused an error.
Cognitive offloading occurs when part of that work is delegated to an external system. Offloading is not automatically harmful. People use reference books, calculators, search engines, compilers, and colleagues because tools extend what they can accomplish. The educational question is whether the delegated step is the capability being assessed, and whether the learner still performs enough of the surrounding reasoning to acquire it. A calculator can remove arithmetic while leaving mathematical modelling intact. A generated proof, essay, or program can remove the modelling, selection, explanation, and debugging that the assignment was meant to elicit.
The risky sequence is easy to recognise once it is made explicit. First, the student encounters a task and asks the system for a solution before diagnosing the problem. Second, the system returns an answer that is coherent enough to adopt. Third, the student implements or edits that answer with limited evaluation because the immediate goal—finishing the task—has been met. Finally, the polished product becomes the evidence shown to the instructor. The missing steps are not visible in the submission: the student may never have formed a tentative explanation, compared alternatives, or discovered an error through their own attempt.
That sequence changes metacognition, the learner’s monitoring and regulation of their own thinking. A student working with a human expert often has to identify the difficulty, ask a targeted question, consider the response, and decide what to do next. When a chatbot provides a ready-made solution, some students move directly from “I need help” to “I will use this.” Diagnosis, evaluation, and iteration are compressed or skipped. The phrase “metacognitive laziness,” used in the research synthesised by the OECD, describes this pattern of reduced monitoring; it is more useful as a description of a learning process than as a moral label.
The immediate product can improve precisely because the learner has been spared the productive difficulty. That creates a misleading signal. A higher score on the assisted task may mean that the tool supplied useful support, but it may also mean that the assessment measured the student’s ability to obtain and present an answer rather than their ability to reason through the problem. If the tool is removed, the unsupported part of the performance becomes visible.
The distinction matters for assessment because an instructor often observes the last state in the sequence, not the sequence itself. A final essay does not show whether the student began with an argument or selected one from a generated menu. A code submission does not show whether the student understood the function or merely iterated until the tests passed. A correct explanation does not establish that the student can produce or verify the explanation again. The product remains relevant, but it is no longer sufficient evidence when the system can perform the central cognitive steps.
What the Turkish mathematics experiment reveals
The clearest numerical example in the OECD Digital Education Outlook 2026 comes from a randomized field experiment with 1,000 students in grades 9 through 11 in Türkiye. Across six 90-minute mathematics sessions, students practiced under one of three conditions: their course notes and textbook, a general-purpose GPT-4 chatbot called “GPT Base,” or an educational version called “GPT Tutor.” The tutoring version was configured to avoid direct answers and support learning, although the report notes that it was not necessarily an adaptive learning system.
During practice, the AI groups did better than students working without GenAI. The reported improvement was 48% with the standard interface and 127% with the tutoring version. If the assessment stopped at the practice exercises, the conclusion would be straightforward: access to the tools made students more successful, and the tutoring configuration looked especially effective.
The researchers did not stop there. They later assessed knowledge in a closed-book environment. Students who had practiced with the general-purpose chatbot performed 17% worse than students who had studied without GenAI. Students who had used the tutoring version performed about as well as the self-study group. The result does not show that AI always damages learning, nor does it show that the tutoring version produced a durable advantage. It shows something more precise: in this population, subject, interface, and study design, success while the tool was available did not predict success after it was removed. The general-purpose answer engine improved the visible practice result while leaving a gap in independent performance.
That gap is the false-mastery trap. The learner, instructor, or institution sees a correct answer and infers a capability that the task did not actually observe. The inference feels reasonable because the product is real. The problem is that the cause of the product has changed. The student may have learned something, but the assessment cannot tell how much of the result came from the student’s knowledge and how much came from the system’s ability to supply the missing steps.
Other studies in the OECD synthesis point to the same distinction without producing identical numbers. In an essay-revision experiment with Chinese higher-education students, the GenAI group achieved the highest task performance, yet the knowledge test did not show corresponding gains. Students in that condition performed fewer metacognitive activities, especially evaluation and orientation. A separate comparison of human advice and chatbot advice found that students receiving human help tended to diagnose their need, ask, evaluate, iterate, and then implement. Some students using the chatbot asked directly for help and implemented the received solution, skipping the middle of the sequence.
A writing study conducted with students from five universities gives the difference another form. Participants wrote a 20-minute essay alone, with a search engine, or with ChatGPT. The essays produced with the LLM were well rated, but only 12% of that group could quote something from their own essay within an hour, compared with 89% in the other two groups. The LLM group also showed lower ability to summarise the essay’s viewpoint, lower ownership, and more similar content across essays. Those results come from one study and should not be treated as a universal measurement of memory or authorship. They do, however, illustrate why a polished object and a learner’s command of that object are different variables.
A simple analogy clarifies the assessment consequence. Suppose someone can deliver a speech by reading a teleprompter but cannot answer a follow-up question about its argument. The speech may be grammatically excellent. It is still weak evidence of spontaneous understanding. Generative AI is a more powerful and flexible teleprompter, able to generate not only words but reasoning-shaped content. The more of the target performance it supplies, the more an output-only assessment risks certifying access to assistance.
Scaffold or shortcut?
The presence of AI does not determine the learning outcome by itself. The interaction design determines what the student must still do. A shortcut starts with substitution: “Solve this for me.” A scaffold starts with a learner who has to make a commitment—an attempt, a diagnosis, a prediction, or a question—and then uses support to extend that effort. In the first pattern, the tool reduces the need to think. In the second, it gives the learner something to think with.
A pedagogical scaffold can take several forms. The system may ask a question instead of giving the next step. It may point to a relevant concept, request a justification, or offer a hint after the student has attempted the task. It may ask the student to compare two approaches, identify an error, or explain why a result is plausible. The point is not to make the AI evasive for its own sake. The point is to keep the cognitive operation being learned in the student’s hands.
The OECD reports promising results from interventions built around this principle. In a Harvard undergraduate physics course, a GPT-based tutor configured to implement active-learning practices produced more learning in less time than an in-person active-learning comparison; students also reported higher motivation and engagement. In a Nigeria after-school programme, students worked in pairs, received teacher guidance on how to use Microsoft Copilot, and achieved a positive medium-sized learning effect. These studies tested particular pedagogies and populations. Their results should not be simplified into “AI tutoring works.” They suggest instead that the relevant intervention is a combination of tool, instructional method, social setting, and student activity.
The same caution applies to the label “educational.” A chatbot can be described as a tutor while still behaving like an answer engine. Effective intelligent tutoring systems traditionally model the learner, the domain, and the tutoring strategy, then adapt support to the student’s current state. Many newer LLM-based systems rely heavily on the language model to perform all of those roles. The OECD notes limited evidence about how reliably such systems maintain a tutoring strategy, prevent hallucinations, or adapt to individual knowledge over time. A friendly tone is not a pedagogical architecture.
One practical test is to ask what happens at each turn. Before the AI responds, has the student attempted anything that can be evaluated? Does the response require a decision, explanation, or verification from the student? Is assistance gradually withdrawn, or does it become the default producer of the work? Can the student later perform the central operation on a new problem without the tool? These questions assess the learning mechanism rather than the system’s marketing category.
Another test is to inspect the order of cognitive activation. The OECD’s summary of the writing research suggests that writing independently first and using an LLM later preserved more activation and recall than starting with the LLM and then continuing alone. This does not establish a universal “always write first” rule. It does indicate that sequence matters. If independent thought is postponed until after the machine has proposed the structure, the student may spend the rest of the task supervising generated content rather than generating and evaluating ideas.
That is why an assessment designed for an AI-rich environment should include a point at which the student has to show the capability without being carried by the tool. The check might be closed-book, but it need not always be. It could ask the student to explain a choice, modify a solution under a changed condition, critique an AI response, reproduce the core method on a new case, or reflect on why one suggestion was rejected. The right check depends on the construct. What matters is that the student must exercise the human capability the original task was supposed to develop.
The two paths in the diagram are not two kinds of student. They are two possible designs of the same interaction. A learner can move between them from task to task. That variability is exactly why a label such as “AI user” tells an assessor very little. The important evidence is what the student did with the assistance, what they learned from it, and what they can still do when the assistance no longer supplies the answer.
A credible assessment therefore treats the removal of AI as a question about inference, not a punishment. If the goal is to certify independent recall, an unaided component is appropriate. If the goal is to certify responsible work with AI, the student should still have to select, verify, explain, and defend the result. If the goal is transfer, a new situation is more informative than another copy of the assisted exercise. In each case, the assessment must make the intended human capability visible.
The next challenge is to design that visibility systematically. Once performance can diverge from learning, educators need more than a general suspicion that the process matters. They need a way to connect a claim about what a student can do to the task, observations, and evidence that justify the claim.