Students who learn with AI progress faster, then struggle when it disappears. The OECD documents this in 2026: AI improves feedback during learning, but the student who becomes too dependent arrives at evaluation unprepared, without the habit of working independently. The problem affects several groups of students, notably those whose educational resources outside AI are limited.

The essentials

  • Unguided use of AI can improve performance while risking reduced cognitive effort and lasting learning outcomes (OECD Digital Education Outlook 2026).
  • When access to AI is removed, the performance advantage observed with the tool can disappear or reverse.
  • An EDUCAUSE 2026 survey of 438 teachers confirms that expectations are shifting, but measuring actual learning outcomes remains difficult.
  • In Latin America, where access to digital tools is profoundly unequal, the effect could amplify already-documented gaps.
  • The solution is not to eliminate the tool: it is to teach students to progressively detach from it before exams.

Feedback without learning the effort

This mechanism of dependency touches on the structure of cognitive effort that a student consents to provide, not merely the amount of feedback received. The decisive point is allowing an attempt at retrieval before providing feedback. Immediate feedback after a retrieval attempt can improve consolidation, depending on the task and setup. The problem lies in the timing of feedback, not feedback itself. The decisive point is allowing a retrieval attempt before feedback, not systematically avoiding immediate feedback.

The student progresses on exercises and answers are more often correct, but this performance rests on external guidance rather than on internalized knowledge.

A gap between performance on assisted homework and performance on exams can occur, but its invisibility during learning is not established. The teacher observes satisfactory results, the student feels competent, and the conditions of this learning do not seem likely to change. It is only at evaluation, when the safety net disappears, that the gap reveals itself. The structural problem is that planned withdrawal of certain pedagogical aids can promote transfer, without confirming a generalized progressive dependency on AI tools. A pedagogical tool effective in the short term can thus be counterproductive in the medium term if no protocol provides for progressively reducing its support.

This logic applies to any support system that does not integrate, in its very design, a withdrawal strategy. AI stands out from other tools by the scale at which it operates and by the discretion with which it imposes itself on work habits, which makes withdrawal all the more difficult to organize without explicit pedagogical decision.

A good teacher corrects an error when it occurs. AI does the same, at scale, without getting tired. That is a real gain. The OECD synthesizes evaluations indicating that adaptive tutorial systems can support learning. On this point, the enthusiasm of EdTech promoters rests on something solid.

But the same report points to a more troubling mechanism. Frequent feedback could reduce the student’s effort to predict errors. It could produce reduced effort in retrieval—this attempt to find an answer from memory without aid, a process that cognitive science identifies as a driver of lasting memory. Learning progresses in terms of immediate performance. The capacity to mobilize one’s knowledge independently stagnates.

Poorly calibrated AI use produces tool dependency. A student who never does math without a calculator can still learn math, provided their teacher knows when to remove the calculator. The same logic applies to AI: explicit pedagogical guidance must define who reduces access to the tool, when, and for what purpose.

At evaluation, everyone brings what they brought

In several school systems, exams place students without access to AI, creating a contrast with daily assisted practices. It is in this gap between daily practice and evaluation conditions that the problem described by the OECD resides.

A student who has learned to work with a constant correction tool arrives at the exam with real competence, but also with a habit of dependency. If this habit has not been addressed, if no one has taught them to function without a safety net, the gap between their learning performance and evaluation results widens.

The EDUCAUSE 2026 survey, conducted with 438 teachers, sheds light on the institutional face of the problem. Teachers observe that student expectations have changed: they expect more guidance, more feedback, more interactivity. But measuring whether these students have truly acquired the targeted competencies remains difficult. Assessment instruments have not evolved at the same pace as learning practices. The gap between the two creates a zone of uncertainty that no one, for now, knows how to truly bridge.

Latin America as a limiting case

This gap between assisted learning and solitary evaluation exists everywhere. In Latin America, it takes a particularly acute form, for a simple reason: access to tools is not uniform.

Some schools, particularly private ones or those in well-equipped urban areas, have been using AI for some time. They develop uses, strategies, habits. In less well-equipped schools, the tool arrives late, sporadically, or not at all. These students do not have the problem of overreliance; they simply lack the tool.

This fracture operates along distinct lines. Some students have learned to depend on a tool absent on exam day. Others have worked without benefiting from the accelerated feedback that AI provides. These groups face different challenges. Quality feedback can improve learning in certain contexts, but its effect on resource inequalities must be established separately.

This is a scenario that labor market automation reproduces at another scale: technology improves conditions for the best positioned, without reaching those who need it most. The mechanism is different; the direction of inequality, the same.

The Inter-American Development Bank has documented these dynamics in its work on EdTech in Latin America. Its analyses emphasize that unequal adoption of educational AI tools, absent intentional deployment in disadvantaged schools, could amplify result gaps between urban and rural areas, between private and public sectors.

Teacher statements and the limits of measurement tools

The measurement difficulty that surveyed teachers point to refers to a problem that extends beyond available instruments alone. Current tools pose challenges in distinguishing assisted performance from competence, but some assessments are designed to measure learning processes as well. A student can produce quality work in an assisted environment without that quality being transferable to a different context.

Instruments designed before the generalization of educational AI were not conceived to establish this distinction. Adapting them would require first clarifying what one seeks to evaluate—the capacity to mobilize knowledge without assistance—which implies rethinking the very purpose of assisted learning itself. This circularity explains why educational systems hesitate to reform assessment methods while simultaneously deploying new tools.

Modifying both simultaneously requires institutional coherence that most systems are not organized to produce. Teachers thus find themselves in an uncomfortable position: they use tools whose benefits they perceive without being able to verify their actual effect on lasting learning outcomes, and they continue to evaluate with instruments that do not give them the information they would need to adjust their practices. This uncertainty has consequences for daily pedagogical decisions. For lack of reliable data, some teachers maintain high levels of assistance to preserve visible performance. The OECD recommends selective and pedagogically grounded use of AI so that it does not replace cognitive effort.

The EDUCAUSE survey points to a tension that adoption figures willingly mask. The 438 teachers surveyed note that their practices have changed—more feedback, more personalization, more interactive sequences. They also observe that their students engage differently, sometimes more, sometimes more passively.

The EDUCAUSE survey highlights persistent uncertainties about assessing learning with AI in higher education. Evaluations without AI can measure the ability to excel without AI, even if they do not cover all complex competencies. Some evaluations can be conducted with AI, which complicates performance interpretation. This is a pedagogical design problem. Some systems integrate AI with governance, monitoring, or assessment adaptations, but practices remain heterogeneous.

Schools can harness demographic and technological shifts provided they anticipate them rather than suffer them. On AI, several Latin American school systems still lack sufficient articulation between pedagogical use and preparation for evaluative autonomy.

Emerging approaches, cautiously

Some Latin American schools are testing a progressive withdrawal approach. AI is available at the beginning of a learning unit, then its assistance is gradually reduced. As evaluation approaches, students work without assistance. The idea: to make weaning a competency in itself, not a shock experienced on exam day.

Results are preliminary and sample sizes small. But the logic is consistent with what cognitive science has long documented about retrieval practice. The OECD recommends selective AI use that preserves independent thinking and core competencies.

A teacher can decide alone to reduce AI access in the weeks preceding evaluation, regardless of available budget. The training necessary to act with intention and institutional willingness to recognize the issue are lacking. In the region, several educational policies prioritize tool deployment. Reflection on proper use, including non-use at the right time, comes afterward.

Measures within reach of decision-makers

The window to correct course is open. Educational AI is still in its early deployment years in Latin America. Habits are not yet fixed. Systems that act now can design from the outset a pedagogical architecture that integrates both assistance and autonomy.

Three levers seem concrete. First, train teachers to calibrate AI use according to the learning phase, not merely to use the tools. Next, revise assessment instruments so they measure competencies that AI cannot substitute: reasoning under constraints, mobilization without guidance, oral argumentation. Finally, concentrate tool deployment as a priority in disadvantaged schools, where AI feedback fills a real void rather than adding to already abundant resources.

These directions are not new in pedagogical debate. What is new is the pressure from OECD data to take them seriously before the gap becomes entrenched. A school system that waits for inequalities to be documented at scale to act has already produced them. The opportune moment is now, while uses are still taking shape.

AI improves learning under conditions that allow thoughtful use: this point is established. The challenge for Latin American decision-makers is to determine whether the region’s educational systems will train their teachers to manage this nuance, or will merely equip their classrooms.

Sources

  1. OECD, Digital Education Outlook 2026, January 2026, https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/01/oecd-digital-education-outlook-2026_940e0dd8/062a7394-en.pdf
  2. EDUCAUSE, Assessment Study 2026 (438 teachers surveyed), cited in Digital Education Outlook 2026; direct URL not available, accessible via educause.edu
  3. Inter-American Development Bank, work on EdTech in Latin America, available at iadb.org