Artificial Intelligence in Learning

What this is. Not a technical document about how to build AI systems, but a pedagogical one about what an AI placed at the service of learning does — and what it must avoid. It opens with the posture that governs it, and continues with fourteen design patterns and eight antipatterns. The premise: AI can dramatically enhance learning when applied with pedagogical criteria, and can produce shallow or false learning when applied as mere automation.

The Guiding Position — “The Almighty Gym”

Intelligence, like a muscle, grows with use and wastes away without it. AI is an engine that can be attached to it either to spare it or to strengthen it.

The analogy is the electric bicycle. It can be ridden as a lazy person’s shortcut — motor at maximum, no pedaling, arriving without effort and without exercise — or as a professional’s machine, built to develop the maximum potential of anyone willing to put in a little additional effort: climbing slopes that were closed off, going farther than one’s own condition allows, training longer than would otherwise be possible. Same bicycle, same motor, and a radically different human being at the end of the ride. The bicycle does not decide which of the two.

Two futures, on the same machine. The first is atrophy: leaning on AI even for basic thoughts, generating texts that are neither read nor understood, avoiding cognitive effort entirely. The second is AI as a cognitive gym: practicing extensively with immediate feedback, exploring more complex concepts because the fundamentals were mastered efficiently, receiving scaffolding that makes it possible to work right at the edge of what one can almost-but-not-yet do alone, sharpening judgement by evaluating and correcting what the AI produces. Both are genuinely possible. The difference between them is not in the AI: it is in the design of how the AI is used.

The almighty gym. We were handed a machine that can do almost anything, and that is exactly what makes it dangerous as a learning tool, not merely powerful: a gym where every machine lifts the weight for you is not, evidently, a place where anyone gets stronger. The two names are canonical and name different things: the almighty gym is the condition we were handed, danger included; the cognitive gym is the future that must be built out of it. The whole task of learning design in this era fits in one sentence: to make the almighty gym work as a cognitive gym.

Whose responsibility it is. This is the part that carries the most weight, and the one almost everybody inverts: responsibility for the right choice does not belong to the person who is learning. Ask a ten-year-old whether they want the easy task or the hard one. It is not a moral defect nor proof that this generation is lazier: it is human nature seeking the path of least resistance, and any adult does the same where nobody designs the friction for them. The responsibility belongs to the adults in the room — the Learning Facilitators (Orientadores de Aprendizaje — the role that accompanies the process of the person who is learning) and the institutions. If the essay can be produced without thinking and still earn the grade, it will be produced that way — and rightly so: we built an assessment that rewards coasting.

Smoother does not mean “frictionless.” Unproductive friction is removed — fighting with the tool, lacking a technique, logistical confusion — and productive cognitive friction is preserved, and sometimes increased: the effort of retrieval, difficulty calibrated to the edge of capability, the work of explaining, evaluating and correcting. AI lowers the cost of trying; education must use that lowering so that more is attempted, not less.

The answer is not a prediction. It is an assignment.

Fourteen patterns and eight antipatterns

Five general principles govern any AI function: the AI does not learn for the person who is learning (the criterion is whether the function transfers cognitive effort or takes it away); desirable difficulty is productive; scaffolding is progressive and designed to fade; there is transparency and traceability, and pedagogical AI never pretends to be human; and AI amplifies the Learning Facilitators, it does not replace them.

On top of those, fourteen patterns — among others: adaptive hints that reveal less the closer the solution is; feedback that meets the quality criteria; cognitive tutoring with a domain model and a model of the person who is learning; metacognitive monitoring in AI contexts; erroneous examples generated to measure; self-explanation prompt design; the transition from worked example to problem; Socratic tutoring dialogue; learning analytics that are actionable for the Learning Facilitator; individualized spacing; facilitated collaboration; a formative assessment cycle; digital worked-example sequences; and contextualized variant generation. Each pattern carries its own evidence label inside the document — strong, moderate or emerging — and none is upgraded.

And eight critical antipatterns: A1 on-demand generation without prior effort · A2 summaries that replace reading · A3 feigning comprehension where there is none · A4 personalization that isolates · A5 optimizing for engagement instead of for learning · A6 dependency that does not fade · A7 AI that claims to be human · A8 unexamined bias.

Evidence

Level: Adapted + synthesized, with a declared Original layer. Adapted — the patterns adapt the AI-and-learning-science domain of Gareth Manning’s open-source library education-agent-skills (CC BY-SA 4.0), and each pattern keeps its own label. Original / design position — the Guiding Position “The Almighty Gym” (Postura Rectora «El Gimnasio Todopoderoso»): a Smoother design position, authored by Carlos Miranda Levy, published in his article “Intelligence Is a Muscle: The Electric Bicycle and the Two Futures of AI in Education”. It is not an evidence claim: it is personal conviction and a design position, in the same register as the CEMI Impact Arc (Arco de Impacto CEMI), and it is never presented as validated pedagogy.

It converges — and only converges — with: desirable difficulties and the effort of retrieval; deliberate and purposeful practice; the Zone of Proximal Development and scaffolding. And it is consistent with the evidence, documented in the canon, that AI can raise immediate performance while learning lags behind, and that unrestricted access can damage it while scaffolded tutoring mitigates the damage. That evidence makes the crutch-future plausible; it does not prove the posture.

Permitted wording: “a Smoother design position, authored by Carlos Miranda Levy” · “aligns with” · “converges with” · “is consistent with”. Prohibited wording: “derived from” · “evidence-based” · “proven by” · “the research shows that AI is a cognitive gym” · presenting the electric bicycle or the gym as a research finding.

Memory and Learning · Learning Activities · Study and Practice Techniques · Experience Layer · Principles of Instructional Design · Evidence Framework · Challenges

Source

protocol/smoother-ai-en-aprendizaje.md — state of 2026-09-17 (open section of the CHANGELOG, after 0.40.0).