Understand
Encode
Relates what is new to previous knowledge and distinguishes the idea from its examples.
Case studies · Students
Learning is not delivering.
It's being able to do it again.
AI can explain, adapt, question and expand practice. The limit appears when it improves the product while replacing the effort that builds understanding, memory and judgment. This case study relates the help to the educational stage, prior knowledge and the real learning objective.
Assisted performance does not equal learning. The useful test is what you can explain, adapt and do again when the help disappears.
Support overview
Secondary
Abstraction increases, but self-regulation and critical reading continue to develop. Ease can be confused with mastery.
Reading
Guidance pedagogical priorities; They are not scores or a clinical prescription.
Learning architecture
The answer is the product.
Learning is changing the system.
Learning leaves a more stable capacity: recognizing, recovering, explaining and using an idea in another context. AI benefits that process when it creates good feedback and more practice; it harms it when it makes reasoning invisible.
Learningtransferiblewith and without AI
01Encode
02Recover
03Explain
04Transfer
Remember
Recover
Try to bring the information without seeing it. The retrieval effort modifies memory.
Make visible
Explain
Put into words why each step works and where it might fail.
Generalize
Transfer
Apply what you have learned to a problem with a different surface, data, or context.
Relevant differences to learn
The same help does not serve all students or in all tasks.
This case study does not classify “types of intelligence.” It looks at six observable conditions that change how AI should be introduced: what the person already knows, what barrier they encounter, what they need to practice, and how they can later demonstrate that they have learned.
Prior knowledge
Those who already know the subject can detect omissions; those who begin may confuse a clear explanation with a complete explanation.
Adjust help
Ask for a prediction or a trial first and grade the clues according to the errors observed.
Find out
Explain and resolve a variant without consulting.
Attention and working memory
A task can be overwhelming due to its format even if the concept is accessible. Split help; converting each step into an automatic response prevents learning to coordinate them.
Adjust help
Reduce noise, show structure and remove support in phases.
Find out
Reconstruct the entire procedure with less and less guidance.
Motivation and self-efficacy
Early help can reinforce “I can't”; A tight track allows progress to be attributed to one's own decisions.
Adjust help
Set achievable challenges, record progress and give feedback on strategy, not a fixed identity.
Find out
Choose and justify the next step in the face of a new blockage.
Language and accessibility
Changing language, size, modality or syntactic complexity can open access without lowering the intellectual objective.
Adjust help
Modify the format barrier and preserve the vocabulary or reasoning that you want to learn.
Find out
Demonstrate understanding in an accessible and agreed upon modality.
Neurodiversity and health
ADHD, dyslexia, autism, disability, anxiety, sleep or fatigue do not produce a single profile or authorize an AI-generated diagnosis.
Adjust help
Agree on support with the student and, when appropriate, with teachers and specialized professionals.
Find out
Observe participation, understanding and autonomy, not just speed.
Context and support available
Time, device, privacy, home language, and support change what a recommendation allows you to do in practice.
Adjust help
Offer an alternative without AI and not penalize those who have fewer resources.
Find out
Check that the evaluation measures learning and not technological access.
The theory is in the common base
Intelligence, constructs and individual differences are explained in the Theoretical Foundations.
Here they are only translated into teaching and evaluation decisions so as not to mix common theory with application to the collective.
Design the type of help
Don't just ask “do I use AI?”
Ask what role he occupies.
The same technology can solve, tutor, discuss or simulate. The mode of interaction determines how much reasoning remains the student's.
Track Tutor
AI regulates aid
It guides without immediately revealing the solution and returns the cognitive work to the student.
Petition to change the distribution
“Don't give me the answer. Ask one question at a time and offer a hint only after my attempt.”
- Fits better
- Acquire procedures, detect a blockage and practice with graduated difficulty.
- Risk
- Hints can become a piecemeal solution if asked too soon.
Orientative profile
Transfer82
own effort76
Helpful feedback88
Pedagogical illustration, not the result of a validated scale.Field experiment · almost 1,000 students
Help can elevate practice and reduce subsequent learning.
In high school mathematics, Bastani and colleagues compared access to GPT-4 without guards, a cued tutor, and a control group. The effect depended on the design.
PNAS, 2025 · complete study ↗During practice with AIgrade improvement
Post-exam without AIversus control
gpt tutor
It does not prove that all AI harms nor that these percentages are reproduced in other ages, subjects or tools.Safeguards largely mitigated the negative effect.
Age, level and previous knowledge
Age guides.
Expertise in the task decides.
Stages are starting points, not rigid boundaries. Within the same age there are enormous differences in development, experience, support, motivation and educational needs.
Visible scaffolding
Secondary
Abstraction increases, but self-regulation and critical reading continue to develop. Ease can be confused with mastery.
Graded clues, comparison of explanations, recovery questions and practice with feedback.
First attempt, intermediate reasoning, extensive writing and checking sources.
After using AI, remake the problem or argument with different data and justify each decision.
13
Access age does not equate to pedagogical maturity. UNESCO's 2023 guidance proposed a 13-year-old threshold for independent conversations with generative AI platforms, along with data protection and age-appropriate design. The applicable regulations, conditions of service and center policy must prevail.
UNESCO ↗Decision laboratory
Design help based on what you need to keep.
This tool does not measure dependency or diagnose learning. Turn four observable variables into an editorial recommendation for deciding when to introduce AI.
Guidance for this task
94/100protection
cognitiva
cognitiva
First without AI; later, track tutor
Make a full attempt or limit the block. Ask for a single clue, continue yourself and finish with a variant without help.
Ruler
The more you need transferir and the less you can verificar, the solution should appear later.
Brief protocol
I → Track → Reconstruct → Transfer
I try and mark the exact blocking point.
I'm asking for a question or clue, not the complete solution.
I close the AI and redo the reasoning with my words.
I resolve a new variant and check the error.
Calibrated trust
It's not enough to get it right.
It's good to know how much you know.
Judgment improves when you separate response and confidence, receive feedback, and record when you were very sure about something false. Practice with three affirmations.
Affirmation 1 de 3
If an answer seems clear and contains links, it is already verified.
?
First decide and calibrate.
The explanation will appear after you set your answer and your confidence level.
predict
Before asking the AI, write down what result you expect and why.
separate
Mark facts, inferences, preferences, and uncertainties with different labels.
Contrast
Find a primary source and a reason why your interpretation might be wrong.
Register
Keep high-confidence errors: they are the most useful map of your judgment.
Learning conditions
The same tool does not create the same opportunity.
The effect depends on resources, time, language, accessibility, quality of support, prior knowledge and the possibility of paying for better tools. “Personalizing” is not enough if you do not observe who is left behind.
AIAccess ≠ benefitThe context transforms the result
Connectivity
Stable device, data, space and time.
Language
Uneven quality between languages, registers and varieties.
Accessibility
Real compatibility with sensory, motor and cognitive needs.
Educational capital
Know how to ask, verify and recognize a good explanation.
Human guidance
Teachers, family, peers and human feedback available.
Power of choice
Being able to reject the tool without losing opportunities.
Educational technology at scale not generative AI
The gains can be real, modest and unequal.
A 2026 study of about 200,000 students estimated a gain of 0.031 standard deviations with about 6.6 hours of Khan Academy per year — about 11 minutes per week — versus not using it. It projected 0.085 SD at 30 minutes per week. Higher-performing students benefited more, in part because they used the platform more and advanced through more skills.
Eames et al., PNAS 2026 ↗00,030,060.09 SD
Observed use · 11 min/week+0,031
Recommended use · 30 min/week+0,085
Questions for centers and teachers
□
Does the activity also work with a free tool or without a personal account?
□
Is there an equivalent alternative for those who cannot or do not want to use AI?
□
Is verification taught or does it presuppose a competence that not everyone possesses?
□
Does the evaluation reveal the process and transfer, not just the quality of the product?
Evidence and limits
What we know, what we infer and what we need to measure.
The evidence on memory and learning is extensive; the causal evidence on generative AI in education is still recent, specific and heterogeneous. That is why design, sample and limits are cited.
01Field experiment · Generative AI
Generative AI without guardrails can harm learning
In high school math, the help architecture changed the outcome: producing better with AI did not guarantee performing better when it was retired.
Bastani et al. · PNAS, 2025 ↗02Experiment · memory and understandingRetrieval practice produces more learning than elaborative studying
Retrieving and reconstructing knowledge improved conceptual learning and inference compared to elaborative study with concept maps.
Karpicke & Blunt · Science, 2011 ↗03Quantitative review · 317 experimentsDistributed practice in verbal recall tasks
Spacing practice is a robust phenomenon; the useful interval depends on how long you want to preserve the learning.
Cepeda et al. · Psychological Bulletin, 2006 ↗04Review · study techniquesImproving Students’ Learning With Effective Learning Techniques
Test practice and distributed practice were rated as highly useful across different ages and abilities.
Dunlosky et al. · Psychological Science in the Public Interest, 2013 ↗05Study at scale · digital learningHow Khan Academy influences student math learning
Sustained use was associated with modest gains; the benefit was uneven and also depended on effective use and progress.
Eames et al. · PNAS, 2026 ↗06International guide · rights and designGuidance for generative AI in education and research
It proposes a humane, age-appropriate approach, with data protection, ethical validation and pedagogical design.
UNESCO · 2023, actualizada en 2026 ↗!
We don't turn a correlation into a brain injury.
Poorer subsequent performance, reduced effort, or an association with more use do not in themselves demonstrate general cognitive impairment. Measures without assistance, temporal monitoring and comparison with real alternatives are needed.