Product ≠ learning
Assisted quality and subsequent ability are different outcomes.
Theoretical foundations
How a mind learns that also uses tools.
This section brings together the concepts that all case studies need: what changes in the brain when we learn, how abilities evolve throughout life, what “intelligence” means, why people respond differently, and what allows the evidence about AI, screens, and social media to really be affirmed.
central idea
AI does not act directly on a single “intelligence.” Modify an activity: change where we pay attention, what we remember, how much we practice, how we check, and what part of the process remains under our control.
Context
Tool
Task
Analysis unitPersonin activity
We don't just ask
“What does AI do?”We ask what changes in this system.Age ≠ expertise
We can be experts in one field and beginners in another task.
Use ≠ dependency
Frequency alone does not reveal loss of autonomy.
Association ≠ cause
A statistical relationship does not identify the mechanism that produces it.
Brain, cognition and learning
The brain does not download files.
It reconstructs, predicts and changes through practice.
Learning modifies the availability and organization of knowledge. For an explanation to become one's own capacity, the person must attend, relate, recover and use. Subjective ease may accompany this process, but it does not demonstrate it.
Attention
Select which information receives deep processing. It is limited, it fluctuates with sleep, stress, motivation and interruptions. AI can reduce noise or become an additional source of fragmentation.
Working memory
It maintains and transforms a small amount of information while we reason. Breaking up a task can help; anticipating all the steps can eliminate the practice of coordinating them later.
Long-term memory
It is not a passive warehouse. Organized knowledge allows you to recognize patterns, understand explanations, and detect errors. Consulting information is not equivalent to integrating it into a recoverable network.
Executive functions
Inhibiting a response, updating information, and changing strategies support planning and self-control. They develop over years and also depend on the situation, not just a supposed force of will.
Metacognition
It is observing and regulating one's own knowledge: estimating confidence, detecting gaps and choosing a strategy. A fluent response can increase the sense of mastery without improving its accuracy.
Learning and transfer
Learning involves a relatively stable change that appears on a subsequent test or in a new case. Performance while the AI is present measures something else: the person-tool system.
From exposure to transfer
Four operations convert available information into usable learning.
They do not always occur in a neat sequence, but they help identify which part AI can support and which part should not be skipped.
Select and hold focus.
Connect, distinguish and explain.
Rebuild without looking.
Use in a new situation.
What AI can contribute
Frequent feedback, adjusted examples, recovery questions, translation, accessibility and varied practice.
What can move
Diamond, 2013 ↗Karpicke & Blunt, 2011 ↗Formulate the problem, tolerate uncertainty, retrieve from memory, construct an explanation and detect an exception.
Development throughout life
There is no brain finished at 18.
There are different trajectories that continue to change.
Some functions mature slowly, others rely increasingly on knowledge, and all are sensitive to health, education, practice, and context. The ages are indicative intervals: they do not allow us to diagnose a person's preparation.
Early childhood
Language, play, shared regulation, sensory exploration and bonding build foundations. Contingent interaction with people and objects has a richness that a generated output does not reproduce.
School childhood
Reading, writing, calculation, knowledge and executive control are consolidated. Help is safer when it accompanies a guided experience and makes the child's effort visible.
Adolescence
Abstract reasoning grows, but self-regulation and social sensitivity continue to change. Accompaniment moves from controlling to teaching criteria and gradual autonomy.
Emerging adulthood
Advanced training and entry to work require building models of the profession. Skipping initial tasks can improve production while impoverishing the experience that supports future judgment.
Adulthood
Knowledge and experience allow you to take better advantage of AI, although convenience can make small losses of practice invisible. Continuing to learn requires recovering, explaining, and resolving exceptions.
Older adulthood
The trajectories are very diverse. Technology can support accessibility, autonomy and connection; there is no basis for claiming that everyday digital use alone causes "digital dementia."
Relative trajectories
Capacities do not share a single maximum.
Speed and some smooth tasksKnowledge and vocabularyRegulation and experience
The lines show general patterns, not exact scores or ages. There is great variability between tasks and people.
The already formed adult
Experience protects judgment, but it does not eliminate the need for practice.
An expert person has models, vocabulary and case memory to evaluate an answer. This allows you to delegate with more security. The risk is gradual: if for months the AI always formulates, summarizes, decides or remembers for it, the absence of visible errors can hide which operations are exercised less.
There is no evidence to say that using AI inevitably produces deterioration or atrophy. Yes, there is a functional reason to check non-delegable skills: what is no longer practiced can become less accessible, while active learning, socially, physically and cognitively demanding activity promote healthy aging.
Useful signalStable productivity + less ability to explain, correct or act without help.
Hartshorne & Germine, 2015 ↗Intelligence and constructs
"Intelligence" does not name a single thing.
And theories are not interchangeable.
An IQ test summarizes performance on a sample of tasks under specific conditions. It is useful for certain questions, but it does not exhaust creativity, situated knowledge, personality, values, experience, health or opportunities. Talking about "types of intelligence" without specifying the model mixes constructs that are defined and measured in different ways.
A level map
Capacity, ability and performance are not synonyms.
AI intervenes on performance in an activity. Inferring from there a stable increase or decrease in intelligence requires independent measurement, time and transfer.
General covariationgstatistical factor
Broad skillsGf · Gc · memory · visual · speed…
Specific knowledge and skillsmathematics · language · craft · mastery
Observed performanceperson × task × tool × context
General capacity and hierarchical models
Correlations between cognitive tasks support a general factor, while models such as CHC describe broad and specific aptitudes. They are psychometric covariation models, not a complete description of the person.
Fluid and crystallized intelligence
They distinguish, in a simplified way, reasoning in the face of new problems and acquired knowledge. They interact: a rich base makes it easier to think and reasoning allows you to continue building it.
Executive functions and working memory
They are related to intellectual performance, but are not synonymous with intelligence. They describe control, maintenance and flexibility mechanisms relevant to learning and acting.
Metacognition
It includes knowledge and regulation of one's own thinking. Its practical value is high: an autonomous person not only responds, he also knows when to doubt, verify or ask for help.
Creativity
It integrates generation of alternatives, originality, knowledge, selection and social context. It's not a one-size-fits-all bar for AI to raise: it can expand on ideas and also anchor on your first examples.
Emotional intelligence
Ability, trait, and competency models do not measure exactly the same thing. It is important to specify the definition before interpreting a score or promising that a tool improves it.
Practical and successful intelligence
Sternberg's proposals emphasize adaptation, creativity and situated resolution. They provide valuable educational questions, although their psychometric separation from other abilities is debated.
Multiple intelligences
Gardner helped recognize diversity of talents, but the evidence does not support eight independent intelligences nor does it allow us to deduce that each student should be taught according to a "type."
AI by functional dimension
The same aid can benefit one operation and displace another.
The decisive column is not "can help" or "can hurt", but how capacity will be proven when the help disappears.
Dimensioncan supportYou can moveTesting
Knowledge
Explain vocabulary, create examples and connect ideas with a previous base.
Replace reading, recovery and construction of own schemes.
Explain and apply without consulting.
Reasoning
Propose counterexamples, variations and graduated clues.
Reveal the structure of the problem before the attempt.
Solve a new variant.
Memory and attention
Organize information and reduce accessory load.
Always outsource steps that must be coordinated by heart.
Rebuild the procedure.
Language
Give feedback, adapt registration and support accessibility.
Uniform your voice and practice your own formulation less.
Defend why each change was accepted.
Creativity
Expand the space of alternatives and combine perspectives.
Anchor the search in probable or conventional examples.
Ideate before and compare final diversity.
Judgment and metacognition
Make assumptions, objections and confidence levels visible.
Inherit the security of a convincing answer.
Calibrate trust with accuracy and sources.
Precision
Useful practice can target memory, reasoning, language, or creativity without assuming that there are fixed “learning styles.” Adapting representation can improve access; pairing instruction with a visual, auditory, or kinesthetic label has no equivalent support.
Critical review of multiple intelligences ↗Individual differences
There is no such thing as an average user.
There is a person faced with a specific task.
Individual differences do not justify putting people into boxes. They serve to anticipate when aid reduces a barrier, when it eliminates a valuable practice, and who can verify what they receive.
Same AI output“Here is a clear summary”The effect depends on what is happening around.
expert person
Compare with your model, detect omissions and save time.
Probable amplificationbeginner person
He recognizes the words, but he doesn't know what is missing or what is debatable.
Apparent understandingPerson with access barrier
A linguistic or format adaptation allows you to enter the content.
Valuable accessibilityPrior knowledge
It is one of the most important moderators. An expert can use a synthesis to work faster and spot errors; a beginner may accept the same synthesis without the knowledge needed to evaluate it.
Design judgementAdjust AI to expertise in that task, not degree or age.
Attention, language and accessibility
Simplification, a change of format, or a supporting reader can free up resources for the core goal. The benefit disappears if the adaptation unnecessarily reduces the complexity to be learned.
Design judgementRemove only the accessory barrier and preserve the cognitive objective.
Motivation and self-efficacy
Believing that effort can produce progress changes persistence. Too early help can confirm "I can't"; a calibrated cue can turn a block into observable progress.
Design judgementDesign successes attributable to the person's decisions.
Neurodiversity and health
There is no universal prescription for ADHD, dyslexia, autism, disability, anxiety, fatigue or cognitive impairment. The needs are heterogeneous and one interface does not perform a valid clinical evaluation.
Design judgementCustomize with the person and, where appropriate, with professionals.
Language, culture and support
The quality of an answer depends on the data and the language; also whether there is someone capable of accompanying or contrasting. Apparent personalization may reproduce biases or ignore local context.
Design judgementCheck representation, understanding and cultural relevance.
Time, privacy and cost
Having access does not guarantee being able to use a tool well. Pressure, precarious employment, caregiving or a shared device modify which practices are realistic and who assumes the risk of an error.
Design judgementEvaluate autonomy with real alternatives and conditions.
General rule
The less the prior knowledge and the greater the consequence of the error, the longer the complete response must be delayed and mediation, traceability and independent verification reinforced.
AI and cognitive offloading
Downloading is not stopping thinking.
It's redistributing where work happens.
Taking notes, using a map, or saving a contact are normal forms of cognitive offloading. The question is not whether we outsource, but whether the strategy improves the objective and preserves the capacity necessary to monitor, recover or act when help fails.
Record
The tool stores information, but the person decides what it represents.
Organize
Reduces accessory load and keeps steps visible.
Guide
The AI asks questions, gives clues and adapts difficulty.
Trigger
Person and system alternate proposals, criticism and review.
Solve
The tool produces much of the result.
Adaptive download
Free up resources for a higher value goal.
- The goal is still human.
- The output can be verified.
- An alternative is preserved in the event of failure.
- The delegated task was not the object of learning.
- The decision and responsibility remain clear.
Fragility pattern
Aid replaces a still necessary capacity.
- Consult before formulating.
- Fluency is confused with accuracy.
- No steps or sources are preserved.
- Withdrawal produces disproportionate blocking.
- Practice is lost in exceptions and errors.
Transfer of control
More delegation requires more verifiability.
difficult to verifyEasy to verify
Protect practiceUse scaffoldingReversible delegation↑ Major consequence of the error
Conceptual decision diagram, not clinical scale.
Dependency is not an automatic effect of outsourcing. It arises when the strategy becomes rigid, is not calibrated with one's own knowledge, or eliminates the opportunities necessary to maintain competence.
Risko & Gilbert, 2016 ↗Screens, social media and IQ
“Screen time” mixes activities that are not cognitively the same.
Reading, creating, chatting, gaming, studying, scrolling through videos, or using an accessibility aid can all happen on the same screen. Duration, content, design, timing, sleep, multitasking, accompaniment, and displaced activities change the interpretation.
Possible effect=
actividad
×design
×persona
×contexto
−the displaced
Observational synthesis
Screens and school performance
In a meta-analysis of cross-sectional studies, total screen time did not show a unique association with performance; television and video games did present small negative associations in some results.
Limit
Open studio ↗Activity, content, and displacement of other practices matter. The cross-sectional design does not establish causality.
Review and meta-analysis
Screen context in early childhood
Recent research distinguishes background exposure, content, coviewing, and caregiver use. Contexts of use are differentially associated with cognitive and psychosocial outcomes.
Limit
Open studio ↗Parental measures and heterogeneity between studies limit simple conclusions by the minute.
Longitudinal cohort
social media and adolescent cognition
In 6,554 adolescents, increasing trajectories of use were associated two years later with somewhat lower scores on reading, vocabulary, memory, and cognitive composite.
Limit
Open studio ↗The differences were small and the adjusted association does not demonstrate that networks are the direct cause.
Observational meta-analysis
Technology and aging
In 57 compatible studies, digital use was associated with lower risk of decline and lower rate of decline in older adults.
Limit
Open studio ↗There may be selection: people with better cognition use more technology. The result does not prove causal protection or any type of use.
What can be said about IQ
Generational changes exist. Attributing them to screens requires something more.
The Flynn effect describes increases in IQ scores throughout much of the 20th century; in some countries and cohorts it has slowed or reversed. The Norwegian study by Bratsberg and Rogeberg supports environmental explanations because the pattern appears within families. “Environmental” includes education, nutrition, health, culture, family structure, migration and many other changes.
That design does not identify the use of screens, the internet or social media as a cause. Proposing them as an explanation is a plausible hypothesis that requires isolating exposure, mechanism, temporality and alternative factors. Furthermore, a change in IQ score does not equate to “the brain becoming less intelligent” in all its capacities.
Bratsberg & Rogeberg, PNAS 2018 ↗Inference traffic light
Yes it allows
Describe an association in a specific population, measurement and period.
Requires caution
Propose displacement, sleep, attention or education as possible mechanisms.
Does not allow alone
Claiming that "screens lower IQ" or that any use causes deterioration.
Socioeconomic development and context
Cognitive autonomy also requires material conditions.
Socioeconomic level is neither an internal characteristic nor a measure of potential. It summarizes correlated exposures and opportunities: income, education, housing, stress, health, nutrition, pollution, language, safety, time, and access to resources. Confusing association with individual essence adds stigma and leads to worse policies.
Resources
Device, connectivity, quality tools, books, space and specialized support.
exposures
Chronic stress, sleep, pollution, nutrition, insecurity and care burden.
Opportunities
Time to practice, human feedback, expectations, networks and cultural experiences.
Use of AI
Ability to choose, verify, protect data and have an alternative.
AI can reduce barriers
Translation, accessibility and on-demand tutoring.
You can offer explanations in another language, convert formats, expand practice and provide support where it is lacking. The benefit is especially valuable if it increases participation and control.
It can also widen gaps
Quality and supervisory capacity are not distributed equally.
Those who have training, time and premium tools can verify and customize better; anyone who relies on free, opaque output may incur more errors and data exposure.
judgement of justice
Do not offer poor automation where others receive rich teaching.
A scalable solution should not replace relationships, education or professional support only for the least resourced groups. Efficiency must be evaluated along with learning, dignity and the ability to complain.
The differences associated with socioeconomic level are probabilistic, modifiable and multi-causal. They do not allow us to infer a person's capacity nor do they legitimize an automatic recommendation based on their origin.
Farah, Neuron 2017 ↗Evidence and readings
A commented library,
not a list of links.
These references support the main concepts of the framework. Each card indicates what it contributes and, when necessary, what it does not allow to conclude. Links open the publication or its stable record in a new tab.
Does the study measure usage, assisted performance, subsequent learning, IQ, health, or self-perception?
Is it an experiment, longitudinal follow-up, cross-sectional comparison, review or survey?
What ages, countries, educational levels, professions and tools does it represent?
Does the result allow us to talk about a cause, association, probable mechanism or just a hypothesis?
Review · mechanisms
Executive Functions
Diamond · Annual Review of Psychology, 2013Framework on inhibition, working memory, flexibility, development and conditions that affect them.
Open post ↗Experiment · learningRetrieval practice produces more learning
Karpicke & Blunt · Science, 2011Distinguishes studying material from recovering and using knowledge later.
Open post ↗Large-scale study · lifespanWhen does cognitive functioning peak?
Hartshorne & Germine · Psychological Science, 2015It shows that different abilities reach maximums at different ages; there is no single cognitive peak.
Open post ↗Review · toolsCognitive Offloading
Risko & Gilbert · Trends in Cognitive Sciences, 2016Describes how actions and tools change internal processing demands.
Open post ↗Psychometric analysisA Psychometric Network Analysis of CHC Measures
McGrew et al. · Journal of Intelligence, 2023Examines relationships between broad aptitudes and the complexity of the hierarchical model.
Open post ↗Critical reviewWhy multiple intelligences theory is a neuromyth
Waterhouse · Frontiers in Psychology, 2023Reviews the lack of evidence for independent intelligences and for prescribing teaching by profile.
Open post ↗Meta-analysis screensScreen Media Use and Academic Performance
Adelantado-Renau et al. · JAMA Pediatrics, 2019Avoid treating all screen time as a homogeneous cognitive exposure.
Open post ↗Cohort · social mediaSocial Media Use Trajectories and Cognitive Performance
Nagata et al. · JAMA, 2025Finds small associations after adjusting for multiple variables; does not completely resolve causality.
Open post ↗Population study · CIFlynn effect and its reversal are environmentally caused
Bratsberg & Rogeberg · PNAS, 2018It supports an environmental origin of generational changes, but does not identify screens as a sufficient cause.
Open post ↗Review · social contextThe Neuroscience of Socioeconomic Status
Farah · Neuron, 2017Analyzes correlates, mechanisms and precautions when interpreting differences associated with socioeconomic level.
Open post ↗Field experiment AIGenerative AI without guardrails can harm learning
Bastani et al. · PNAS, 2025Separates performance during help and subsequent learning without AI; the design of the tutor modifies the effect.
Open post ↗Survey · work with AIThe Impact of Generative AI on Critical Thinking
Lee et al. · CHI, 2025Describes 936 experiences of professionals; it is self-report and does not prove longitudinal loss of capacity.
Open post ↗Metaanalysis · agingTechnology Use and Cognitive Aging
Benge & Scullin · Nature Human Behaviour, 2025The observed association contradicts simple claims of “digital dementia,” without demonstrating causal protection.
Open post ↗!
Current frontier of evidence. Generative AI changes faster than longitudinal research. There are solid studies on learning, cognitive offloading, memory or development, and initial evidence on concrete AI systems. Connecting both levels is reasonable to formulate prudent recommendations, but it does not allow us to claim that we already know the cognitive effects of decades of use.