Generative AI Is Changing the Cognitive Ecology of Education, New Psychology Paper Argues

A calculator changed:
calculation.
Google changed:
information retrieval.
Generative AI may change something deeper.
Where thinking happens.
That is the provocative argument in a new open-access perspective published in Communications Psychology on 14 September 2026.
Researchers Laura Desiree Di Paolo, philosopher and cognitive scientist Andy Clark, and Thomas Wachter argue that generative AI should not simply be understood as:
another educational tool.
Instead, schools should be understood as:
cognitive ecologies—
environments where students, teachers, technologies and social practices jointly shape how knowledge is:
accessed;
produced;
evaluated.
Generative AI alters that ecology because it is unusually:
active;
persistent;
general-purpose. �
Nature
That framing is powerful.
Because the biggest educational question may not be:
Should students use AI?
It may be:
Which parts of thinking should remain cognitively owned by the student?
Humans Have Always Outsourced Cognition
We write things down because:
memory is limited.
We use:
maps.
Calculators.
Calendars.
Books.
Search engines.
None of this destroyed:
human intelligence.
Tools can extend:
cognition.
But Generative AI Is Different
Calculator waits for:
equation.
Search engine gives:
documents.
Generative AI can:
interpret the problem;
suggest the question;
generate the answer;
explain it;
rewrite it;
evaluate it;
continue the conversation.
It participates across:
multiple cognitive stages.
That is new.
Imagine an Essay
Traditional student:
reads question.
Interprets it.
Researches.
Forms argument.
Writes.
Edits.
Now:
student asks AI:
“What does this question mean?”
AI explains.
“Give me arguments.”
AI generates.
“Find weaknesses.”
AI critiques.
“Write structure.”
AI structures.
“Draft.”
AI drafts.
“Improve.”
AI improves.
At the end:
the essay exists.
But we need another question.
Where did the student’s thinking happen?
Output Is Not Learning
This distinction may become the most important educational principle of the AI era.
Education traditionally uses:
output
as evidence of:
cognition.
Essay exists.
Therefore student:
thought.
AI breaks that inference.
Excellent output can now be produced with:
limited underlying understanding.
Schools Need New Evidence of Learning
Perhaps:
oral defence.
Process logs.
Live problem solving.
Reflection.
Application to novel cases.
Not simply:
submit polished document.
AI Can Also Improve Learning Enormously
This is why banning everything is too simple.
Imagine student who does not understand:
statistics.
Teacher has:
30 students.
AI can explain the same concept:
five ways.
No embarrassment.
No impatience.
At:
11 PM.
That is extraordinary educational access.
Personalised Scaffolding Could Be Transformative
Student:
“I don’t understand.”
AI:
Which part?
Student:
“Why correlation isn’t causation.”
AI creates:
example.
Student still confused.
AI tries:
another.
Human tutoring historically provides this.
But human tutoring is:
expensive.
AI could make parts of it:
abundant.
Yet Scaffolding Has a Rule
Eventually:
remove scaffold.
If support remains forever,
learner may never internalise:
the skill.
The AI Tutor Should Sometimes Become Quiet
This is where psychology matters.
Good teaching does not always answer.
Sometimes teacher asks:
“What do you think?”
That creates:
productive struggle.
Productive Struggle Matters
Learning is not maximised by:
making every cognitive task effortless.
Difficulty can force:
retrieval;
reasoning;
integration.
AI systems optimised only for:
helpfulness
may remove useful friction.
Imagine Two AI Tutors
Tutor A:
“Here is the answer.”
Tutor B:
“Before I answer, tell me what you think the first step is.”
B may produce:
slower task completion.
But potentially:
more learning.
Efficiency and Education Can Conflict
Very important.
Business often wants:
fastest correct output.
Education sometimes wants:
slow enough cognition
for learning to occur.
Using the same AI design for both environments would be:
a mistake.
AI Needs Learning Mode
Not just:
answer mode.
A learning-mode system might:
ask questions;
withhold solutions;
provide hints;
test retrieval;
adapt difficulty.
Now AI becomes:
pedagogical infrastructure.
Teachers Become Even More Important
Paradoxically.
If information generation becomes abundant, teachers shift from:
information delivery
toward:
cognitive architecture.
What should students practise?
When should AI assist?
When should AI disappear?
How do we verify:
understanding?
These are sophisticated professional decisions.
Teachers Become Designers of Thinking Environments
That is a much stronger framing than:
“AI will replace teachers.”
A great teacher understands:
learner;
development;
motivation;
classroom relationships;
curriculum.
AI adds:
another cognitive participant.
Teacher orchestrates:
the ecology.
The Extended-Mind Question
Andy Clark is well known for philosophical work around the idea that cognition can extend beyond:
the biological brain
into tools and environments.
Generative AI makes this idea extremely concrete.
If I use AI to remember:
my projects;
organise ideas;
challenge arguments;
retrieve information,
where does:
my cognitive system
end?
Interesting.
But Education Has a Special Requirement
Adults may reasonably outsource:
routine cognition.
Students are still:
building cognition.
That distinction matters enormously.
An Expert Can Outsource Differently From a Beginner
Senior programmer understands:
architecture.
AI writes boilerplate.
Fine.
Beginner programmer lets AI write:
everything.
Problem.
The beginner may never build:
the internal model
needed to recognise when AI is:
wrong.
Expertise Requires Internal Structure
You need enough knowledge to:
evaluate external intelligence.
Otherwise:
confidence becomes borrowed.
This Creates the AI Verification Paradox
People say:
“Students can use AI as long as they check it.”
Check it with:
what knowledge?
🤣
Verification requires:
domain competence.
Therefore foundational knowledge remains essential.
We Cannot Outsource Everything Then Demand Critical Thinking
Critical thinking needs:
material to think with.
A student who knows nothing about:
history
cannot critically evaluate a plausible but false historical answer.
Knowledge and reasoning:
work together.
Memorisation Is Not Dead
Perhaps:
less central.
But not:
irrelevant.
The brain needs:
stored knowledge
to notice:
patterns;
contradictions;
relationships.
You cannot Google every thought while:
having the thought.
AI May Change What We Memorise
Less:
isolated detail.
More:
conceptual frameworks.
But some internal knowledge remains:
cognitive infrastructure.
The Social Ecology Matters Too
Education is not only:
student + information.
It includes:
peers.
Teachers.
Discussion.
Disagreement.
Group work.
Belonging.
AI tutoring cannot simply replace:
the social classroom
without changing:
the psychology of learning.
Students Learn From Other Students’ Confusion
Someone asks:
“Why?”
You had not considered:
why.
Now everyone thinks.
A perfectly personalised AI tutor may remove:
shared intellectual friction.
There Is Value in Learning Publicly
Explaining.
Defending.
Changing your mind.
Hearing:
another perspective.
These are social cognitive skills.
Therefore Personalisation Has Limits
A classroom where every student receives:
perfect individual content
might improve some learning outcomes
while weakening:
shared knowledge;
discussion;
community.
Again:
optimisation has trade-offs.
Assessment Must Change
If AI can write:
excellent essays,
schools should not spend the next decade building:
better AI detectors.
That becomes an arms race.
Instead ask:
What does this assessment actually measure?
Assess Thinking More Directly
Student submits:
essay.
Then:
five-minute defence.
“Why did you choose this evidence?”
“What is the strongest objection?”
“Apply your argument to this new scenario.”
Now:
understanding becomes visible.
AI Can Be Included Transparently
Student records:
where AI helped.
Then reflects:
What did AI get wrong?
What did you change?
Why?
This turns AI use into:
metacognitive learning.
The Goal Is Not AI Avoidance
It is:
cognitive ownership.
Can the student explain:
the final reasoning?
Can they reconstruct:
the argument?
Can they operate:
without the tool when necessary?
Then augmentation may be:
healthy.
Perhaps We Need a New Educational Question
Not:
Did you use AI?
Ask:
What cognitive work did you retain?
Much better.
A Cognitive Work Map
Task:
research essay.
Student owns:
question formation;
source evaluation;
argument;
final judgement.
AI assists:
search;
counterarguments;
grammar.
Now responsibilities are:
visible.
Different Ages Need Different Boundaries
Primary-school child learning:
arithmetic
needs different AI support from:
doctoral researcher analysing literature.
Development matters.
One AI policy for:
everyone
makes little sense.
Younger Learners Need More Cognitive Construction
Adults may use:
GPS.
Child still needs to understand:
space.
Adults use:
calculator.
Child still learns:
number.
The same principle may apply to:
generative reasoning.
We Need Research, Not Panic
Education has survived:
books;
calculators;
Wikipedia;
Google.
But each changed:
learning practice.
Generative AI is unusually broad, so the redesign may be:
larger.
Final Thoughts
The new Communications Psychology perspective gives us a better vocabulary for this moment.
AI is not simply:
a tool entering the classroom.
It is entering:
the cognitive ecology.
That means it changes relationships between:
student;
teacher;
knowledge;
assessment;
effort.
And the most important educational question may become:
Where should cognition live?
Some thinking can safely move into:
tools.
Humans have always done this.
But some thinking needs to remain:
inside the learner
because the purpose of education is not simply to produce:
answers.
It is to produce:
a person capable of understanding why the answer makes sense.
The best educational AI therefore may not be the system that gives students:
the most help.
It may be the system intelligent enough to know:
which help should eventually disappear.
Research note: Di Paolo, Clark and Wachter’s open-access perspective, Educating minds with generative AI, was published in Communications Psychology on 14 September 2026. It argues that generative AI redistributes epistemic labour and restructures the cognitive ecology of education. It is a theoretical Perspective, not an experimental study demonstrating that a specific classroom intervention improves or harms learning. �
Nature
Read the Communications Psychology paper⁠�