The Thinking We Hand Off: What AI Takes, What It Frees, and Where Both Should Go

A borrowed word

In Spanish I have been calling this externalidad cognitiva, a cognitive externality, and I trained as an economist, so I know I am bending the word. In economics an externality is an effect on third parties: the factory’s smoke that the neighbours breathe. Here I mean something closer to home: the thinking a learner hands to the machine. I will come back to the economist’s meaning in the last piece of this series, because it asks who pays. This one is about what is handed over and what is freed.

The research has a plainer name for the behaviour: cognitive offloading, which the psychologists Risko and Gilbert define as “the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand.” Their examples are deliberately ordinary: tilting your head to read a rotated image, programming a phone to remind you of an appointment.¹ We have always offloaded. What is new is how much of the thinking can now leave, and how little of it we notice leaving.

So here is the list of what we hand off, as I see it. The items are not equal, and the verdict at the end ranks them.

1. Remembering

The oldest handoff, and the one that worries me least. In 1998 the philosophers Andy Clark and David Chalmers built a famous argument around it. Otto has Alzheimer’s disease and carries a notebook everywhere: “When he learns new information, he writes it down. When he needs some old information, he looks it up.” For Otto, they argue, the notebook plays the role memory plays for everyone else, and in cases like his the mind extends into the world.² That is the strongest version of the enthusiast’s case, and it deserves to be taken seriously.

There is evidence that this handoff frees something real. In three experiments, the psychologists Storm and Stone found that saving one file before studying a new one “significantly improved memory for the contents of the new file.” Saving the first file got it out of the way of the second.³

But read both sources to the end, because each carries a condition. Storm and Stone’s effect “was not observed when the saving process was deemed unreliable.”³ And among the features that make Otto’s notebook part of his mind, Clark and Chalmers list a fourth, which they themselves call arguable: the information “has been consciously endorsed at some point in the past, and indeed is there as a consequence of this endorsement.”² Otto wrote his notebook. When the AI writes the notebook for you, it holds things you never knew, and its reliability is precisely what nobody can promise. The best case for offloading carries the two conditions that AI output, by default, does not meet.

Hand off remembering, then, when what is stored is something you put there, or something you can check.

2. The mechanical scaffolding

I taught myself to code at nine on a Texas Instruments TI-99/4A, and at fourteen I took a couple of courses at IBM, COBOL and databases on a System/34. They were the only programming courses I ever took. On that base of variables, routines and structure I can still read languages I never studied. The syntax was only ever the vehicle. That is lived experience more than settled science, and I hold it with that honesty, but it tells me something about what can be handed off: the syntax costs little to give away once you hold the structure underneath it.

That is how we wrote it into the program we are running with young people this year, «Crecer con IA», Grow with AI. When participants build with AI, the technical scaffolding that is not the subject of the Learning Unit is handed off on purpose: creating the window, a file format, starting a page. The specification, the decision about what to change, the explanation of the error and the proof that it works are not handed off. And the autocomplete stays off in the first coding Units until the participant can write the structure alone, because for someone who does not hold the base yet, the syntax is part of how the base gets built. What this frees is the thing the year is for: more rounds of testing with real people.

3. The first draft

Writing is how many of us find out what we think. Hand off the draft and you hand off that moment. One small study points the same way: adults who wrote essays with ChatGPT “fell behind in their ability to quote from the essays they wrote just minutes prior.” It is a preprint with a small sample, still under review, and I read it as a signal, not a verdict.⁴

What a handed-off draft frees is real: time to revise, to compare, to argue with the result. The trouble is that the freed time is spent on a text the writer did not think through, which is a strange thing to revise. The trade I would make is the reverse one: write the draft yourself and hand off the critique of it. Ask the machine to mark what is weak, then fix it with your own words.

4. The practice

This is the handoff that frees nothing, though it looks as if it does. The clearest evidence I know comes from a field experiment with nearly a thousand high-school mathematics students. With GPT-4 available during practice, their grades on the practice problems rose 48% with a standard chat interface and 127% with a tutor designed to safeguard learning. When access was taken away, the students who had used the standard interface performed 17% worse than students who never had it. The authors’ diagnosis is blunt: without guardrails, students “attempt to use GPT-4 as a ‘crutch’ during practice problem sessions.” And the negative effects “are largely mitigated” by the safeguards in the tutor version.⁵

Two lessons sit in that paper. The gains while the tool was there were performance, not learning. And the same model, set up to give hints instead of answers, largely avoided the harm. A second experiment, with 1,200 adults online, points the same way: AI-written explanations helped compared with seeing only the answers, and “the benefits were largest for those who attempted problems on their own first.”⁶ Attempt first; then the machine. That is the order that keeps the practice yours. I made the case for why the quantity of practice matters in No Shortcuts to Mastery; handing it off is the one shortcut that leaves nothing behind.

5. The judgment

The last item is the one I would never hand off while it is what the learner is there to build: the question, the judgment of what is true, the voice of one’s own text. A survey of 666 adults found a significant negative correlation between frequent use of AI tools and critical thinking, mediated by increased cognitive offloading.⁷ It is a correlation, not a cause. It may be that AI wears judgment down, or that people less inclined to judge lean on the tools more. Either reading lands in the same place for anyone designing learning.

And here a handoff can be turned around. The study material we give participants asks them not to request explanations but to tell the AI: don’t explain it to me; ask me questions until I get to the why myself. The same exercise works without any AI, with cards and a friend, and the page says what changes: only who asks the questions. Hand off the asking; keep the answering. And distrust the machine’s verdict on you: it tends to please, and if it says you are ready, the real test is tomorrow, without it.

The verdict

Ranked from safest to costliest: remembering, when you wrote what is stored or can check it; the mechanical scaffolding, when you already hold the structure; the critique of your draft, but not the draft; and never the practice or the judgment while they are what the task is for.

That last phrase is the whole line. Every task carries some things and is for others. Handing off what a task carries frees something real. Handing off what it is for frees nothing; it only buys a better performance while the tool is in the room. And the line moves with the learner: syntax is something a task carries for me and something it is for in a fourteen-year-old’s first coding Unit.

What is freed goes nowhere on its own

Now the other half of the question, the half the warnings skip. Even when the handoff is right, the time and attention it frees have no direction of their own. Notice that in Storm and Stone’s experiments the freed memory had somewhere to go: there was a second file waiting to be studied.³ Learners do not always have the second file.

In «Crecer con IA» we draw a board with two columns: MORE QUICKLY / MORE EASILY on one side and MORE, BETTER, FURTHER on the other, with one question underneath: did this improve the instruction, or did it improve the question? Every minute AI saves goes in one of the two columns, and a minute that lands in “faster” and stays there gets discussed. Each participant keeps a three-line record for every Unit:

I kept ___ · I handed off ___ · I used what was freed for ___

And the design rule behind it, which is a design position of the program and mine, not a research finding: in every Learning Unit that uses AI we declare what effort is kept on purpose, what is handed off on purpose, and where what is freed should go. Freed time with no declared destination is taken by the “faster” column. If your institution wants to write those declarations into its own learning experiences, that is work we do with schools and universities.

The objection I take most seriously

It goes like this. Deciding what a task is “for” is exactly the judgment a novice cannot make; that is what makes them a novice. Asking a fifteen-year-old to choose what to hand off is asking the learner to make the expert’s call, and leaving learners to decide alone is the arrangement that serves disadvantaged learners worst — I quoted the OECD saying as much in my essay on PISA. Your three lines are a self-report that a teenager can fill in with whatever sounds right.

I concede the first part entirely. The learner does not draw the line alone, at least not at the start. The declaration is made first by the adults who design the Learning Unit, before anyone sits down, and the Learning Facilitator holds it in the room. The three lines do not hand the decision to the learner; they make the handoff visible so it can be discussed and corrected, and the test the next day, without AI, checks it against something nobody can talk around. Over the year, the learner takes more of the declaration over. That is the point: building the judgment rather than replacing it. The second part of the objection I cannot answer yet. The three lines have not been evaluated. They are a design decision, and we will know what they do when we have watched them in the room.

Who gets the freed part

One more study, because it opens the question this series ends on. Repeating the saving experiments, Buchli and Storm found that people with a higher working-memory capacity got a larger benefit from saving: a significant positive correlation, across three experiments.⁸ It is a laboratory task with word lists, not a classroom. But if it holds anywhere near learning, then what offloading frees is not distributed evenly: it goes furthest for those who already have the most room to use it.

Which is where the economist’s meaning of the word comes back. This is the middle piece of three: the effort that teaches, the thinking we hand off, and who pays for it. The first, Raise the Friction, is about the friction we keep on purpose. The third will ask who pays for both, and I do not know the answer yet.

Every study cited here was checked against its primary text or the authors’ own abstract, and each is quoted at its own strength: preprints, correlations and laboratory tasks are named as such. The three-line record, the two-column board and the design rule are positions of the program and the author, not findings. All sources are listed below.

Four perspectives

References

Every source below was checked against its primary text or the authors’ own abstract. Quotations stay in English, the language they were verified in. Preprints, correlational studies and laboratory tasks are marked as such.

1. Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. doi.org/10.1016/j.tics.2016.07.002 The definition and the two examples are from the authors’ abstract.
2. Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. doi.org/10.1093/analys/58.1.7. Otto’s notebook: §4, “From Cognition to Mind”; the four features, the fourth called “arguable” by the authors: §5, “Beyond the Outer Limits”. Quoted from the authors’ copy (consc.net/papers/extended.html), which has no page numbers and gives the pages as 10–23; the publisher’s record gives 7–19.
3. Storm, B. C., & Stone, S. M. (2015). Saving-enhanced memory: The benefits of saving on the learning and remembering of new information. Psychological Science, 26(2), 182–188. doi.org/10.1177/0956797614559285 Three laboratory experiments. Both the effect and its reliability condition are in the authors’ abstract.
4. Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. arxiv.org/abs/2506.08872 Preprint, under review, not peer-reviewed. Adults aged 18 to 39; a small sample. Suggestive, not conclusive.
5. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi.org/10.1073/pnas.2422633122 A field experiment in one school. The 48%, 127% and 17% figures and the quotations are from the authors’ abstract; the gains are performance while the tool was available.
6. Kumar, H., Rothschild, D. M., Goldstein, D. G., & Hofman, J. M. (2025). Math education with large language models: Peril or promise? In Artificial Intelligence in Education (AIED 2025), Lecture Notes in Computer Science, pp. 60–75. doi.org/10.1007/978-3-031-98459-4_5 A pre-registered online experiment with adults, not school learners. The explanations also helped, less, when they came before the attempt.
7. Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. doi.org/10.3390/soc15010006 Surveys and interviews; a correlation, not a cause.
8. Buchli, D. R., & Storm, B. C. (2025). Working memory capacity and the saving-enhanced memory effect. Memory, 33(8), 1015–1030. doi.org/10.1080/09658211.2025.2551230 Three laboratory experiments with word lists; a correlation.