Skip to content
Spiral Staircase
Buy the book
Menu

Reference topic 44: The Luxury Trap Accelerates

Cognitive offloading: what GPS, search and AI do to memory

EmergingEvidence grade: Emerging. Early, small, animal-only, or actively contested.14 sources5 graded recommendationsTopic hub: Phones and attention

Published Updated

In 2011, Betsy Sparrow, Jenny Liu and Daniel Wegner published a set of experiments in Science about what a person does with a fact when a machine is holding it for them.1

The method was ordinary. Volunteers typed a run of short trivia statements into a computer. Some were told the computer would save what they typed; others, that the file would be erased the moment they finished. Then everyone tried to recall as much as they could.

Mismatch card

The system

Effortful cognition: memory retrieval, navigation, reasoning and the cognitive reserve that a lifetime of such effort appears to build.

What it expects

Daily problems that must be solved with the equipment inside the skull, in a world where nobody else holds the answer.

What it gets

An external store for every fact, a route for every journey and, lately, a competent draft of almost any piece of thinking, available before the effort starts.

What happens in the gap

Unknown at the level of brain health, and measurable already at the level of learning: work that looks better while the person doing it learns less.

The handrail

Tools that withhold the answer (●●○); unaided first passes and occasional self-powered retrieval (●○○)

Strength of evidence

●○○ One strong field trial on learning, plus small observational and preprint studies; no evidence yet on long-term cognitive outcomes, because the exposure is too recent to have any.

What this topic covers

  • ·The Trap Built Out of Good News
  • ·What the Effort Was Buying
  • ·The Classroom That Split in Two
  • ·The Mismatches Now Arriving
  • ·The Timescale Problem

The practical version is in the book: Chapter 12, Attention in the Age of the Phone.

The handrail

Topics 40 to 42 hold the full programs, and Topic 43 makes the case for changing defaults instead of relying on willpower. What follows is specific to handing thinking to machines.

None of this applies to assistive technology. For someone using these tools to read, write, speak or manage a diagnosed condition, the tool is not an elevator past effort but a ramp to the floor everyone else already stands on.

Each recommendation carries the strength of the evidence behind it. How the grades work.

  1. Use AI like a spotter, not an elevator

    ProbableEvidence grade: Probable. Good observational evidence with a plausible mechanism, or small or few trials.

    A gym spotter lets you do the lift and steps in only when you stall. To learn or retain something, choose or set up tools that withhold the finished answer and prompt the next step. In the Turkish trial that version produced the largest practice gains with no exam penalty, so the advice is a differently shaped AI, not less of it.6 If the goal is only to ship the output, none of this applies.

  2. Do your own retrieval sometimes

    EmergingEvidence grade: Emerging. Early, small, animal-only, or actively contested.

    Try the recall before the search. Take the familiar route without the navigation. The reasoning is that retrieval and spatial memory respond to use. Two small studies and a plausible mechanism support it, not evidence that the practice protects against later decline.14

  3. Take the first pass unaided

    EmergingEvidence grade: Emerging. Early, small, animal-only, or actively contested.

    If writing or analysis is a skill you mean to keep, draft badly by yourself first, then use the assistant to improve what exists. This rests on a small preprint and the wider pattern in research on learning, not on a trial.7

  4. Favor whole food over an engineered replacement when the microbiome, your gut's community of microbes, is the point

    ProbableEvidence grade: Probable. Good observational evidence with a plausible mechanism, or small or few trials.

    A shake or bar can be nutritionally complete and still lack the physical structure the gut ecosystem works on.9 This is about kinds of food, not any particular product, and says nothing against medical nutrition products used on clinical advice.

  5. Set the defaults once rather than deciding daily

    ProbableEvidence grade: Probable. Good observational evidence with a plausible mechanism, or small or few trials.

    Evidence grade: Probable. Good observational evidence with a plausible mechanism, or small or few trials. for the underlying habit research· see Chapter 43

    The phone in another room during work, the assistant off for the first draft, one route a week walked without a map. A rule set at a calm moment costs nothing to follow later; a decision made fifty times a day costs something every time.

The skeptic's corner

The case against this topic, from the research itself.

Every information technology has set off this worry, and the first recorded version is the best. In Plato's Phaedrus, Socrates argues that writing will produce forgetfulness, because those who learn it will trust outside marks instead of their own memory.14 He was right about the mechanism and wrong about the verdict: the Sparrow result confirms his objection in a laboratory, about a technology nobody proposes to give up. Slide rules and pocket calculators drew the same complaint from arithmetic teachers, who expected mental arithmetic to decline, and the sky held.

The rate argument is the book's own synthesis of verified facts, not a published finding. A reader may fairly reply that adoption speed and biological harm are different things: a harmless tool adopted fast is still harmless.

The AI evidence is thinner than its coverage suggests. The 666-person survey is a self-reported snapshot, and its published correction replaces a duplicated table without changing the conclusions.5 The MIT study is a preprint with 18 completers.7 The Turkish trial is the only strong field experiment in the set, and its one correction fixes an author's affiliation.6 The GPS work rests on 50 people, 13 of them retested.4 None of it establishes that AI use damages a healthy adult brain, and this chapter does not claim so.

Harari's idea has a demographic hole. The luxury trap runs on populations growing to consume the surplus, and populations no longer reliably do. The global total fertility rate, the average number of children per woman, was 2.25 in 2024, down from 3.31 in 1990 (Topic 32).13 The trap may still close through economic expectation rather than Malthusian arithmetic (population outrunning its food), but that is a different mechanism needing different evidence.

Finally, the case against bypassing the senses draws on predictive processing, the view that the brain builds its picture of the world by predicting its sensory input and learning from the misses. That case is a hypothesis this book finds plausible and cannot support with a verified citation. Treat it as a reason for curiosity, not as a warning.

Terms this topic defines

Related topics

Sources cited above

  1. 1.Sparrow, B., Liu, J., and Wegner, D.M. (2011). "Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips." Science, 333(6043), 776-778. https://doi.org/10.1126/science.1207745 (a series of laboratory experiments with undergraduate participants; the summary here follows the paper's reported findings on recall of saved versus erased information and on recall of folder locations)
  2. 14.Plato. Phaedrus, approximately 274c-275b (Stephanus pagination). No translation is quoted here; the argument is paraphrased.
  3. 5.Gerlich, M. (2025). "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking." Societies, 15(1), 6. https://doi.org/10.3390/soc15010006; a correction was subsequently published as Gerlich, M. (2025). "Correction: Gerlich, M. AI Tools in Society..." Societies, 15(9), 252. https://doi.org/10.3390/soc15090252. The correction replaces a duplicated table (Table 4 had repeated Table 3); the author states that the scientific conclusions are unaffected. (participants spanned a range of ages and education levels; cross-sectional design, with AI use and cognitive offloading self-reported; offloading was the mediating path; younger participants reported the heaviest reliance and scored lowest, and more education was protective)
  4. 7.Kosmyna, N. et al. (2025). "Your Brain on ChatGPT." MIT Media Lab. arXiv preprint 2506.08872. https://arxiv.org/abs/2506.08872 (n=54, of whom 18 completed all sessions; not peer reviewed as of writing)
  5. 6.Bastani, H., Bastani, O., Sungu, A., Ge, H., et al. (2025). "Generative AI without guardrails can harm learning: Evidence from high school mathematics." Proceedings of the National Academy of Sciences, 122(26). https://doi.org/10.1073/pnas.2422633122 (A field experiment with roughly 1,000 Turkish high-school students: unrestricted GPT-4 access improved practice performance 48% but lowered later exam scores 17% versus control, while a guardrailed tutor improved practice performance 127% with exam scores similar to control. A correction was published: Bastani et al. (2025). "Correction for Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics." PNAS, 122(34). https://doi.org/10.1073/pnas.2518204122; it corrects one co-author's affiliation and does not change the results described here.) (the unrestricted arm used GPT-4 directly; the guardrailed arm used a version of the same underlying model that withheld direct answers)
  6. 4.Dahmani, L., and Bohbot, V.D. (2020). "Habitual use of GPS negatively impacts spatial memory during self-guided navigation." Scientific Reports, 10, 6310. https://doi.org/10.1038/s41598-020-62877-0 (n=50 tested in the laboratory; 13 retested roughly three years later)
  7. 13.United Nations Department of Economic and Social Affairs, Population Division (2024). World Population Prospects 2024. (Global total fertility rate 2.25 in 2024, down from 3.31 in 1990.)
  8. 9.Sonnenburg, E.D., and Sonnenburg, J.L. (2019). "The ancestral and industrialized gut microbiota and implications for human health." Nature Reviews Microbiology, 17(6), 383-390. https://doi.org/10.1038/s41579-019-0191-8 (a review; no single numeric estimate of diversity loss is taken from it here)

This topic carries 14 notes in the research manuscript; the 8 above are those behind the material on this page. Full source register.