Agency, the evidence

Cognitive offloading:
does AI make you dumber?

You tilt your head to read a sideways sign, set a phone reminder, write a shopping list, ask a chatbot. Each is cognitive offloading, and most of it is sensible. The question people now type into search boxes is whether the last one makes them dumber. The studies give a narrower answer than the headlines: they do not show people getting dumber in general, but a ready answer can take the place of the practice that builds a skill.

What cognitive offloading is

The term comes from cognitive psychology. In a 2016 review in Trends in Cognitive Sciences, Evan Risko and Sam Gilbert defined it as using a physical action to change what a task asks of the mind, so that it demands less. Their own examples are plain: tilting your head to see a rotated image, programming a phone to remind you of an appointment. The review asks what makes people offload and what it does to them afterwards, and it treats offloading as an ordinary, everyday behaviour worth studying in its own right.

Often it pays. In three experiments, Benjamin Storm and Sean Stone (2015) found that saving one file before studying a new one improved memory for the new file, but not when people could not trust the save to work. A store you can rely on frees attention for what you still have to learn. The trouble starts when the tool holds the very thing you meant to learn.

AI widens what can be handed over: not a phone number but a first draft, an argument, a calculation, a diagnosis. A newer term marks the step beyond offloading. In a 2026 preprint, Steven Shaw and Gideon Nave call it cognitive surrender: adopting an AI’s output with minimal scrutiny. On our reading, the difference is who does the judging. When you offload, you hand over the work and still check the result. When you surrender, the check goes too.

The claim card: ten studies, dated

News reports on this question keep returning to the same few studies, usually without their size or their status. Here they are in order of publication, with both. Status matters more than the headline: a randomised experiment can show cause within its own setting, a correlational survey cannot, and a preprint has not yet passed peer review.

Checked on 31 October 2026 at the source: the publisher, PubMed or Europe PMC, arXiv, OpenReview, PsyArXiv or the Open Science Framework. Points are percentage points. Statuses change; a preprint may since have passed review.

What the evidence shows, and what it does not

No study in the card shows that AI makes people dumber in general. The two surveys are snapshots: people who use AI tools more rated their own critical thinking lower (Gerlich), and workers more confident in AI reported less critical thinking (Lee). Neither can say which came first, and people who already check less may simply lean on AI more. The brain study is a small preprint, and it recorded activity during one task, not intelligence (more on it below).

What the better evidence shows is narrower. When a tool supplies the answer during the practice that builds a skill, people do worse once the tool is gone. Bastani and colleagues found this in a randomised trial in a school. Liu, Christian and colleagues found it in short online experiments with adults, where those who had had AI help also gave up sooner; that paper was peer-reviewed for the COLM 2026 conference. In the laboratory, Grinschgl and colleagues showed the trade at its simplest: offloading made the task faster and the memory weaker. The Polish colonoscopy study points the same way in a skilled profession, but as an observational before-and-after comparison it cannot prove cause.

How the tool is used changes the result. In the Bastani trial, the tutor that gave hints instead of answers largely avoided the loss on the exam. In Grinschgl’s third experiment, people who had to offload as much as possible but knew a memory test was coming almost made up the cost. In Lee’s survey, workers confident in their own ability reported more critical thinking, not less. And when the AI is wrong, trusting it costs directly: in Shaw and Nave’s first experiment, accuracy fell 15 points when the assistant erred.

Our reading. The honest summary is not that AI rots the brain. It is an old rule of learning: what you do not practise, you do not keep, and a ready answer is the easiest way to skip the practice. The open chapter of Modern Hermet puts it in one line: “Delegating practice has a deferred price.” The evidence is young. Most of the experiments are short, many use students, and the tools have changed since the data were collected. This paragraph is an interpretation of the card, not a further finding.

What the headlines leave out

Part of the Google effect did not replicate. Sparrow, Liu and Wegner reported four experiments in 2011. In the first, people who had just faced hard trivia questions were slower to name the colour of computer words, as if computers had come to mind. The Social Sciences Replication Project could not reproduce it with 234 people (Camerer and colleagues, 2018), and neither could Guido Hesselmann with 117 (2020). The memory experiments, about recalling facts versus where to find them, were not part of either replication, so these two neither confirm nor refute them.

The survey with the striking numbers measured self-ratings. In Gerlich’s published questionnaire, critical thinking is a set of self-ratings on a six-point scale, such as how often you evaluate sources or compare information. The methods section says the items draw on the Halpern Critical Thinking Assessment, but the questionnaire contains no test. The article reports r = -0.68 between AI use and critical thinking in its text and Table 6, and -0.49 for the same pair in its correlation matrix (Table 5), as published on 3 January 2025. The correction of 10 September 2025 replaced a different table, Table 4, which had duplicated Table 3; the author states that the conclusions are unaffected.

The 83% is 15 people. The figure most quoted from the MIT study is 15 of the 18 people in the ChatGPT group who, in the first session, had trouble quoting a sentence from the essay they had written minutes earlier; none of the 18 quoted one correctly. By the third session, 6 of the 18 still could not. The study is a preprint (version 2, 31 December 2025), and a comment posted the same month questions its sample size, its EEG methods and its reporting. The authors, in an FAQ on the project’s website, answer no to the question whether language models are making us dumber, and ask readers not to use words such as brain rot.

How old the worry is: Theuth and Thamus

The fear that a tool for thinking weakens the mind that uses it is at least as old as Plato. In the Phaedrus, written around 370-355 BCE, Socrates tells a story he says he heard. Theuth, Plato’s name for the Egyptian god Thoth, offers King Thamus the art of writing as a remedy for memory and wisdom. Thamus turns the claim around: those who learn letters will stop exercising their memory and will trust marks outside themselves. His verdict: “You have invented an elixir not of memory, but of reminding” (275a, tr. H. N. Fowler). Such pupils, he adds, will seem wise without being wise. Phaedrus at once suggests that Socrates makes up stories of Egypt as he pleases (275b).

Our reading. Thamus’ complaint is about practice, the same thing the modern studies keep returning to, and it was made against a tool that became one of the great aids to thought. Both halves are worth keeping. We know his verdict only because Plato wrote it down. A story in a dialogue is not evidence, and we do not claim it is the first argument of its kind. The full passage, its translations and its later readers are on the myth of Theuth; the popular version, “Socrates said writing destroys memory”, has its own dated card in Ask for the date.

One small move for this week: answer first, then ask

Pick one kind of task you want to stay good at: a type of email, a calculation, a piece of code, a summary of a report. This week, each time it comes up:

1. Answer first. Before you open the model, spend two or three minutes on your own attempt: a draft, a guess, an outline, three sentences. A wrong attempt still counts.
2. Then ask for a critic or a tutor, not the answer. Paste your attempt and ask what is wrong with it, or ask for a hint. Look at a full solution only after you have finished your own.
3. Keep one round without it. Once this week, do the task from start to finish without AI and notice where you get stuck. That is the part you were handing over.

Why it should help. Three findings from the psychology of learning. People remember what they generate better than what they only read, the generation effect: a meta-analysis of 86 studies put the advantage at about 0.4 of a standard deviation (Slamecka and Graf, 1978; Bertsch and colleagues, 2007). Trying to answer before seeing the answer improves learning of the answer, even when the attempt fails (Kornell, Hays and Bjork, 2009; Richland, Kornell and Kao, 2009). And recalling beats rereading after a delay of days, although rereading feels more effective at the time (Roediger and Karpicke, 2006). Reading a model’s answer is closer to rereading than to recalling. The closest evidence with AI itself is the tutor in the Bastani trial, which gave hints instead of answers. The same habit is part of the five questions before you ask AI.

What this does not promise. No study has tested this exact habit with AI; the learning studies used word lists, trivia questions and short texts. It costs minutes, so keep it for skills you want to keep: for work you have no need to learn, offloading is the sensible choice. It does not make the model’s answers right, and it will not catch errors in a field you do not know. It is not a productivity method; if it slows you down, that is what practice costs.

Where to go next

What to hand to AI and what to keep for yourself: Human agency in the AI age. What to do when nobody can say where this is going: acting under uncertainty. The old story in full: the myth of Theuth.

Fair questions

What is cognitive offloading?

Using a tool or a physical action to reduce the mental work a task demands: writing a list, setting a phone reminder, tilting your head to read a rotated image, using a calculator or a chatbot. The definition used here comes from a 2016 review by Evan Risko and Sam Gilbert in Trends in Cognitive Sciences. It is a normal strategy, and often a useful one.

Does AI make you dumber?

The studies so far do not show that it does in general. Surveys find that people who use AI more, or are more confident in it, report less critical thinking, but they cannot show cause. The stronger evidence is narrower: when AI supplies the answers during practice, people do worse once it is taken away, as in a randomised trial with nearly 1,000 high-school students (Bastani and colleagues, 2025). A tutor version that gave hints instead of answers largely avoided that loss.

What is the difference between cognitive offloading and cognitive surrender?

Offloading, in Risko and Gilbert’s sense, is using a tool or an action to lower the mental demand of a task; you can still check the result. Cognitive surrender is a term from a 2026 preprint by Steven Shaw and Gideon Nave for adopting an AI’s output with minimal scrutiny. In their first experiment, accuracy with an AI assistant rose 25 points when it was right and fell 15 points when it was wrong, compared with working without one. On our reading, the difference is who does the judging.

Did the MIT study show that ChatGPT damages the brain?

No. Your Brain on ChatGPT (Kosmyna and colleagues, 2025) is a preprint on 54 young adults from Boston-area universities writing essays, 18 of whom came back for the last session. It used EEG to record brain connectivity during the task, not damage, and the authors’ own FAQ answers no to the question whether language models are making us dumber.

How can I use AI without losing my skills?

Decide which skills you want to keep. For those, make your own attempt first, then ask the AI to criticise it or give a hint rather than the answer, and do one task a week without it. Research on learning supports the mechanism, attempting before seeing the answer, but no study has tested this habit with AI, and for work you do not need to learn, offloading is fine.

Sources & further reading

• Risko, E. F. and Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences 20(9), 676-688. doi:10.1016/j.tics.2016.07.002
• 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. doi:10.1126/science.1207745
• Camerer, C. F. et al. (2018). Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015. Nature Human Behaviour 2, 637-644. doi:10.1038/s41562-018-0399-z. Replication report for Sparrow et al. (experiment 1 only): Open Science Framework.
• Hesselmann, G. (2020). No conclusive evidence that difficult general knowledge questions cause a “Google Stroop effect”. A replication study. PeerJ 8, e10325. doi:10.7717/peerj.10325
• Storm, B. C. and 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:10.1177/0956797614559285
• Grinschgl, S., Papenmeier, F. and Meyerhoff, H. S. (2021). Consequences of cognitive offloading: boosting performance but diminishing memory. Quarterly Journal of Experimental Psychology 74(9), 1477-1496. doi:10.1177/17470218211008060
• Gerlich, M. (2025). AI tools in society: impacts on cognitive offloading and the future of critical thinking. Societies 15(1), 6, published 3 January 2025. doi:10.3390/soc15010006. Correction: Societies 15(9), 252, 10 September 2025, doi:10.3390/soc15090252
• Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. and Wilson, N. (2025). The impact of generative AI on critical thinking: self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. CHI ’25, article 1121. doi:10.1145/3706598.3713778
• Kosmyna, N. et al. (2025). Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task. Preprint, not peer-reviewed: arXiv:2506.08872 (version 1, 10 June 2025; version 2, 31 December 2025). The authors’ FAQ: brainonllm.com. Comment by Stankovic, M., Hirche, E., Kollatzsch, S. and Doetsch, J. N., preprint: arXiv:2601.00856.
• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning: evidence from high school mathematics. PNAS 122(26), e2422633122. doi:10.1073/pnas.2422633122. Correction: PNAS 122(34), e2518204122, 26 August 2025, doi:10.1073/pnas.2518204122 (one author’s affiliation only).
• Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterology & Hepatology 10(10), 896-903. doi:10.1016/S2468-1253(25)00133-5; abstract on PubMed. Correction: Lancet Gastroenterology & Hepatology 10(11), e12, November 2025, doi:10.1016/S2468-1253(25)00294-8 (a covariate had been left out of the multivariable analysis in the appendix; the corrected estimates are those for AI exposure and patient age).
• Shaw, S. D. and Nave, G. (2026). Thinking, fast, slow, and artificial: how AI is reshaping human reasoning and the rise of cognitive surrender. Preprint, not peer-reviewed, posted January 2026. doi:10.31234/osf.io/yk25n_v1
• Liu, G., Christian, B., Dumbalska, T., Bakker, M. A. and Dubey, R. (2026). AI assistance reduces persistence and hurts independent performance. Conference on Language Modeling (COLM 2026), published on OpenReview 8 July 2026; text also on arXiv:2604.04721 (version 1, 6 April 2026, a preprint; version 6, 13 October 2026).
• Slamecka, N. J. and Graf, P. (1978). The generation effect: delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory 4(6), 592-604. doi:10.1037/0278-7393.4.6.592
• Bertsch, S., Pesta, B. J., Wiscott, R. and McDaniel, M. A. (2007). The generation effect: a meta-analytic review. Memory & Cognition 35(2), 201-210. doi:10.3758/BF03193441
• Kornell, N., Hays, M. J. and Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition 35(4), 989-998. doi:10.1037/a0015729
• Richland, L. E., Kornell, N. and Kao, L. S. (2009). The pretesting effect: do unsuccessful retrieval attempts enhance learning? Journal of Experimental Psychology: Applied 15(3), 243-257. doi:10.1037/a0016496
• Roediger, H. L. and Karpicke, J. D. (2006). Test-enhanced learning: taking memory tests improves long-term retention. Psychological Science 17(3), 249-255. doi:10.1111/j.1467-9280.2006.01693.x
• Plato, Phaedrus 274c-275b, translated by Harold N. Fowler, text on the Perseus Digital Library. Date of composition, 370-355 BCE: Harvey Yunis (Cambridge, 2011), as reported by Christopher Moore, Bryn Mawr Classical Review 2012.07.55.
• Modern Hermet (The Hermeticum, 2026), chapter 8, Technology (open in full): the line on delegating practice.
• The weekly move is our own arrangement. It is not a treatment and makes no therapeutic or productivity claim.