Workbench / 08 Critique / The offloading audit
Am I still doing the thinking?
Twelve questions about four habits, about ten minutes, nothing stored. At the end you get a reading for each habit, the studies behind it, and a plan short enough to keep. This is not a test of intelligence. It is a check on where the thinking is happening: in you, or in the window.
Offloading is not the enemy. Unnoticed offloading is.
What the word means
Cognitive offloading is “the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand.” That is the 2016 definition from Evan Risko and Sam Gilbert, and their examples are homely: tilting your head to read a rotated page, setting a reminder on your phone. Writing itself is offloading. So is a library. Humans have always put memory and calculation into the world so that our heads could do something else.
The trouble starts when the offloading happens without our noticing, and when the thing we hand over is the thing we were supposed to be building. In 2011 a study in Science found that people remembered trivia less well when they believed a computer would save it, and remembered where it was saved better than what it said. In 2020 a study of drivers found that a lifetime of GPS use predicted worse spatial memory, and that heavier users declined faster over three years. Nobody decided to stop being able to find their way home. It just happened, one turn at a time.
Generative AI is offloading at a different scale, because the thing it takes is not a fact or a route but the act of composing, reasoning and judging. Whether that costs you anything depends on how you set it up. Which is what the audit is for.
Sources: Risko & Gilbert, TICS 2016 · Sparrow, Liu & Wegner, Science 2011 (one experiment did not replicate in 2018; the direction held) · Dahmani & Bohbot, Scientific Reports 2020 · checked Sep 2026
The audit
Four habits, three questions each
Answer for how you have actually worked in the last month, not how you would like to. Nobody sees this but you.
Habit 1 · Verify
When an AI gives me a fact, a number or a citation, I open the source before I use it.
I have caught an AI being wrong about something in my own field in the last month.
When the answer sounds confident and well-organized, I take that as a sign it is right.
Habit 2 · Retain
An hour after finishing a piece of work with AI help, I could explain its argument without looking.
I take notes by hand, or rewrite the model’s output in my own words, rather than pasting.
I could not reproduce, on an exam with no tools, work I got a good grade for with them.
Habit 3 · Attempt first
I try the problem myself before the first prompt.
When I use AI for coursework, I ask it to explain or quiz me rather than to produce the answer.
I reach for the model the moment a task feels hard.
Habit 4 · Own the judgement
In the last week I have disagreed with an AI answer and been right.
I could say, for a piece of AI-assisted work, which decisions were mine.
There is a skill I used to do by hand that I would now struggle to do without the tool.
Your reading
A plan you can keep
The evidenceWhat the studies actually measured
10 studies · checked Sep 2026| Study | Who and how | Finding | Handle with care |
|---|---|---|---|
| “Your Brain on ChatGPT” Kosmyna et al., MIT Media Lab, June 2025 | 54 people wrote essays with an LLM, a search engine, or nothing, wearing EEG caps. | The LLM group showed the weakest brain connectivity; in session one, 15 of 18 LLM users could not correctly quote from the essay they had just written. Self-reported ownership was lowest in that group. | Preprint, not peer-reviewed; only 18 finished the last session. The “55% reduced connectivity” number circulating online is not in the paper. |
| “Generative AI Can Harm Learning” Bastani et al., PNAS, June 2025 | Randomized trial, nearly 1,000 high-school maths students in Turkey, 2023–24. | Unrestricted ChatGPT raised practice performance 48% and lowered the no-AI exam score 17%. A “GPT Tutor” built to give hints rather than answers raised practice 127% and left exam scores essentially unchanged. | One school; the strongest evidence here that design, not the tool, decides. |
| AI tools and critical thinking Gerlich, Societies, Jan 2025 | Survey of 666 people in the UK. | Frequent AI use correlated with lower critical-thinking scores, mediated by cognitive offloading; younger participants showed higher dependence. | Correlational, not causal; a correction to one table was published in September 2025. |
| Knowledge workers and critical thinking Lee et al., CHI 2025 (Microsoft / CMU) | 319 workers, 936 real examples of AI use. | Higher confidence in the AI went with less critical-thinking effort; higher confidence in oneself went with more. | Self-reported effort, not measured performance. |
| AI assistance and coding skill Anthropic, Jan 2026 | Randomized, 52 engineers learning a new Python library. | AI-assisted group scored 50% on a comprehension quiz against 67% for hand-coders, and finished only about two minutes faster. | Small sample, mostly junior engineers, comprehension measured immediately. Published by the vendor, against its own product. |
| Experienced developers with AI METR, July 2025 | Randomized, 16 experienced open-source developers, 246 real tasks. | Developers took 19% longer with AI tools and believed afterwards that they had been 20% faster. | Small, expert, specific tooling; the perception gap is the finding. |
| Deskilling in colonoscopy Budzyń et al., Lancet Gastro. & Hep., Aug 2025 | 19 experienced endoscopists, 1,443 non-AI colonoscopies before and after routine AI use. | Detection rate in unassisted procedures fell from 28.4% to 22.4% after AI exposure. | Observational; authors say other factors may have contributed. |
| Handwriting vs typing Van der Weel & Van der Meer, Frontiers in Psychology, Jan 2024 | 36 students, 256-electrode EEG. | Handwriting produced far more elaborate connectivity in regions tied to memory and learning than typing did. | Measures brain connectivity, not grades or retention. |
| How students use Claude Anthropic Education Report, Apr 2025 | 574,740 academic conversations analyzed by the company. | About 47% were “direct” requests for an answer or output with minimal engagement; the model was mostly doing higher-order work (creating, analyzing). | Company analysis of its own product, not a survey of students. |
| Digital Education Outlook 2026 OECD, Jan 2026 | International policy review. | Completing tasks with generative AI “does not necessarily translate into learning”; offloading cognitive work to chatbots risks “metacognitive laziness.” | Quotations via the European Commission’s summary of the report; check the wording against the PDF before citing it in an essay. |
Design beats willpower
The most useful thing in the table is the tutor condition in the Turkish study. Same students, same model, same month. The version that answered questions made them worse on the exam; the version that gave hints and made them try did not. Nobody’s character changed between the two rooms. The setup did.
That is also the 2010 finding about automation bias, which predates chatbots by a generation: the habit of following a decision aid even when it is wrong “cannot be prevented by training or instructions.” You will not out-discipline the tool. You can arrange the work so that the thinking has to pass through you: try first, write your own version, verify one thing, say which decisions were yours. That is what a lab notebook is for, and why the Institute’s TRAIL notebook asks what the AI contributed, what you decided, and what still needs checking.
Parasuraman & Manzey, Human Factors 2010 · checked Sep 2026
How many of us
- 95% of UK undergraduates use generative AI and 94% use it for assessed work; 12% paste AI text straight in, up from 8% a year earlier. HEPI 2026
- 85% of U.S. college students use it; 55% say the effect on their learning and critical thinking is “mixed.” Inside Higher Ed, 2025
- 54% of U.S. teens have used a chatbot for schoolwork, up from 26% using ChatGPT in 2024 and 13% in 2023. Pew, Feb 2026
- 86% of 3,800 students in 16 countries used AI in their studies in 2024; 58% felt they lacked AI skills. Digital Education Council
Different populations and years; not one trend line. Checked Sep 2026.