You ask a model to check a clause in a contract. It answers in two tidy paragraphs, with total confidence. Do you read the clause again yourself?
Most people do not, and the research says that is the problem. Hand a model the sums, the search and the routine paperwork. Keep the checking, the weighing of risk and the choice. Those are the skills that weaken fastest when a model does them, and they are the same skills you need to catch the model when it is wrong.
We are The Long Game Project. We design wargames and exercises where teams practise hard decisions, and we built Shortlist, a tool that matches board games to the thinking skills they exercise. For Shortlist we scored 75 thinking skills against the research on what happens when people hand thinking to a tool. Twenty-seven of them matter most and wear down fastest.
Every tool that thinks for you charges a small fee
Calculators and satellite navigation each took a small fee from the skill they replaced.
In 2019, researchers gave undergraduates an on-screen calculator rigged to show wrong answers. Very few showed any suspicion, even when the answers were absurd.1 People who used GPS heavily over their lives had worse spatial memory. Only 13 of the original 50 were tested again three years later, but among them, heavier GPS use went with a steeper decline.2
The fee was worth paying. The calculator was nearly always right, and the suggested route nearly always got you home. A tool that is nearly always right makes people more capable, even while one skill fades behind it.
A model changes that deal, because it is not nearly always right.
A model makes you better when it is right and worse when it is wrong
At Wharton, Steven Shaw and Gideon Nave gave 1,372 volunteers reasoning puzzles and an AI assistant whose answers they had quietly seeded with errors.3 When the assistant was right, people who used it got about 25 percentage points more puzzles right than an unaided group. When it was wrong, they got about 15 points fewer right.
The authors' reading is that people stopped checking. They call it cognitive surrender.
The help does not train the skill either. In a trial with 480 consultants at Boston Consulting Group, people used ChatGPT on data-science work outside their expertise and came close to expert results.4 Tested later without the tool, they did no better than colleagues who had never used it. The authors compare the model to an exoskeleton. It carried the load and built no muscle.
With a model in the loop, your result depends on the model being right, and you stop noticing how much.
Four ways a skill wears down
The research on cognitive offloading, which means using a tool or the world to do part of a thinking job, is older than the chatbot.5 It points to four separate ways a skill fades.
Recall fades when you stop recalling. Memory gets stronger each time you pull something out of it, not each time you store it. GPS showed this for routes. An EEG study of essay writing now suggests it for prose: people who wrote with ChatGPT struggled to quote accurately from essays they had just written.6 That study is a preprint, not yet peer reviewed.
Confident output switches off scrutiny. With an automated aid in the room, people make two kinds of error: they follow bad advice, and they miss problems the aid did not flag.7 8 Training helps with the first, but less than you would hope. In one experiment with 197 people, 80.7% said they had noticed the AI making mistakes, and many still relied on its advice.9 Very little fixes the second.
The routine work was the training. In 1983 Lisanne Bainbridge named the irony of automation: take the routine work off the operator, and their skill fades just before a rare failure needs more of it.10 11 Medicine now has the number. Four Polish endoscopy centres brought in AI help to spot adenomas, the growths that can turn into bowel cancer. Afterwards, in the procedures doctors did without the AI, the share that found at least one adenoma fell from 28.4% to 22.4%.12 It is an observational study, so it shows the drop, not a proven cause.
Trust takes the place of thinking. Among 319 knowledge workers, more trust in a model went with less critical thinking.13 Their work did shift toward checking the model's output, which is the right move, as long as the checker still has the skill. A survey of 666 people found the same link.14 Both are surveys of what people report, so treat them as a warning sign, not a measurement.
Twenty-seven skills sit where the cost is highest
Shortlist sorts its 75 thinking skills into five domains of 15: cognitive (reasoning and analysis), social (working with people), executive (managing your own time, attention and effort, in the psychologist's sense of the word), operational (running a process) and creative (making and testing ideas). We scored each skill on two counts, from 1 to 5:
- Importance: how strongly the research links the skill to good decisions, lasting learning, leadership and steady performance.15 16 17 18
- Atrophy risk: how likely the skill is to weaken when people hand it to a model, weighted by how good the evidence is.
These are our scores, built from the papers in this piece. They are a judgement, not a measurement, and two careful people would move some of them by a point. Crossing the two at 4 gives four bands. A skill that scores 4 or more on both is one to keep doing yourself.
27 of 75 thinking skills matter most and wear down fastest when a model does them.
Foundational
40High importance · lower atrophy risk
Defend: keep practising
27High importance · high atrophy risk
Background
7Lower importance · lower atrophy risk
Watch: let it help, then check
1Lower importance · high atrophy risk
Source: LGP scores for the 75 skills in Shortlist's taxonomy, May 2026 · scored from the cited papers · importance and atrophy risk scored 1 to 5 · 4 or more counts as high
Cognitive and creative skills make up 20 of the 27 in the Defend band. Social skills have just one.
Creative skills score high because coming up with ideas and reframing a problem are now the first things many people ask a model to do. Social skills score low because, so far, few people hand a negotiation or a hard conversation to a chatbot. If they start, the scores move. The one social skill among the 27 is presenting, because slides are the part people already generate.
Five skills score the maximum on both counts. These are the ones to protect first.
| Skill | What it does | Try this |
|---|---|---|
| Critical evaluation | Tests a confident answer against the evidence | Before you accept an AI answer, write one reason it could be wrong |
| Risk assessment | Weighs uncertainty against how often things usually happen | Write your own range first, then ask the model for its range and make it argue the difference |
| Decision making | Chooses between options when the values conflict | Let the model rank the options. You rank the values |
| Focus | Keeps attention on a hard problem | One hour, one tab, one task, before you open the model |
| Reframing | Sees the problem as a different problem | Write two other ways to state the problem before you prompt |
Focus rests on the thinnest evidence of the five: the EEG preprint above and Nicholas Carr's argument in The Shallows.19 It is on the list on our reading: a model is always one tab away, and every hard minute is a chance to hand the problem over.
All 75 scores are in Figure 2, with what each skill does and how strong the evidence is.
Sort by atrophy to see what to keep. Sort by offload to see what to hand over.
| What it does | ||||
|---|---|---|---|---|
| Conceptual ReframingCreative · strong evidence | 5 | 5 | -2 | Seeing the problem as a different problem |
| Critical EvaluationCognitive · strong evidence | 5 | 5 | -2 | Challenging confident outputs against evidence |
| Decision MakingCognitive · strong evidence | 5 | 5 | -2 | Choosing under uncertainty across conflicting values |
| Focus MaintenanceExecutive · strong evidence | 5 | 5 | -2 | Sustaining attention through difficulty |
| Risk AssessmentCognitive · strong evidence | 5 | 5 | -2 | Calibrating subjective uncertainty against base rates |
| Memory DevelopmentCognitive · strong evidence | 5 | 4 | -2 | Effortful retrieval that builds durable memory |
| Spatial ReasoningOperational · strong evidence | 5 | 4 | -1 | Building and using internal maps |
| Numerical ProcessingOperational · strong evidence | 5 | 3 | +1 | Estimating and sanity-checking numeric outputs |
| Creative Problem SolvingCreative · strong evidence | 4 | 5 | -1 | Generating non-obvious approaches |
| Design ThinkingCreative · moderate evidence | 4 | 5 | -1 | Holding user, problem, and solution at once |
| Flexible ThinkingCreative · strong evidence | 4 | 5 | -2 | Switching frames as evidence shifts |
| Hypothesis TestingCreative · moderate evidence | 4 | 5 | -1 | Designing a test that could change your mind |
Source: LGP scores for the 75 skills in Shortlist's taxonomy, May 2026 · scored from the cited papers · evidence: strong means a trial or a long study, weak means an argument from older tools
You can hand over the work only while you can still check it
The third score, offload, asks a different question: how well does a model do this job today? Split the 75 skills by it and you get three lanes.
46 of 75 skills are ones to keep doing yourself. Social skills lead.
keep ithand over, then checkhand over
Source: LGP scores for the 75 skills in Shortlist's taxonomy, May 2026 · scored from the cited papers · keep: offload score -1 or -2 · check: 0 · hand over: +1 or +2
Social skills top the keep lane even though they barely wear down. A model cannot do them, so nobody hands them over. The middle lane is the risky one: let the model do the routine part, then check the result. That lane assumes the person checking can still tell a plausible answer from a correct one.
That checking is a skill, and it runs on the expertise that AI-assisted work stops building. Nolan Lovett, who studies professional learning, calls this the validation tether.20 The endoscopists did not forget how to look at a screen. They lost some of the unaided judgement their oversight depended on.12
Offloading is fine when the tool is reliable. A saved file frees your memory for the next thing you learn, but only if the file is really there when you go back.21 A model in 2026 is reliable at formatting a citation and unreliable at judging a risk. Treat the middle lane as rented. You keep it only while the skills that do the checking stay in practice.
Offloading has paid off before, for well-defined jobs
This is the strongest case against everything above. Socrates complained that writing would ruin memory. People said the same about calculators and the web. Each time, people ended up more capable, not less, and the old skill that faded was rarely missed. Offloading is often the rational choice: it frees attention for harder work.5
For well-defined jobs, that argument wins. Hand over the unit conversions and the meeting transcripts without guilt.
The difference now is that the model also does the checking-level work, and its errors arrive fluent and well formatted. (The calculator never wrote you a persuasive paragraph about why its wrong answer was right.)
The evidence is also young, and that cuts both ways. Several of the studies above are working papers, some are small, and the surveys measure what people say about themselves. We expect some of the scores to change as better studies land.
Choosing is the part a model cannot do for you
The skills a model replaces well have something in common. They are well-defined operations on well-defined inputs: arithmetic, translation, search, boilerplate. The skills it cannot replace, even when its output looks like a replacement, need a choice under uncertainty, between options you cannot undo, when the values conflict. That is a fair definition of strategy.17
A century ago Frank Knight split risk, where you can measure the odds, from uncertainty, where you cannot even list the outcomes.22 Statistical tools work on the first. Daniel Kahneman and Gary Klein later showed that expert intuition is only reliable where the world is regular and the feedback is fast.15 Most strategic choices, such as which market to enter, who to hire or when to give ground in a negotiation, happen where the world is irregular and the feedback is slow.
In that territory nobody guesses well, you or the model. What is left is the choice itself: which risk you will carry and which value wins. A model can lay out the options. Someone still has to own the choice.
The Oxford philosopher Carissa Véliz argues that a prediction about a person can make itself true: refuse someone credit because a model predicts they will struggle, and they struggle.23 A self-fulfilling prediction never shows up as an error. Whether to make the prediction at all is a question the model cannot answer for you.
You can offload the sums, the routine drafting, the search and the first cut of the analysis. You cannot offload the choice, because the values are the question, and the model takes the question as given.
Three things to try this week
Ration the help. In a 12-week chess study, students who got only the system's own scheduled AI hints improved about twice as much as students who also had a button to ask for more.24 The students with the button knew they were leaning on it too much and leaned anyway. Decide the rule in advance. The model can draft the status update. It does not draft the first version of a plan, a decision paper or an argument, because writing that draft is how you find out what you think.
Try first, then ask. Spacing practice out, recalling before you re-read and practising in varied conditions all make learning stick.16 A model removes all three by default. Put them back. Try the recall before you ask. Write the rough draft cold before the model edits it.
Write one reason it could be wrong. Surrender starts when you stop testing the premise. Before you accept an answer, write one reason it might be wrong, then go and look for it. Treat a fluent answer as a reason to check.25 Then practise acting on it. Noticing a mistake and refusing the advice are two separate skills.9
The skills worth keeping need practice, and the choosing needs it most. That is hard, because real strategic choices come a few times a year and the feedback arrives years later. A wargame or a tabletop exercise shortens that loop. A team makes the call under pressure, sees what it cost within the hour and talks it through, with no model at the table.
Practise the choosing where the feedback is fast.
Shortlist has the full scoring table, all 75 skills with their scores and sources, and a short diagnostic that shows where to start. If your team needs to practise the choosing itself, under pressure, talk to us. Think a score is wrong? Tell us, with the source.
Footnotes
-
LaCour, M., Cantú, N. G. & Davis, T. (2019). When calculators lie: A demonstration of uncritical calculator usage among college students and factors that improve performance. PLOS One 14(10), e0223736. Link ↩
-
Dahmani, L. & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports 10, 6310. Link ↩
-
Shaw, S. D. & Nave, G. (2026). Thinking - fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School. SSRN. Link ↩
-
Wiles, E., Krayer, L., Abbadi, M. et al. (2024). GenAI as an exoskeleton: Experimental evidence on knowledge workers using GenAI on new skills. Working paper, SSRN. Link ↩
-
Risko, E. F. & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences 20(9), 676-688. Link ↩ ↩2
-
Kosmyna, N. et al. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. MIT Media Lab, arXiv:2506.08872. Link ↩
-
Skitka, L. J., Mosier, K. L. & Burdick, M. (1999). Does automation bias decision-making? International Journal of Human-Computer Studies 51(5), 991-1006. Link ↩
-
Parasuraman, R. & Manzey, D. H. (2010). Complacency and bias in human use of automation. Human Factors 52(3), 381-410. Link ↩
-
Vicente, L. & Matute, H. (2023). Humans inherit artificial intelligence biases. Scientific Reports 13, 15737. Link ↩ ↩2
-
Bainbridge, L. (1983). Ironies of automation. Automatica 19(6), 775-779. Link ↩
-
Onnasch, L., Wickens, C. D., Li, H. & Manzey, D. (2014). Human performance consequences of stages and levels of automation. Human Factors 56(3), 476-488. Link ↩
-
Budzyń, K., Romańczyk, M., Kitala, D. et al. (2025). Endoscopist deskilling after exposure to artificial intelligence in colonoscopy: a multicentre observational study. The Lancet Gastroenterology & Hepatology. Link ↩ ↩2
-
Lee, H.-P., Sarkar, A., Tankelevitch, L. et al. (2025). The impact of generative AI on critical thinking. CHI 2025, Microsoft Research. Link ↩
-
Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies 15(1), 6. Link ↩
-
Kahneman, D. & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist 64(6), 515-526. Link ↩ ↩2
-
Bjork, E. L. & Bjork, R. A. (2011). Making things hard on yourself, but in a good way. In Gernsbacher et al. (eds), Psychology and the Real World. Link ↩ ↩2
-
Rumelt, R. (2011). Good Strategy Bad Strategy. Crown Business. Link ↩ ↩2
-
Duckworth, A. L. & Seligman, M. E. P. (2005). Self-discipline outdoes IQ in predicting academic performance of adolescents. Psychological Science 16(12), 939-944. Link ↩
-
Carr, N. (2010). The Shallows: What the Internet Is Doing to Our Brains. W. W. Norton. Link ↩
-
Lovett, N. C. (2026). The tragedy of the cognitive commons: How AI could disrupt the regeneration of professional expertise. Human Resource Development Review, online first. Link ↩
-
Storm, B. C. & Stone, S. M. (2015). Saving-enhanced memory. Psychological Science 26(2), 182-188. Link ↩
-
Knight, F. H. (1921). Risk, Uncertainty, and Profit. Houghton Mifflin. Link ↩
-
Véliz, C. (2026). Prophecy: Prediction, Power, and the Fight for the Future, from Ancient Oracles to AI. Doubleday. Link ↩
-
Poulidis, S., Bastani, H. & Bastani, O. (2025). Self-regulated AI use hinders long-term learning. Working paper, SSRN. Link ↩
-
Facione, P. A. (2011). Critical Thinking: What It Is and Why It Counts. Insight Assessment. Link ↩