Why AI Rewards Critical Thinking and Imagination

I was the kid with a Raymond E. Feist saga open under the desk. School and I never agreed on terms. The formal channel — abstract instruction, maxims delivered dry, “consider other perspectives” written on a slide as though the act of reading the words were the skill — went in one ear and out the other. Fiction was the escape hatch: not non-fiction, not the improving book, but thousands of invented lives stacked quietly end to end. I left school convinced I was bad at thinking, because the only kind of thinking school measured was the kind I couldn’t do.

Whatever I actually learned, I learned by doing. Photography first, then the editing that makes a photograph worth keeping, then a wedding photography business built on both — each one picked up in the field rather than a classroom, shot badly and shot again until it wasn’t. I taught myself SEO at thirty-three, on live client work, and it became the spine of what I do now. None of it arrived through a syllabus.

Ideas were never the scarce part for me; execution was. Without it, an idea is just weather — the bottleneck was always hands, hours, the slow tax of turning a hypothesis into something you can show. The LLM harness deleted that bottleneck. I can now run, in an afternoon, the kind of experiment that used to take a quarter. The flattering explanation is that AI multiplies people like me — that it takes talent and scales it. I went to the research expecting that to be confirmed.

The research says I am half wrong.

The leveling result

On routine output, AI is not a multiplier. It is a leveler. Noy and Zhang ran the cleanest test: 453 professionals, randomized access to ChatGPT on real writing tasks. Time fell by about 40%, quality rose by about 18% — and the weakest writers gained the most. Initial inequality was nearly half-erased^[1]. Most treated participants simply pasted the model’s output, and the essays they edited scored no better than the raw ChatGPT draft. The model, the authors concluded, mostly substitutes for worker effort.

Doshi and Hauser found the same shape in creativity: 293 writers given AI story ideas. With five ideas, novelty rose 8.1% and usefulness 9.0% — but the entire gain concentrated in the least-creative writers; the most inherently creative writers gained almost nothing^[2]. Dell’Acqua and colleagues found it a third time at BCG: 758 consultants, +40% quality, and the bottom half of performers gained 43% against 17% for the top half^[3].

This is the concession I have to make before anything else. On fluency, on polish, on competent output, AI amplifies everyone — and it amplifies the people with the least the most. It does not multiply talent so much as raise the floor and squeeze the middle toward a competent average.

So what did it multiply?

But the same BCG study had a second half, and it is the one that matters. On a task deliberately designed so the AI would mislead — a case outside the “frontier” of what the model can do safely — users of the AI were 19 percentage points more likely to produce a wrong answer than people working unaided^[3]. On the easy task the model closed the gap; on the hard task it inverted it.

Lee and colleagues sampled 319 knowledge workers across 936 real episodes of AI use and found the same pattern from the other direction: the more confident people were in the AI, the less critical thinking they did; the more confident they were in themselves, the more they did^[4]. The work of thinking hadn’t vanished — it had migrated. It now lives in verification (is this true?), integration (does this fit?), and stewardship (is this the right thing to want?). The authors are explicit that this is correlational, not causal. But the direction is consistent with everything above.

AI collapsed the price of production. It did not collapse the price of judgment. What it multiplied, for someone like me, was not skill but direction — the ability to point the machinery at the right question and to refuse it when it answers confidently and wrong. The bottleneck moved upstream, to exactly the two things I was never short of: knowing what to ask, and noticing what shouldn’t be believed.

The sameness problem

There is a second finding in Doshi and Hauser that I keep coming back to. The stories that writers produced with AI help didn’t just get a little better — they got more similar to each other, by 8.9 to 10.7%^[2]. Individually, every writer was better off. Collectively, the field drifted toward the same average. The authors called it a social dilemma: rational for each, bland for all.

This is the bridge to my own working life. I work in AI search and content. When everyone drafts from the same model — when the same scaffolding generates the same outlines, the same hedged framings, the same “here’s why this matters” closer — output regresses to the model’s mean. Distinctiveness becomes the only scarce asset left, and distinctiveness is precisely the part the model does not supply. It cannot imagine what it doesn’t already contain; it can only recombine its training data along well-trodden paths. Which means the thing that separates one piece of work from the next is no longer execution. It is whatever you brought that the model couldn’t.

Can the gap be closed? Partly

The honest question, for anyone who reads this far, is whether the two capacities at the bottom of all this — critical thinking and imagination — can be taught. I went into the literature expecting a clean yes. What I found is a qualified, revealing, half-yes.

Critical thinking is trainable, but modestly, and mostly inside a domain. Abrami and colleagues synthesized 341 studies: generic critical-thinking instruction produces an effect of g = 0.30 — small. Embed the same instruction inside a substantive discipline and it nearly doubles, to g = 0.57^[5]. The formats that worked combined explicit principles with applied practice: dialogue, authentic problems, mentoring. Dispositions — the wantingto think — barely moved at all: g = 0.23, the weakest of every outcome they measured.

Debiasing training does a little better than the genre’s reputation suggests. Morewedge and colleagues found that a single interactive session cut committed biases by 31.9–46.3% immediately, and still by 23.6–34.8% at eight to twelve weeks — and the effects generalized across problem formats people had never been taught^[6]. Sellier, Scopelliti, and Morewedge followed it into the field: graduate business students who’d played the debiasing game were 19% less likely to pick the confirmation-biased solution on an unannounced case weeks later^[7]. Chang and colleagues showed that under an hour of probabilistic-reasoning training improved forecasting accuracy by 6–11% (in Brier score) and held up across four years^[8].

Across every one of these results, one pattern holds: abstract instruction is the weakest form. Domain-embedded practice beats maxims; realistic exercises beat inspiration; interactive games with feedback beat videos^[5][6][9]. Willingham argues flatly that critical thinking is not a generic, transferable skill — it rides on stored domain knowledge, and the maxims do little without it^[9]. School taught thinking the one way that barely works. My struggle with formal education wasn’t a capacity problem. It was a mismatch with the delivery mechanism.

Creativity training tells a similar, slightly harsher story. Scott, Leritz, and Mumford’s synthesis found strong effects on test performance — overall Δ = 0.68, divergent thinking 0.75, creative problem solving 0.84 — but only for programs teaching specific cognitive heuristics through realistic domain exercises; inspiration-only approaches did worse^[10]. The catch is transfer. McKay and colleagues found that organizational creativity training reliably moves learning outcomes (g = 0.73) but transfers to actual on-the-job behavior at a nonsignificant g = 0.34, with gains decaying from 0.86 to 0.40 over time^[11]. And Kim, in the study I had to get the DOI right for, found that divergent-thinking scores predict real creative achievement at only r = 0.216^[12]. The thing training improves is a weak proxy for the thing that matters.

Stanovich and West round it out: actively open-minded thinking predicts the ability to evaluate arguments independent of prior belief, over and above cognitive ability^[13]. Intelligence anchors part of the variance; disposition anchors another part. Both are sticky.

The two capacities don’t even fail training the same way, which is worth admitting. Critical thinking’s wall is disposition: the willingness to use the skill barely moves. Imagination’s is transfer: the gains are real, even large, but they sit on tests and rarely reach real work. Different walls, same practical result — the technique is teachable; what decides whether it counts mostly isn’t.

A hypothesis about where imagination comes from

Schacter and Addis propose that imagining the future is a recombination of episodic memory: we assemble new experiences out of parts of remembered ones, and patients with hippocampal amnesia are impaired at imagining new experiences at all^[14]. Madore, Addis, and Schacter showed that a brief exercise in detailed recollection selectively boosted divergent thinking and the episodic richness of imagined future events^[15]. The capacity to imagine seems to run on the same machinery as the capacity to remember, and it draws on the library of specifics you’ve stored.

I want to be careful here, because the verified evidence doesn’t contain a study linking fiction reading to imagination in adults, and I’m not going to pretend it does. What I’ll offer instead is a hypothesis, narrower than a finding: if imagination is the recombination of episodic memory, then what feeds it is raw material — diverse, specific, vivid episodes stored against later use. By that logic the general principle matters more than my biography. A reader of dense non-fiction, a traveller who keeps notes, a mechanic who has rebuilt a hundred engines — anyone who stocks the library with particulars is doing the same thing. The fiction I read was one particularly immersive way to accumulate them: continents with their own histories, magic bound by its own rules, characters who wanted something, failed at it, and recovered, or didn’t. Every volume was another set of episodes to draw on. The model recombines its training data; I recombine my life. The lesson I take from that is not to outsource the noticing.

Nobody is immune, including me

I am not the hero of this story. Cau and Spano ran a setup where people got AI help, and some of them also got calibration feedback — told, essentially, when their confidence was miscalibrated. The two-stage system did improve calibration. It also increased overreliance on the AI^[16]. Even being told you’re over-trusting doesn’t stop you over-trusting; in this design it nudged you toward more of it. Overreliance is the default drift.

The practical upshot for my own work is unglamorous. The value I get from the harness scales with how hard I argue with it. Every draft it produces is a draft I treat as suspect until I’ve checked the citations against primary sources, traced the claims to the paragraph that supports them, and asked where it would be confidently wrong if it were going to be. Most of the work is the arguing. The model does the easy part.

Handle the scary studies the way the essay preaches

The field’s scariest findings are exactly the ones a careful reader should distrust.

Gerlich’s 2025 paper is the one most often quoted as proof that AI use erodes critical thinking, reporting a correlation of r = −0.68. That number exists in the paper. But the paper contradicts itself: its own full correlation matrix gives −0.49 for the same pair, and the other pairs disagree between tables by similar margins. No p-values attach to the headline table. It’s cross-sectional, UK-only, based on self-report items, and a named critique has called out its causal language. I don’t lean on it^[17].

Kosmyna and colleagues’ “Your Brain on ChatGPT” EEG study — the one that coined “cognitive debt” — had 54 participants, 18 in the crossover session, and is a preprint. The authors published a FAQ begging journalists not to write that ChatGPT “makes your brain lazy or dumb”^[18]. It’s hypothesis-generating, not established.

A fair rebuttal is that the case for erosion is overdrawn. Disposition is partly trainable — g = 0.23 is small but not zero — and the debiasing and forecasting results show that focused, interactive practice can move behavior for weeks; small gains, practiced repeatedly in the same domain, compound. Nor does anything in the evidence fix AI as a substitute for judgment. It could be designed to scaffold it: to surface its own uncertainty, demand a verification step, make the noticing cheaper. That is a real possibility, and the studies here describe the tools we have, not the tools we could build. But on the present tools the direction is the one Lee and colleagues measured: more confidence in the AI, less critical thinking. The disposition is sticky, not fixed; how the tools get built and used decides which way it drifts.

The disposition this essay is defending is the disposition needed to read the research about it.

What the detour built

Competence, polish, fluency: free. The model gives them to everyone, and gives them most generously to the people who had the least. Judgment: expensive, and rising, because the same model that hands you a clean draft will hand you a clean wrong answer on the hard task and feel exactly as confident about both. And the wanting to think — the disposition to argue with the confident wrong answer, to refuse the first good-enough draft, to notice the gap between what the model contains and what the work needs: not for sale. The literature says you can teach the skill and mostly can’t teach the wanting, and AI hasn’t changed that. It has only made the difference more consequential.

The arc closes with the kid under the desk. School was optimized, with the best institutional intentions, for the precise thing AI just made nearly free: competent, fluent, rule-following output on abstract prompts. It drilled that, and it punished the daydreaming — the fiction, the hands-on tinkering, the refusal to learn from maxims — as though it were the opposite of real learning. What the research now says is that the daydreaming stocked the library, the tinkering was the format that actually works, and the wanting, the part that wouldn’t sit still for a slide, is the one thing the model can’t give you and the one thing the price list rewards most.

The wanting survived school, I think, because it was never schooled. It was the part of me that was busy with something else while the formal channel did its best with me. AI made the formal channel’s product free. It left exactly what the detour built.

References

1. Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586
2. Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), Article eadn5290. https://doi.org/10.1126/sciadv.adn5290
3. Dell’Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838
4. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & 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. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1–22). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713778
5. Abrami, P. C., Bernard, R. M., Borokhovski, E., Waddington, D. I., Wade, C. A., & Persson, T. (2015). Strategies for teaching students to think critically: A meta-analysis. Review of Educational Research, 85(2), 275–314. https://doi.org/10.3102/0034654314551063
6. Morewedge, C. K., Yoon, H., Scopelliti, I., Symborski, C. W., Korris, J. H., & Kassam, K. S. (2015). Debiasing decisions: Improved decision making with a single training intervention. Policy Insights from the Behavioral and Brain Sciences, 2(1), 129–140. https://doi.org/10.1177/2372732215600886
7. Sellier, A.-L., Scopelliti, I., & Morewedge, C. K. (2019). Debiasing training improves decision making in the field. Psychological Science, 30(9), 1371–1379. https://doi.org/10.1177/0956797619861429
8. Chang, W., Chen, E., Mellers, B., & Tetlock, P. (2016). Developing expert political judgment: The impact of training and practice on judgmental accuracy in geopolitical forecasting tournaments. Judgment and Decision Making, 11(5), 509–526.
9. Willingham, D. T. (2008). Critical thinking: Why is it so hard to teach? Arts Education Policy Review, 109(4), 21–32. https://doi.org/10.3200/AEPR.109.4.21-32
10. Scott, G., Leritz, L. E., & Mumford, M. D. (2004). The effectiveness of creativity training: A quantitative review. Creativity Research Journal, 16(4), 361–388. https://doi.org/10.1207/s15326934crj1604_1
11. McKay, A. S., Reiter-Palmon, R., Coombes, S. M. T., & Coombs, J. E. (2024). A meta-analysis of creativity training in organizational settings. Creativity and Innovation Management, 33(4), 587–602. https://doi.org/10.1111/caim.12605
12. Kim, K. H. (2008). Meta-analyses of the relationship of creative achievement to both IQ and divergent thinking test scores. The Journal of Creative Behavior, 42(2), 106–130. https://doi.org/10.1002/j.2162-6057.2008.tb01290.x
13. Stanovich, K. E., & West, R. F. (1997). Reasoning independently of prior belief and individual differences in actively open-minded thinking. Journal of Educational Psychology, 89(2), 342–357. https://doi.org/10.1037/0022-0663.89.2.342
14. Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: Remembering the past and imagining the future. Philosophical Transactions of the Royal Society B: Biological Sciences, 362(1481), 773–786. https://doi.org/10.1098/rstb.2007.2087
15. Madore, K. P., Addis, D. R., & Schacter, D. L. (2015). Creativity and memory: Effects of an episodic-specificity induction on divergent thinking. Psychological Science, 26(9), 1461–1468. https://doi.org/10.1177/0956797615591863
16. Cau, F. M., & Spano, L. D. (2025). Beyond awareness: Investigating how AI and psychological factors shape human self-confidence calibration [Preprint]. arXiv. https://arxiv.org/abs/2511.17509
17. Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. https://doi.org/10.3390/soc15010006
18. 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 [Preprint]. arXiv. https://arxiv.org/abs/2506.08872