The Part That Stays Mine

Most days I work with a model open in a second window. It drafts, it lists options, it fetches the thing I half-remember, and I keep my hands free for the part I actually care about, which is deciding what is any good. It feels like a clean division of labor, and mostly it is. But every so often I catch a smaller, less comfortable thing: a draft I approved too fast, a direction I took because it was already written rather than because it was right. Nothing dramatic. Just a quiet sense that some muscle I used to use had gone slightly slack.

I went looking for the science on what that muscle is, and found that it is not one muscle but four, each studied for decades under its own name. What follows is the map I built, and the argument it left me with. The thesis is simple to state and harder to live: the durable value in AI-augmented work is not the execution the machine took over. It is the attention, the intuition, the insight, and the judgment I bring to deciding what the execution was for. Offloading the first is fine. Offloading the second is how you quietly stop being worth augmenting.

The cheap part and the expensive part

In an earlier piece I argued that the intuition is the cheap part and the testing is what makes it knowledge. This is the same argument, aimed the other way. If a model can generate the draft, the layout options, the first pass at the research, then generation is now cheap for me too. What stays expensive is the framing that decides which draft to ask for, and the judgment that decides whether the one I got is right. So the interesting question stopped being "what can it do" and became "what is the expensive part it cannot do for me, and am I keeping it sharp."

The cognitive science answers that in layers. Focus is the substrate everything else runs on. Intuition is what fast recognition feels like from the inside, when it has been earned. Insight is the sudden change of representation that gets you past a wall. And metacognition is the layer that watches all three and decides which to trust. You could read the whole literature as a description of the expensive part, written before there was a cheap part to contrast it with.

Focus is the thing that has to survive first

Everything downstream depends on protecting a small pool of resources, so focus comes first. Working memory is severely limited when it handles anything genuinely new, which is the core of John Sweller's cognitive load theory (1988). Expertise, in that account, is mostly the accumulation of schemas that let you treat what used to be ten separate things as one chunk, so the load drops. This is why a familiar problem feels light and a novel one feels heavy, and why externalizing memory onto notes and diagrams is not a crutch but good practice.

The more uncomfortable finding is about switching. Sophie Leroy's work named attention residue, "the persistence of cognitive activity about a Task A even though one stopped working on Task A," and showed the residue is worst when Task A was left unfinished (2009). Part of your mind stays stuck on the thing you did not close. A notification-driven workday is expensive for a reason you can measure. Cal Newport built the popular case for protecting concentration in Deep Work (2016), and Mihaly Csikszentmihalyi described the far end of it, the absorbed state he called flow (1990), where challenge and skill are matched and self-consciousness drops away.

I wrote a whole essay about how my own attention actually behaves, so I will not relitigate it here. The point for this argument is narrower. A model does not fragment my attention the way notifications do, but it offers a subtler version of the same cost: it makes it very easy to not finish the thought myself, to hand off the hard middle of a problem the moment it gets heavy. The residue then is not from switching tasks. It is from never having fully held the task in the first place.

Intuition is recognition, and recognition has conditions

The thing designers call taste, the fast sense that one option is right before you can say why, has a boring and reassuring explanation. Gary Klein's fieldwork with firefighters and nurses produced the recognition-primed decision model (Sources of Power, 1998), where experienced people mostly do not compare options. One commander told him, "I don't make decisions." He saw the situation as an instance of a pattern he already knew, recognized a workable move first, and mentally simulated it before acting. Intuition, in this view, is recognition, not magic.

Recognition has a memory substrate. Chase and Simon's chess study90004-2) showed masters could reconstruct real game positions far better than novices, but had no advantage at all on random ones (1973). The masters were not seeing more, they were seeing in chunks, meaningful patterns stored by the thousands. Expertise is a large library of those chunks, which is what lets recognition be both fast and, under the right conditions, correct.

Under the right conditions is the whole game. Kahneman and Klein, a skeptic and an advocate, ran an adversarial collaboration and agreed on when to trust a gut call: only in an environment regular enough to contain valid cues, and only after enough practice with real feedback to have learned them (2009). Their sharpest conclusion is the one I keep taped to the inside of my head. Subjective confidence is not a reliable indicator of accuracy. A warm feeling of certainty is not evidence that you are right. This is exactly the test I need when a model hands me something plausible: not "does this feel right," but "is this a domain where my feeling has earned the right to be trusted, and have I actually checked." (One honest caveat on the practice side: the 1993 deliberate-practice claim that experts are mostly made by effortful practice was partly walked back by a 2019 replication, which found it matters but explains less variance than first claimed. Practice builds the library. It is not the only thing that does.)

The glimpse before the answer

The part I find most beautiful is also the best evidence that some of my thinking happens without me. When a solution is close, can you feel it before you can state it? The answer turns out to depend on what kind of problem it is. Metcalfe and Wiebe had people rate their "warmth," their felt closeness to a solution, every fifteen seconds. For step-by-step problems, warmth rose smoothly as they approached the answer, so their sense of progress was accurate (1987). For insight problems, warmth stayed flat and then jumped at the last moment. The answer arrived as a flash they could not see coming.

The brain shows the same seam. Using the compound-remote-associate task, Kounios and Beeman found that insight solutions carry a burst of high-frequency gamma activity around 40 Hz over the right anterior superior temporal gyrus, roughly 300 milliseconds before the person presses the button to say they have it, preceded by an alpha-band quieting over visual cortex, as if the brain briefly dims the outside world to hear a faint internal signal (2014). The gamma burst appears for insight and is absent for the same answer reached analytically. Insight is consciously abrupt but unconsciously prepared. The answer is assembled below awareness and then handed up.

Two practical things follow, both of which I now trust more than I used to. First, incubation is real. Sio and Ormerod's meta-analysis confirmed that setting a problem aside improves later solutions, with longer prior struggle producing a larger effect and low-demand breaks helping more than high-demand ones (2009). In art high school we used to step back from a canvas for a few days on purpose, because prolonged immersion makes you lose your grip on it and you have to come back with new eyes. The lab result is that stepping back is not avoidance, it is processing. Second, Stellan Ohlsson's representational-change theory explains the wall itself: an impasse is caused by a starting representation that quietly excludes the answer, and getting past it requires relaxing a constraint you imposed on yourself or breaking a whole into recombinable parts. He also names partial insight, where you break the impasse just enough to resume stepping forward. That is the closest thing I have found to a description of what a good glimpse actually does. It does not hand you the answer. It reopens the search.

This is the faculty most exposed to offloading, because a model is very good at ending an impasse for you before your own representation has had time to change. You get unstuck without ever having restructured anything, which is faster and, for the kind of problem where the restructuring was the point, worse.

The architect layer

Above the other three sits the layer that decides among them. John Flavell named metacognition, "cognition about cognitive phenomena," the monitoring and regulation of your own thinking (1979). It is the capacity to know whether a strategy is working, to switch when it is not, and to decide when to trust an intuition, when to push, and when to set a thing down and let it incubate. Everything above is only useful if this layer is running, because this is the layer that applies the Kahneman and Klein test instead of just having a feeling.

Widen the same move from a single problem to a whole practice and you get systems thinking, Peter Senge's discipline of seeing structures rather than events and finding the leverage points where a small, well-placed change produces a large one (The Fifth Discipline, 1990). The classic methodology sits under both: Newell and Simon framed problem solving as search through a problem space with means-ends analysis (1972), and Gick and Holyoak showed how analogy supplies candidate structures90013-4), while also showing that people routinely fail to notice a relevant analogy without a cue (1980). What I like to call the architect layer is just this: the move from executing each instance to monitoring the process and designing the structure that produces the instances. It is also, not coincidentally, the exact layer the work migrates to once a machine can execute.

The machine, and the debt

There is a respectable philosophical warrant for treating a model as part of my thinking rather than a tool beside it. Clark and Chalmers argued in "The Extended Mind" that an external resource which is reliably available, easily accessed, and automatically trusted functions as a genuine part of the cognitive system (1998). By that standard a model I reach for a hundred times a day is not a lookup, it is cognition, extended. I find that honest. It is also the reason the risk is real rather than sentimental, because if the thing is part of my cognition, then how I use it changes my cognition.

The upside is measured and large. The Harvard and BCG field experiment with 758 consultants using GPT-4, now published in *Organization Science*, found that on tasks inside the model's range, consultants completed 12.2% more work, 25.1% faster, at 40% higher quality, with the biggest gains going to the weakest performers (Dell'Acqua et al., 2023/2025). But on a task deliberately placed outside the model's range, AI users were 19 percentage points less likely to get it right, because the frontier of what the model can do is what they call jagged: it excels at one task and quietly fails at a very similar-looking one. They describe two working styles that succeed, which map cleanly onto my own habits: Centaurs, who split the work by task, and Cyborgs, who interleave it move by move. Either way the durable human job is the same, and it is the architect layer. Frame the problem, sense which side of the jagged edge a task falls on, and judge the output.

The downside is where I have to be careful, and where the evidence is newer and softer, so I will flag it as such. A survey of 319 knowledge workers found that higher confidence in the AI predicted less critical thinking, while higher confidence in oneself predicted more, and that people reported critical thinking felt like less effort when a model was in the loop (Lee et al., CHI 2025). A separate correlational study of 666 people reported a similar association between heavier AI use and lower critical-thinking scores (Gerlich, 2025), with the usual caveat that correlation is not cause. And an MIT Media Lab EEG study had 54 people write essays with a model, a search engine, or nothing, and found that the brain-only writers showed the strongest, most distributed neural connectivity and the model users the weakest, that the model users reported the lowest ownership of their essays, and that they often could not quote a line they had just "written" (Kosmyna et al., 2025). That last study is a small, non-peer-reviewed preprint and a published comment has urged a more conservative reading, so I hold it as suggestive, not settled. The authors call the pattern cognitive debt, the interest you pay later for offloading you did not need to do now.

Put the measured upside and the soft downside together and you do not get a verdict, you get a discipline. The gains are real and they live inside the frontier. The erosion is plausible and it comes from crossing the frontier without noticing, or from offloading the expensive part because the cheap part was right there.

What stays mine

So I have a rule now, and it is smaller than the reading that produced it. Offload execution, keep judgment. Use the model for drafts and options and search, and keep problem framing, intuition, insight, and the metacognition that governs them for myself, on purpose, including the parts that would be faster to hand over. Run the Kahneman and Klein test before I trust either my gut or the model's fluent output: is this a domain where I have earned real feedback, and have I actually checked. Treat an impasse as a signal to change my own representation before I ask for one to be handed to me. And watch the one indicator that turned out to matter most.

That indicator is ownership. The MIT writers who had leaned hardest on the model could not quote their own sentences, and I think that is the whole warning in one image. If my confidence in the tool is rising while my ability to explain my own work is falling, that gap is the debt coming due, and the fix is boring and reliable: go back and do a hard thing unaided until I can feel the muscle again. The essays across my file-ownership and my voice were both, underneath, about the same instinct, and I only see it now that there are three of them. The thing worth protecting was never the files or even the voice. It was the part of the thinking that stays mine when the machine does the rest, and the discipline of noticing the moment it starts to leave.

References

• Chase, W. G., & Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4(1), 55–81. https://doi.org/10.1016/0010-0285(73)90004-2
• Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7
• Csikszentmihalyi, M. (1990). Flow: The psychology of optimal experience. Harper & Row.
• Dell'Acqua, F., McFowland, E. III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2025). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Organization Science. https://doi.org/10.1287/orsc.2025.21838
• Ericsson, K. A., Krampe, R. Th., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406. https://doi.org/10.1037/0033-295X.100.3.363
• Flavell, J. H. (1979). Metacognition and cognitive monitoring. American Psychologist, 34(10), 906–911. https://doi.org/10.1037/0003-066X.34.10.906
• 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 (correlational; causality not established)
• Gick, M. L., & Holyoak, K. J. (1980). Analogical problem solving. Cognitive Psychology, 12(3), 306–355. https://doi.org/10.1016/0010-0285(80)90013-4
• Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526. https://doi.org/10.1037/a0016755
• Klein, G. (1998). Sources of power: How people make decisions. MIT Press.
• Kosmyna, N., et al. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv preprint. https://doi.org/10.48550/arXiv.2506.08872 (n = 54; non-peer-reviewed preprint; treat as suggestive)
• Kounios, J., & Beeman, M. (2014). The cognitive neuroscience of insight. Annual Review of Psychology, 65, 71–93. https://doi.org/10.1146/annurev-psych-010213-115154
• Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking. CHI 2025. https://doi.org/10.1145/3706598.3713778 (n = 319; self-report)
• Leroy, S. (2009). Why is it so hard to do my work? Organizational Behavior and Human Decision Processes, 109(2), 168–181. https://doi.org/10.1016/j.obhdp.2009.04.002
• Macnamara, B. N., & Maitra, M. (2019). The role of deliberate practice in expert performance: Revisiting Ericsson, Krampe & Tesch-Römer (1993). Royal Society Open Science, 6(8), 190327. https://doi.org/10.1098/rsos.190327 (replication; smaller effect)
• Metcalfe, J., & Wiebe, D. (1987). Intuition in insight and noninsight problem solving. Memory & Cognition, 15(3), 238–246. https://doi.org/10.3758/BF03197722
• Newell, A., & Simon, H. A. (1972). Human problem solving. Prentice-Hall.
• Newport, C. (2016). Deep work: Rules for focused success in a distracted world. Grand Central.
• Ohlsson, S. (1992). Information-processing explanations of insight and related phenomena. In M. Keane & K. Gilhooly (Eds.), Advances in the psychology of thinking (Vol. 1, pp. 1–44). Harvester-Wheatsheaf.
• Senge, P. M. (1990). The fifth discipline: The art and practice of the learning organization. Doubleday/Currency.
• Sio, U. N., & Ormerod, T. C. (2009). Does incubation enhance problem solving? A meta-analytic review. Psychological Bulletin, 135(1), 94–120. https://doi.org/10.1037/a0014212
• Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285. https://doi.org/10.1207/s15516709cog1202_4