The Rewards Without the Hunt
When AI does the practice, who does the thinking?
Human beings are built to learn from experience. Our brains change with what we go through, and the whole biochemistry of the body is tuned to react to a challenge and keep what it learns from it. Effort, error, the next attempt: that loop is how we became who we are.
Nothing in that machinery is there by accident, least of all the pleasure of getting something right. The rush of satisfaction we feel when we succeed, however small the success, is part of how we learn. Neuroscience has described it with remarkable precision: dopamine neurons respond most strongly not to a reward as such, but to an outcome that turns out better than we expected, which is exactly the signal a brain needs to update what it knows (Schultz, Dayan & Montague, 1997). The reward is how experience gets written into us.
I often think about this in the most basic terms. When our ancestors lived in caves and had to hunt to feed their families, every hunt taught them something: how the animal moved, how the weather turned, which plan failed and why. Now imagine that a piece of meat had appeared at the mouth of every cave each morning, just for asking. Life would have been easier for a while. But when the climate shifted and conditions turned hostile, as they did again and again, nobody would have known how to find food. We would probably not be here.
That is why I believe the technological revolution we are living through is different from the ones before it. It is centred on the outcome. It does not give us better spears, exoskeletons to run faster or shields to protect ourselves. It works inside our heads, on the very activity that made us adaptable. And it can do something no tool could do before: hand us the reward without the hunt.
This idea is at the centre of an essay I have been writing, Who Thinks For You? (The Cognitive Divide), and it rests on a handful of experiments. They matter as much for the people we lead as for the students who took part in them.
The Reward Without the Hunt
The first experiment contains the whole argument in miniature. Hamsa Bastani and colleagues gave nearly a thousand high school students in Turkey access to GPT-4 during their maths practice, in a randomized trial published in the Proceedings of the National Academy of Sciences (Bastani et al., 2025). While they had the AI, students did what you would expect: their practice scores rose 48% above the control group. Then came the exam, without the tool. The students who had practised with the chatbot scored 17% lower than those who had never had it.
Plus 48% with the tool, minus 17% without it. The AI improved their results and weakened their ability. And here is the part that connects back to the cave: the students did not notice. Asked afterwards, they believed they had learned more. They had received the reward, the feeling of success, without the experience that the feeling is supposed to mark.
The experiment had a second group, and that is where the hope lies. They used the same model, configured as a tutor: hints instead of answers, questions instead of solutions. During practice they did even better, 127% above the control. On the exam without AI, the loss that sank the first group practically disappeared. The tutor still gave them the satisfaction of solving the problem, but only after they had done the work.
One technology. Two configurations. Two very different outcomes for the people using it.
Didn't We Say the Same About Calculators?
It is a fair objection. Socrates worried that writing would weaken memory, and every generation since has feared that some new tool would make us lazy. But earlier tools were prostheses. The plough extended the arm, the book extended memory, the calculator executed a rule we already understood. All of them left the essential work of understanding, judging and deciding where it had always been, with us. Generative AI is the first tool that can take raw information and hand back a conclusion, with the feeling of having reached it ourselves. That is the frontier that has moved.
Extracting or Exercising
The line this creates does not separate people who use AI from people who don't. Access gaps close within a few years; they always have. The new line runs through the way we use it.
The first way is extraction: you ask the machine for the product. The answer, the draft, the decision. The thinking happens inside the machine, and you receive the result and sign it. Every use works, often better than what you would have produced yourself. What you don't see is the price: the process the machine ran is exactly the process your brain needed to run to stay in shape.
The second way is exercise: you ask the machine for the process. The hint, the objection, the correction. The thinking stays in your head; the machine structures it, challenges it and speeds it up, like a coach adjusting the load. Psychologists call this kind of useful effort, which feels slower but leaves more behind, a desirable difficulty (Bjork, 1994).
A second study shows how subtle the slide can be. Yizhou Fan and colleagues gave 117 university students a writing task with different kinds of support (Fan et al., 2025). The group using ChatGPT produced the best essays, better even than the group helped by a human expert, but learned no more than students who had no help at all. The researchers called the cause metacognitive laziness: what gets delegated first is not the task but the watching over it, the planning, checking and self-correcting.
The third study is the good news. At Harvard, Gregory Kestin's team built an AI tutor for a large introductory physics course and compared it, in a randomized trial, with active learning in class, one of the best forms of teaching we know (Kestin et al., 2025). Students learned more with the tutor, in less time. The detail that matters is how it was built: the tutor had every solution loaded and was not allowed to give them. One hint at a time, and the student had to try first. The team did not remove the effort. They designed it in.
Put the three side by side and the pattern is clear. The same technology pushes towards atrophy or towards amplification depending on one thing: where the effort happens. If it moves to the machine, the product improves and the person shrinks. If it stays with the person and the machine organizes it, both improve.
Why This Is a Leadership Question
Picture two people on your team who use an AI assistant for every report, every analysis and every difficult email. After three months both are producing better work, faster, and both feel good about it. Then the tool goes down for a day. One of them carries on. The other cannot. Nothing in the quarter's metrics would have told you which was which.
That is the workplace version of plus 48, minus 17. For leaders and learning teams, it turns a vague worry into three concrete decisions.
1. What to Delegate
Not every task needs to be a learning opportunity. Formatting a table, summarizing a long thread or translating a routine message are tasks where only the output matters, and handing them over is sensible. The useful test is whether a task is one where a person is expected to build or keep judgment: assessing a supplier, weighing a trade-off, writing the recommendation someone senior will rely on. Delegating those entirely can be efficient this quarter and expensive over several years. A simple start is to ask each team to name three tasks they will fully delegate and three they will keep doing themselves, because they want to stay good at them.
2. What to Keep Training
If the open chatbot is the problem and the tutor is the answer, the lesson for training is about configuration. Wherever people are learning something that matters, the AI they use should be set up to ask first: before giving me an answer, ask what I think and why; show me where my reasoning has a gap; give me a hint before the solution. Most major assistants already allow a standing instruction like this. It also means protecting some practice without the tool. Airlines keep pilots flying by hand even though the autopilot is excellent. Teams can do the same: work a problem through unaided, then compare the reasoning with the model's. The comparison is where the learning happens.
3. How to Measure the Difference
This is the hardest, because the usual metrics point the wrong way. Output volume, turnaround time and even quality ratings all improve with unrestricted use of the assistant, exactly as the practice scores did. Three simple indicators help: an unaided check, a short periodic task done without the tool and scored like the assisted version; an explanation test, in which the person walks a colleague through why a recommendation is right and what would change their mind; and a delegation log, a light record of what each person handed to the model this month. None of this needs surveillance or new software. It makes an invisible trade-off visible before it compounds.
What the Evidence Does Not Show
These studies involved students in structured courses over short periods. They do not show that professionals who use AI heavily lose skills at the same rate, or that the effects last for years. Experienced employees have a stock of judgment to draw on, and many tasks really do reward speed over practice. The honest conclusion is that the mechanism is credible and the early evidence is consistent with it, not that we know its size at work. That is a reason to measure, not to wait. Checking costs little. Discovering, after years of unexamined delegation, that a team can no longer evaluate what the model produces costs a great deal.
The Question to Take Back to Your Team
The lesson is not to use less AI. It is to decide which parts of the work the tool takes and which stay with the person, and to configure the tool to match. Used as an answer machine, AI is the meat at the mouth of the cave: the reward, without the hunt. Used as a tutor, it asks for the reasoning first, and the reasoning is what stays with you when the tool goes away.
So here is a question to try this week: of the work your team did with AI last month, which parts could each person still do, and explain, without it? If nobody knows, that is the first thing to find out.
A note on how this article was written: I drafted it with the help of AI tools, under my direction, and I reviewed the text and checked the cited studies against their published versions.
References:
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about knowing (pp. 185–205). MIT Press.
Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530.
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458.
Schultz, W., Dayan, P., & Montague, P. R. (1997). A neural substrate of prediction and reward. Science, 275(5306), 1593–1599. https://doi.org/10.1126/science.275.5306.1593
Recommended Resources:
Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In Psychology and the Real World. Worth Publishers.
Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
Mollick, E. (2024). Co-Intelligence: Living and Working with AI. Portfolio.
Oliver Henares is a strategy and innovation consultant based in Barcelona, with more than twenty years of experience in strategy, innovation, brand and foresight for companies including BMW, Microsoft, Nike and IKEA. He is co-founder of Mimetik, an AI studio, and teaches at the Culinary Institute of Barcelona. His essay Who Thinks For You? (The Cognitive Divide) examines how generative AI is changing the way people think. Contact: oliver@mimetik.ai
