The Ancestors Ran an Experiment, and the Variable Was Effort

Consider the great hill pilgrimages, among the largest gatherings of devotion on earth, drawing tens of millions of pilgrims each season. What is striking about them is how much of the hardship is deliberate. Before a pilgrim sets out, he keeps a long fast, often forty days of austerity: simple vegetarian food, abstinence, plain dress, bare feet, and a daily rhythm of prayer. He carries a cloth bundle of offerings balanced on his head for the journey. The final approach is made on foot, and the sanctum is reached only by climbing a steep flight of sacred steps, and only by those who have kept the vow.

None of this is logistically necessary. You could pave a road. The difficulty is not a relic of an older, poorer time that better technology would remove. The difficulty is the design. Anthropology has a name for what such an ordeal does to a person: liminality, the term Arnold van Gennep and later Victor Turner gave to the threshold state in a rite of passage, where the hard crossing is precisely what dissolves the old self and forms a new one (van Gennep, 1909; Turner, 1969). A pilgrimage is a machine for transformation, and effort is its fuel.

And here is the detail that matters most for everything that follows. The discipline is voluntary, and it is carried inside the person. No official inspects whether you fasted for those forty days. You did it because the discipline was yours.

The Struggle Is the Learning, Not the Price You Pay For It

For most of history this was intuition. It is now one of the most robust findings in the science of learning. The psychologist Robert Bjork named the principle a desirable difficulty: conditions that make learning feel slower and harder in the moment often produce stronger, more durable learning over time (Bjork, 1994). Ease is pleasant. Ease is also forgettable.

The cleanest demonstration is the testing effect. In a now-classic experiment, Henry Roediger and Jeffrey Karpicke had students learn prose passages either by rereading them or by putting the text away and trying to recall it. Five minutes later the re-readers looked like the winners, remembering 83 percent against the recall group's 71 percent. A week later the result had reversed completely: the re-readers could produce only 40 percent of the material, while the students who had struggled to retrieve it held 61 percent (Roediger and Karpicke, 2006). The effort of pulling knowledge out of your own head, rather than pushing it in again, is what makes it stay.

Look closely at that five-minute figure, because it is the whole danger in miniature. Rereading felt like mastery, and it produced the least durable learning of all. It is an illusion of fluency.

The same lesson turns up wherever it is tested. We remember what we generate better than what we are handed, the generation effect (Slamecka and Graf, 1978). And students made to wrestle with a hard problem before being taught the method, even when they fail to crack it, end up understanding the idea more deeply and transferring it more flexibly than students who are taught the method first, what Manu Kapur calls productive failure (Kapur, 2014). The struggle is not the toll you pay on the way to learning. The struggle is the learning.

Effort Is Also How Meaning and Character Are Made

That settles the learning half of the question. The second half, that effort is where satisfaction and meaning come from, has evidence just as old and just as strong.

In 1959, Elliot Aronson and Judson Mills ran a study that has never gone out of date. Volunteers underwent either a severe or a mild initiation to join a discussion group. The group was the same for everyone, and deliberately dull. Yet those who had endured the harder initiation rated it as more worthwhile (Aronson and Mills, 1959). Psychologists call it effort justification: we value things in proportion to what they cost us. Half a century later the same instinct showed up in the most ordinary setting imaginable, when people who assembled their own flat-pack furniture valued it far above identical ready-made pieces, the IKEA effect (Norton, Mochon and Ariely, 2012). The self-made shelf is not prized because it is flawless; it is prized because its maker can see their own effort in every slightly crooked joint.

The furniture studies carry the hinge of this whole argument: the effect appeared only when people finished the task. Stopped halfway, the affection vanished. Effort creates value only when it completes. That is the line between difficulty that builds a person and difficulty that merely punishes one, and walking it is the whole art of teaching.

This is the happiness you cannot download. A summit reached by helicopter is a photograph; a summit reached on foot is a memory you become. And it is more than happiness. A child who has earned something through difficulty learns a fact about themselves that no delivered result can teach: that they can meet hardship and come out the far side of it. That is character, and character is built on precisely the ground that instant gratification paves over.

A Tool Built for Productivity, Pointed at the Wrong Target

None of this is an argument against AI. It is among the most powerful tools we have ever built, and in education it is genuinely transformative: a tireless explainer, a patient reader of a struggling student's draft, a way to give one teacher the reach of ten. We cannot do without it, and we should not want to. The question was never whether to use it, but what we aim it at.

No. The claim is not that hardship is good. It is narrower and sharper than that: the specific difficulties that do the cognitive and moral work, retrieval, generation, productive struggle, seeing a thing through to completion, are exactly the ones a frictionless tool is most eager to remove. Keep those. Strip out the rest. The broken sandal on the trail is friction worth removing; the climb is not.

Here is the trap. AI is, above all, a productivity engine. It removes friction and delivers the finished result, and for most of the economy that is the entire point. But a school's product is not the finished essay. A school's product is the changed mind that should have been built in the writing of it, and the character formed in the difficulty. You can raise a student's output and lower their learning at the very same time, and from the outside the work will look better than it ever has.

The early evidence is sobering. A 2025 study of 666 people found a significant negative association between frequent reliance on AI tools and critical-thinking scores, with cognitive offloading, the habit of letting the tool do the thinking, as the mechanism in between (Gerlich, 2025). Correlation is not causation, but the direction is hard to wave away. More vivid still, a 2025 study at the Massachusetts Institute of Technology (MIT) Media Lab measured the brain activity of students writing essays with and without a large language model (LLM). The LLM group showed markedly weaker neural engagement, and asked moments later to quote a single sentence from the essay they had just produced, 83 percent could not. They had generated the work without ever owning it (Kosmyna and colleagues, 2025). The study is a preprint and its authors urge caution, but it puts a number on what every teacher already senses: the answer arrived, and the learner did not.

The evidence, at a glance

Colour key: blue, effort that strengthens learning; amber, a conditional effect; red, an early signal that offloading may weaken it. The 2025 findings are recent and, in the MIT case, not yet peer-reviewed: direction, not proof.

It is, for access. AI can hand every child the explanation that used to sit behind a paywall, and that is a real gain. But the thing wealthy families were buying was never the answer. It was the tutor who refused to give it and made the child do the work. Real equity is not equal access to answers. It is equal access to the struggle that builds a mind, and that is the harder thing to provision.

Abstinence Is a Habit, Not a Feature

Which invites the obvious fix: why not build the restraint into the tool? Make the AI refuse to write the essay, cap the hints, force the struggle first.

Some of that helps at the margin. But it mistakes where restraint actually lives. A tool that refuses to help is, in most situations, simply a broken tool, and a student who wants the answer will always find another tool that gives it. Restraint cannot be engineered into the software. It has to be cultivated in the person. This is the deepest thing the pilgrimage knows. The discipline was never enforced at the temple gate; it was carried up the hill inside the pilgrim. Abstinence is a habit, not a feature.

So the task for an educational institution is not to hunt for the AI that withholds. It is to form students who know when not to reach for it, the way a serious musician understands that the point of practice is not to produce the sound but to become someone who can. Responsible use of AI is not, in the end, knowing how to use the tool.

This reframes what good AI in education even looks like. The right model is not the servant with the plate. It is the guide on the trail. A good guide does not carry you to the summit, which would rob you of the climb. A good guide knows the mountain, sets the pace, catches you when a fall would be pointless rather than instructive, and, above all, knows when to step back and let you struggle. Learning science has described the good tutor this way for decades: as scaffolding that is deliberately withdrawn as competence grows (Wood, Bruner and Ross, 1976). Benjamin Bloom's celebrated finding that one-to-one tutoring can lift the average student above 98 percent of a conventional class came from tutors who diagnosed, stretched, and made students work, not tutors who supplied answers (Bloom, 1984). That is the real bar for AI in education: not the fastest route to the output, but the wisest companion on the climb.

Concretely, that is the difference between an assistant that writes the thesis statement and one that asks the question exposing why the student's draft does not yet hold, then makes the student close the gap themselves. The first hands over the summit. The second walks the trail beside them.

None of this is the default an institution will drift into. Almost every incentive that bears on a school pushes the other way: rankings, grade inflation, throughput, cost per seat, and the parent who wants to see polished work tonight. The frictionless path is the one the system quietly rewards, which means the difficulty that forms a student is now something an institution has to choose, and defend, on purpose.

What this asks of an institution

Preparing a Child for a World That Will Not Arrive on a Plate

There is a final reason this matters, and it reaches well past test scores into the kind of adults we are forming. Children who grow up with every answer delivered at the press of a button do not merely miss the content. They absorb a model of the world: that effort is optional, that results are instant, that wanting a thing and getting it are the same act. The world they will actually inherit runs on the opposite logic. It is volatile, uncertain, complex, and ambiguous, more so with each passing year, and it will not arrive on a plate. It will demand patience, a tolerance for not-knowing, and the willingness to stay with a hard problem long after the first easy answer has failed. Those are muscles. They are built only by use, and they waste away when a machine does the lifting.

A school that removes every struggle in the name of efficiency is not protecting its students. It is disarming them, sending them into an uncertain century trained for a world that does not exist, one where everything is easy and immediate and given. Preparing a child for the future cannot mean smoothing every difficulty out of the present. It has to mean choosing, with great care, which difficulties are worth keeping.

Our ancestors chose theirs deliberately. They could have cut a road to the top of the hill. Instead they kept the steep stone steps, and the forty-day fast, and the long walk, because they understood the one thing we are now at risk of automating away: that the climb is not the obstacle standing in front of the meaning. The climb is the meaning.

The question every institution now has to answer is a simple one. Will we sell our students the helicopter, or will we walk beside them, and have the wisdom to let them climb?

References

1. Aronson, E., & Mills, J. (1959). The effect of severity of initiation on liking for a group. Journal of Abnormal and Social Psychology, 59(2), 177-181.
2. Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing about Knowing. MIT Press.
3. Bloom, B. S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6), 4-16. Link
4. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. Link
5. Kapur, M. (2014). Productive Failure in Learning Math. Cognitive Science, 38(5), 1008-1022. Link
6. 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 (preprint, not yet peer-reviewed). Link
7. Norton, M. I., Mochon, D., & Ariely, D. (2012). The IKEA effect: When labor leads to love. Journal of Consumer Psychology, 22(3), 453-460. Link
8. Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249-255. Link
9. Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592-604.
10. van Gennep, A. (1909). The Rites of Passage; and Turner, V. (1969). The Ritual Process: Structure and Anti-Structure.
11. Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89-100.