I believe some degree of difficulty is necessary to make learning stick. This conviction, built up over years of practice, deserves serious scrutiny, especially now that the rise of artificial intelligence has changed the game.
The trial, an old method of learning
Humans have always learned through a succession of attempts, most often followed by mistakes, and therefore through an accumulation of difficulties, sometimes frustrations. Software developers know this feeling well: try, try again, rewrite the code, until suddenly the problem gives way. What satisfaction, in that exact moment. Athletes feel the same thing when they train relentlessly, grit their teeth, sweat, hurt themselves to refine a movement and shave off a tenth of a second. Musicians know it too, in those hours spent on a stubborn passage, loosening fingers that refuse to obey. Sport, music, computing: so many fields where learning is not handed over, it is earned.
Entrepreneurship is no exception to this rule, and is probably one of the fields where it plays out most consistently, as Nassim Taleb’s essay on responsibility and risk reminds us. Building a company means facing, every day, a share of decisions to make under uncertainty, risks to take on, mistakes to correct on the fly.
Philosophers already understood this
Aristotle formulated this intuition more than two thousand years ago: we do not become courageous by reading treatises on courage, but by performing courageous acts. Virtue, like skill, is acquired through repeated practice, through what he called a hexis: a disposition forged in action, not bookish knowledge. You do not become an entrepreneur by studying entrepreneurship, but by weathering your own failures.
The philosopher of science Gaston Bachelard, later, went as far as making the obstacle the very condition of knowledge. In La formation de l’esprit scientifique (1938), he shows that the scientific mind is never built through the calm accumulation of facts, but always against something, against a resistance that must be identified and overcome. Difficulty is therefore not a regrettable accident along the path of learning. It is, structurally, its engine.
Socrates, finally, made it the very method of knowledge: in his dialogues, he never handed over a ready-made answer, but led his interlocutor to aporia, that moment of confusion where one no longer knows. It is precisely out of this confusion that genuine understanding is born.
What science confirms
These ancient intuitions are confirmed by contemporary research. As early as 1994, psychologist Robert Bjork, of UCLA, theorized the notion of desirable difficulties: adding a dose of challenge to a learning activity produces better long-term retention, even though it apparently slows down immediate performance. Bjork drew a clear distinction between performance, what one is capable of doing here and now, often with help, and genuine learning, what remains once the help is removed. This distinction echoes the broader one Daniel Kahneman draws between our two systems of thought: the comfortable speed of system 1 and the effort of system 2, the one that, precisely, is what makes learning stick. This distinction becomes decisive in the age of AI.
Closer to us, a study by the MIT Media Lab published in June 2025 gave this question an almost clinical treatment. Researchers followed fifty-four people across three sessions spaced four months apart, split between writing with no assistance at all, writing with a standard search engine, and writing with a language model. Result: the AI-assisted group produced essays the fastest, but with the weakest neural connectivity between the areas associated with attention, working memory and decision-making, that is, the set of functions researchers group under the term executive engagement. More striking still: more than eighty percent of ChatGPT users proved unable to quote a single passage from the text they themselves had written just minutes earlier. The researchers speak of genuine cognitive debt: what assistance spares us today, it makes us pay for tomorrow, in the form of difficulty recovering, unaided, a full and complete mental activity.
Lev Vygotsky’s zone of proximal development describes the space where this work happens: what the learner cannot yet do alone, but can accomplish with help. It is in this zone of tension, neither too easy nor too hard, that progress happens.
The age of AI: what if difficulty disappeared?
This is what makes my own current experience so unsettling. I am a daily and convinced user of artificial intelligence. Yet I observe a clear shift in my own coding practice: I barely write code anymore. I steer artificial intelligences instead, a shift I describe in more detail in my take on AI and software development after 40 years of code. And I feel conflicting emotions about it. On one side, a sense of power and performance, perhaps partly illusory. On the other, the loss of that particular satisfaction of finding the solution myself after hours of effort. I sometimes wonder whether my exchanges with Claude, ChatGPT or other models short-circuit the Socratic aporia, handing me the answer before I have had time to get a little lost in the problem.
Nietzsche saw in the overcome obstacle the very material out of which a solid self is forged. His best-known formula, “what does not kill me makes me stronger,” is not a mere witticism: it carries the idea of a self-overcoming that can only happen through a resistance genuinely experienced. What, then, becomes of a mind from which all resistance is systematically removed? This is precisely the question I ask myself when I probe a possible intellectual laziness, comfortably settled behind the alibi of increased performance.
My conclusion, without a verdict
So should we give up artificial intelligence and go back to traditional learning methods? I do not believe so, and I have no wish to, for that matter: it lets me go faster and further in my projects. But I actively try to preserve this share of difficulty, even if it means displacing it, perhaps unconsciously, toward other territories: sport, an artistic practice, anything artificial intelligence cannot yet help me accomplish in my place.
This may be where the real answer lies. To collaborate effectively with an artificial intelligence that relieves us of so much effort, we probably need to preserve, elsewhere, the habit of difficulty, of hurting ourselves a little from time to time. And it is perhaps precisely in the fields where AI is not yet present that we forge the experience that will let us, tomorrow, steer these systems as something other than mere spectators of our own delegated thinking.
As Malebranche wrote: “It is by applying oneself that one perfects oneself.” The question is perhaps not whether difficulty is necessary, it is. But rather choosing, consciously, the trials we consent to, a question that echoes, in its own way, that of the Adaptation Coefficient: adapting without ever giving up the effort that makes change genuinely acquired.
Related articles: