Learning Hurts: How AI Is Robbing Us of the Struggle
TL;DR:
- Educational research consistently shows that deep learning requires effort, mistakes, and deliberate practice. Without these, long-term retention fails.
- Widespread AI usage is causing "cognitive atrophy": we delegate thinking to algorithms and our skills erode without us noticing.
- History shows this pattern clearly — the printing press, the calculator, television, GPS — what matters is never the tool itself, but the conscious decisions around how to integrate it.
There's a phrase circulating in developer communities that has stuck with me: "When AI does in seconds what took you years to learn, it hurts." That sentence follows me. Not because it's a complaint, but because it's evidence. It hurts because something is being lost.
What hurts isn't that the machine is faster. What hurts is that the value of what you learned — the hours of frustration, the mistakes, the late nights staring at a problem you couldn't solve, the satisfaction of finally figuring it out on your own — is being socially devalued. The expectation is no longer that you learn. The expectation is that you use a tool that does the work for you.
As a sociologist, I don't care much about technology for its own sake. I care about what technology does to people, to institutions, and to the power relations that organize knowledge. And what artificial intelligence is doing to our relationship with learning has a name in my discipline: it's not novel, it's a pattern.
Why Learning Hurts (and Why That's Good)
The research on how we learn is old, robust, and consistently ignored by those building AI products.
In 1885, Hermann Ebbinghaus demonstrated something every student has experienced: without review, we forget approximately 75% of what we learn within the first 24 hours. The forgetting curve is not a brain defect — it's a feature. Forgetting is the mechanism that forces the brain to decide what deserves to be remembered. For something to stick, it has to cost something.
This "cost" has a name in academic literature: desirable difficulties, a concept developed by psychologists Robert Bjork and Elizabeth Bjork. They discovered that strategies that make learning feel harder in the moment — like mixing topics instead of studying them one at a time — produce dramatically better long-term results. In one classic study, students who practiced with interleaved problems scored 63% on a test one week later, compared to 20% for those who practiced with single-topic blocks.
The discomfort, the "pain" of not understanding something right away, is not a signal that something is wrong. It is a signal that the brain is forming new connections. Cognitive neuroscience research documents that students often interpret intense cognitive effort as a sign of poor learning, when in reality it's the exact opposite.
The problem is that the tech industry has built an ecosystem whose explicit goal is to eliminate that discomfort.
Learning Takes Real Time: The Numbers Don't Lie
The hours required to reach even a basic level of competence in any discipline are hard to ignore:
| Domain | Estimated time to intermediate level |
|---|---|
| Programming | ~300 hours to exit the beginner stage |
| "Easy" languages (French, Spanish) | 600–750 hours (U.S. Foreign Service Institute) |
| Complex languages (Chinese, Arabic) | ~2,200 hours (FSI) |
| Expert skills (music, elite sports) | Thousands of hours of deliberate practice (Ericsson, 1993) |
These numbers are uncomfortable for an industry that promises instant results. Nobody wants to hear that becoming a junior programmer requires 300 hours of conscious practice. It's more profitable to sell the idea that an AI tool can do the work for you — that you already "know" because ChatGPT gave you the answer.
But there is a fundamental difference between getting an answer and knowing how to get an answer. That difference is what's at stake.
Sociology Has Seen This Pattern Before: A History of Technologies That "Facilitate"
Here is the point most often omitted in debates about AI and education: we have been here before. Several times. And sociology has kept the record.
The Printing Press and the Panic About Memory (15th Century)
When Gutenberg democratized the printing press around 1450, Renaissance humanists were alarmed. The scholar Conrad Gessner decried an "overwhelming abundance of books" that was "harmful to the mind." Erasmus of Rotterdam worried that easy access to written knowledge would degrade the quality of thought. Why struggle to memorize if you could simply look up the reference?
The fear was not irrational. The ars memorativa — memory cultivated as an art — genuinely declined. Monks who memorized entire Bibles, rhapsodes who could recite the Iliad in full: those practices became unnecessary. But what emerged in their place was something different and powerful: critical reading, cross-referencing, argument construction from written sources. The printing press did not destroy thought. It transformed what kind of thinking was valuable.
The question is: are we being equally patient and clear-eyed about AI?
The Calculator and the "Death" of Mathematics (1970s–80s)
In the 1970s, when pocket calculators became accessible, the debate in schools worldwide was identical to today's. Teachers argued students would lose the ability to do mental arithmetic. Reformers said that freeing students from mechanical calculation would let them focus on higher-order mathematical reasoning.
Both sides were right.
Subsequent research showed that basic mental arithmetic did decline among students who used calculators without restriction from an early age. But it also showed that when the calculator was integrated with conscious pedagogical design — first understanding the concept, then using the tool to explore more complex problems — mathematical reasoning improved. A 2003 meta-analysis by Ellington examining 54 studies found that students using calculators outperformed non-users in conceptual understanding and problem solving, provided the curriculum was designed to accompany the tool thoughtfully.
The lesson was not "calculators yes" or "calculators no." It was: the context of use determines everything.
Television and Fragmented Attention (1950s–90s)
Neil Postman, in Amusing Ourselves to Death (1985), documented how television was transforming American public discourse. Images replaced arguments. Entertainment colonized politics, education, and religion. Attention fragmented because the medium demanded fragmentation.
Postman was not entirely wrong. But he wasn't entirely right either. Television did not destroy complex thought: it displaced it toward other formats. Documentaries, in-depth debates, investigative journalism on screen all emerged. What it did destroy, as he predicted, was a certain kind of slow, deliberate public discourse.
The sociology of media learned from that episode that communication technologies do not replace cognitive faculties: they redistribute which ones get exercised and which ones atrophy. It's never all-or-nothing. It's a reordering.
GPS and Spatial Memory: The Most Documented Case
Perhaps the cleanest example of what AI is doing to us is a more recent and specific one: GPS and spatial memory.
A 2020 study published in Nature Communications followed GPS users over four years and documented what many drivers already sensed: assisted navigation reduces hippocampal activity and the ability to build internal "cognitive maps." Users who once found their way intuitively began to depend completely on the device. When the GPS failed, they were disoriented in areas they had known for years.
GPS is not bad. The problem is that no user received instructions on how to use it without losing the underlying skill. Nobody designed the integration with long-term autonomy in mind.
Sound familiar?
Cognitive Offloading: How AI Is Making Us Dependent
Cognitive offloading describes the act of delegating mental processes to external tools. It's something humans have always done: writing on paper, using calculators, searching Google. But the current scale and speed of delegation are qualitatively different.
A study on physicians performing colonoscopies found that after working with AI-assisted detection systems, their ability to detect polyps without the AI had significantly declined. The tool had not helped them learn better: it had made them dependent.
Harvard Gazette warned in 2025 that "excessive reliance on AI" could lead to "cognitive atrophy" and reduced critical thinking ability. The pattern repeats in educational contexts: students who use chatbots for academic tasks show worse retention and more superficial analysis after routine use.
From sociology, the concept that best captures this is Pierre Bourdieu's cultural capital. Bourdieu argued that embodied knowledge — skills that become second nature after years of practice — is a form of capital that positions individuals in the social field. AI does not democratize that capital: it creates the illusion of access without genuine incorporation. A student who obtains correct answers without understanding them has not acquired cultural capital. They have acquired a prosthetic.
And prosthetics, when removed, leave a void.
What AI Is Doing to Children and Teenagers
Discussing AI without discussing children, adolescents, and young people is, to me, an unforgivable omission. They are the ones growing up in an ecosystem where instant answers are the norm, where frustration is eliminated before it appears, where there is no room for the discomfort of not knowing.
The Korean-German philosopher Byung-Chul Han describes in The Burnout Society how we are losing the ability to experience negativity — the "no," the obstacle, the limit — which is essential for the formation of a subject. A generation growing up without experiencing resistance from the real world, without having to persist, without failing and trying again, is not being "empowered" by technology: it is being deprived of the basic conditions for developing intellectual autonomy.
The sociologist Harry Braverman documented in Labor and Monopoly Capital (1974) the deskilling process that accompanied industrialization: machines did not only automate tasks, they fragmented artisanal knowledge so that workers no longer understood the complete process. The shoemaker who made an entire shoe was replaced by workers who only stitched one part, without knowing how the whole was constructed. The result was a cheaper, more replaceable, and fundamentally less autonomous workforce.
AI in education has the potential to reproduce exactly that pattern — not with bodies, but with minds.
The Illusion of Knowledge: Dunning-Kruger Amplified by Algorithms
There is a pattern that recurs in testimonials from developers who adopt tools like GitHub Copilot or Cursor. At first, they feel incredibly productive. Code writes itself. Bugs fix themselves. But over time, some begin to notice something unsettling: they no longer understand the code they are producing.
AI doesn't just write code: it also writes the illusion that we understand what we're doing.
The Dunning-Kruger effect describes how people with low competence in an area tend to overestimate their ability. AI introduces a dangerous variant: it's not that we believe we know more than we do — it's that the tool gives us the result before we even have the chance to realize we don't know.
Ivan Illich called this counterproductivity: the moment an institution or technology begins to undermine the very objective that justified its existence. A school that produces students who hate learning is counterproductive. A healthcare system that makes people sick is counterproductive. An educational tool that generates the illusion of knowledge without real knowledge is, in Illich's terms, profoundly counterproductive.
And the most concerning part: the indicators of counterproductivity are not noticeable immediately. They appear years later, when the system is already entrenched and the costs of change are enormous.
The Misuse of AI in Education: A Critique with Evidence
Let's be direct: the AI education industry has a structural incentive problem. Success metrics — time on platform, "sessions completed," active users — do not measure learning. They measure engagement. And engagement is optimized by eliminating friction, not creating it.
The result is predictable. Platforms that promise to "learn to code in 30 days" with AI that completes the student's code before they can make a mistake. Virtual tutors that give the answer before the student has properly formulated the question. Tools that simplify the problem so much it is no longer the original problem.
The political scientist Langdon Winner asked in his classic essay Do Artifacts Have Politics? (1980): are technologies designed with embedded values that benefit some and harm others? The answer, in the case of these platforms, is yes. They are designed to maximize user retention, not cognitive development. This is not malice — it is market rationality. But its consequences are political.
The most ironic part is that AI has precisely the opposite potential. A well-designed system could:
- Generate problems calibrated to each student's exact level, within the "zone of proximal development" Vygotsky identified as the optimal learning space
- Provide hints instead of answers, guiding without resolving
- Identify each student's specific conceptual errors and design exercises to address them
- Simulate real contexts where knowledge must be applied before it is urgently "necessary"
That would be transformative use. What we largely see is something else.
Using AI Better: A Proposal Rooted in History
I am not calling for us to abandon AI. That would be as naive as asking a society to abandon electricity. The history of communication technologies shows that wholesale rejection never works: the printing press survived Erasmus, the calculator survived conservative teachers, the Internet survived 1990s moralists.
What can change is the consciousness with which we integrate these tools.
For educators and trainers: Recover intentional pedagogical design. AI as a "practice partner," not a shortcut. First the individual attempt, then comparison with the AI's answer, then critique of that answer. The process of comparison and critique is where real learning happens.
For developers and professionals: Establish deliberate practice sessions without assistance. Like pilots who practice manual landings despite the autopilot, we need to maintain the "mental muscle" that AI allows us to neglect. Alternating between tool use and manual work is not romantic nostalgia: it is cognitive hygiene.
For families and teenagers: Teach thinking with AI, not replacing thought. Show that AI makes mistakes, hallucinates, has no consciousness. Cultivate a skeptical, active relationship with technology. Researcher danah boyd insists that digital literacy isn't about knowing how to use tools — it's about understanding how tools use us.
For AI product designers: Rethink success metrics. Does the tool help the user learn? Or does it make them dependent? Design for long-term autonomy, not immediate convenience. There are examples of doing this well — adaptive tutoring systems like Sal Khan's Khanmigo — and the design difference is striking.
Conclusion: Pain as a Sign of Intellectual Life
Technology is never neutral. The sociology of technology has documented this since the 1970s: every tool carries an embedded philosophy, a worldview, an idea of what it means to be human.
Current AI, as it is being designed and deployed at scale, carries an embedded idea: that knowledge can be obtained without effort. That difficulty is a bug, not a feature. That efficiency is the supreme value of learning.
The historical and educational evidence says otherwise. The printing press transformed memory but created critical reading. The calculator atrophied mental arithmetic but potentially freed higher-order reasoning — when used thoughtfully. Television fragmented attention but also created new forms of visual narrative.
In each case, the outcome depended on conscious integration decisions. In each case, there was a window — sometimes short — to make those decisions before the technology became naturalized and its effects became difficult to reverse.
We are in that window now, with AI.
Learning hurts because it requires us to face what we don't know, to make mistakes, to persist, to sleep with unsolved problems and wake up with partial solutions. That process — uncomfortable, slow, frustrating — is not a bug in learning. It is learning itself.
Future generations deserve better than artificial intelligence that saves them the effort of thinking. They deserve tools that help them think better — not tools that think for them. That is not nostalgia. It is a social design decision. And society — not just the industry — has the right to make it.
Maren Schulz — Sociologist of technology. Technology is never neutral. 🔬
