Schools, universities, vocational training - all of them are built for a world in which certain skills were scarce and expensive. They no longer are. What stays, what becomes newly important, what have we overlooked so far?
At first the question seems absurd. Of course you still have to learn - reading, writing, arithmetic, thinking. But the question isn't absurd. It's real, and it grows more pressing every day in millions of classrooms, lecture halls and training companies.
Pupils writing homework with ChatGPT. Students having term papers generated in two hours. Apprentices photographing maths problems with their phone and reading off the answer. Schools react in two modes: ban it or ignore it. Both are wrong. The third mode - rethinking - has barely begun.
This essay doesn't try to regulate AI in school. It asks the question behind it: what is education actually for - in a world where machines can calculate, write, translate, research and summarise? Whoever doesn't ask this question gets asked it by reality.
The school system as we know it is a child of the 19th century. It arose in a world where knowledge was scarce (books expensive, libraries rare), where producing texts meant physical labour (writing by hand, typing with ten fingers), and where every calculation - wages, taxes, surveying - had to be done in the head or with paper and pencil.
School answered that, and it did it well. It trained generations in what was scarce and expensive: taking in and retaining knowledge, producing texts, working through problems, mastering languages. Whoever was good at these disciplines had an advantage in life, because these disciplines were demanding - mastering them set you apart from others.
Now they no longer are. Knowledge is retrievable in two seconds. A machine produces texts at the push of a button. Wolfram Alpha has solved maths problems for twenty years, ChatGPT for three. Translation was an elite profession - today trivialised. What school trained with great effort has become cheap in one stroke. And the question isn't: "Is that bad?" The question is: What was the real thing these exercises were supposed to train?
The term for it is the "grammar of schooling" - the invisible form made up of timetables, subjects, year groups. In Tinkering Toward Utopia (1995), David Tyack and Larry Cuban showed how astonishingly stable this grammar has remained over more than a century. Reforms come and go, the form stays - shaped by a world in which answers were scarce and expensive.
Why this form worked so well is explained by Daniel Willingham in Why Don't Students Like School? (2009) from a cognitive-psychology perspective. His point: the brain isn't primarily made for thinking but for avoiding thinking. Where answers are scarce, the effort of learning pays off. Where they become cheap, the old incentives go off balance - and with them the old school.
The simplest reaction sounds reasonable and is dangerous. It goes like this: if AI can calculate, pupils no longer need to learn to calculate. If AI can write, you no longer need to learn to write. If AI can translate, foreign languages are superfluous. More time for the important things - creative work, social skills, holistic thinking.
The problem with this logic: it confuses the tool with the purpose. Maths isn't calculation. Writing isn't typing. Learning a foreign language isn't translating. These activities are vehicles - they train something connected to them, but not to the finished product.
Whoever solves an equation themselves isn't training the result. They're training structural thinking: recognising that this problem has a certain form, that a certain operation helps here, that in the end you have to check whether the result makes sense. Whoever writes an essay isn't training the finished document. They're training the clarifying of their own thinking - noticing where a thought doesn't yet hold, where it's contradictory, where you don't actually know yet what you want to say. If you take away the vehicles, the real thing falls away with them.
For mathematics the classic reference is Heinrich Winter. In Mathematikunterricht und Allgemeinbildung (1995) he formulated the three "basic experiences" that have shaped the German discourse ever since: mathematics as a tool of the world, as a school of thinking, as a cultural experience. None of them is "being able to calculate". Calculation was always the vehicle - the purpose lay behind it.
For writing, research has drawn the same arc. In The Psychology of Written Composition (1987), Carl Bereiter and Marlene Scardamalia distinguish between knowledge telling - simply writing down what you already know - and knowledge transforming, writing as a cognitive process that changes understanding. If AI takes over only the first mode, that's not bad. If it also replaces the second, it becomes dangerous.
When the activities of the old school become cheap, what it cannot replace becomes expensive. Six abilities whose value will rise in the coming years - and which today barely appear in most curricula.
Judgement. An AI delivers a fluent, plausible answer. Is it correct? Is it complete? Did it leave out important aspects? Whoever can judge that gains. Whoever can't gets deceived without noticing.
Being able to ask questions. A better question leads to a better answer. That was always true - with AI it's merciless. Whoever only asks vague questions gets vague answers back. The art of formulating a question sharply becomes a core competence.
Tolerating complexity. When AI handles the easy 80 percent of a problem, the remaining 20 percent aren't easier. They're the truly difficult ones - exactly those where several solutions are possible, where judgement is needed, where patience is required.
Scepticism. AI always sounds convincing. Even when it talks nonsense. Whoever hasn't learned to read with healthy distrust loses their bearings. And in the classroom we see it daily: pupils trust the chatbot more than the teacher.
Relationship and trust. What a machine can't do because it isn't a person. Withstanding conflict, listening to another view, being able to rely on someone - all of that becomes rarer and more valuable the more our communication is mediated by machines.
Taste. When two answers are both correct - which is the better one? That's the cultural question AI fundamentally cannot answer. It can deliver the average. The outstanding taste a human has to develop.
Why AI answers seem so seductive is explained by Daniel Kahneman in Thinking, Fast and Slow (2011). His central model: System 1 - fast, intuitive thinking - takes linguistic fluency as a sign of truth. AI systematically exploits exactly this confusion, not on purpose, but because it naturally phrases things fluently - even when it's wrong. Judgement means: deliberately activating the second, slow level of thinking.
How widespread these errors actually are in top models is documented by the AI Index Report (Stanford HAI, annual) - one of the best sources for showing that AI output isn't randomly faulty but systematically so. In 2024 UNESCO presented, with the AI Competency Framework for Students, a framework in which judgement, ethics and critical evaluation stand not as an add-on but as the core of learning.
Here comes an insight that still finds too little room in the education debate: to use AI well, you need exactly what AI can't teach you. You need your own practice in what the machine does.
Whoever has never written an essay themselves can't judge whether the AI text is good. Whoever has never solved an equation themselves doesn't see that the AI solution has a sign error. Whoever has never mastered a language from the inside doesn't notice that the AI translation misses the tone. The competence to judge presupposes the ability that's supposedly no longer needed.
From this follows a recommendation that at first seems contradictory: we need more practice, not less. But more targeted, with a different aim. No longer to master the finished product - but to develop the judgement that enables you to check a machine's product.
That changes everything about teaching. Tasks a pupil can delegate to the phone no longer train anything. Tasks where they have to take an AI result, check it, improve it, situate it - those train exactly what counts. The new homework isn't "calculate", it's "check what the AI calculated".
The cognitive-psychology basis is provided by John Sweller with Cognitive Load Theory (from 1988). His point: complex problems can only be solved if schemas exist in long-term memory that working memory can draw on. These schemas arise through practice. Tools don't replace them - they can only use them more efficiently when they're there.
In Visible Learning (2009), John Hattie analysed over 800 meta-studies on learning effects. Direct instruction and targeted practice with feedback are among the most effective educational measures of all. Daniel Willingham adds the decisive point in Critical Thinking: Why Is It So Hard to Teach? (2007): critical thinking isn't a general skill that could be trained independently of domain - it presupposes knowledge of the topic. Whoever wants to judge AI output has to understand something of the topic themselves.
One of the most common arguments in the debate is: "Let's tackle this later, when the pupils are older." That's a mistake for two reasons.
First: the pupils have long been using AI. In the fifth and sixth grades, where many teachers still believe this is an upper-school topic, ChatGPT accounts are long active - sometimes from an older sibling, sometimes secretly, sometimes quite openly. The question isn't whether children come into contact with AI, but whether they do so with or without guidance.
Second: whoever first works with AI as an adult handles it very differently from someone who learned at twelve what hallucination means and why every AI output needs checking. It's about a basic competence - comparable to reading or writing. Whoever develops it early has it. Whoever picks it up late stays unsure.
Concretely this doesn't mean "all primary pupils get a ChatGPT account". It means: already in primary school it has to be addressed that there are machines that sound seemingly clever but make mistakes. In middle school they have to practise working sensibly with them - and checking their answers. In upper school they have to reflect on what they mean - for science, work, society.
Whoever wants to know how widespread AI actually is among young people should look at the JIM study by the Medienpädagogischer Forschungsverbund Südwest - collected annually and representatively for 12- to 19-year-olds in Germany. The KIM study by the same institute does the same for 6- to 13-year-olds. Whoever reads these figures can no longer seriously claim AI is an upper-school topic.
In the AI Competency Framework for Students (2024), UNESCO explicitly recommends AI education from primary school on, staged by age group. By comparison, the German KMK strategy paper "Bildung in der digitalen Welt" (2016, updated 2021) is visibly overtaken by the AI wave - written in a world where ChatGPT didn't yet exist.
Whoever thinks AI in education is a question for teachers and education politicians misjudges its scope. Today's pupils are tomorrow's employees, applicants, customers and entrepreneurs. But the question isn't acute only in ten years. It's acute today.
We see it in every business we work with. Employees who use AI without judgement are a danger - they copy texts that aren't right, take over figures that are wrong, rely on sources that aren't sources. They're exactly the same deficits we see in schools, only with consequences that can cost a business customers, money or reputation.
From this it follows: education doesn't end with school. Any business that today lets employees work with AI without investing in their judgement builds on sand. There are plenty of trainings for operating AI - what's missing are trainings on the question of how to check, situate and improve AI output. That isn't a technical question. It's an education question.
When we at wendwerk build software for businesses, we always think in this layer. Tools that an employee without judgement can't put at risk. Processes with built-in checkpoints. AI functions that show their uncertainty instead of hiding it. Good tools can't replace missing education - but they can make missing education less dangerous.
One of the few books that systematically thinks the education question and the economic question together is Artificial Intelligence in Education by Wayne Holmes, Maya Bialik and Charles Fadel (2019). Their core thesis: what is laid down - or missed - in school continues at the workplace, in both directions. School and business are not separate worlds on this question.
How large the competence gap among employees in Germany has by now become is documented by the ongoing Bitkom studies on AI in companies - AI use is rising faster than the ability to deploy it sensibly. The IAB (Institute for Employment Research) investigates in parallel which activities are being technologically substituted - and which aren't. The takeaway: exactly the abilities we named above - judgement, scepticism, taste - are among the non-replaceable ones.
It isn't about "digitalising" school. It's about asking anew what education is actually for - now that the old school promise no longer holds.
The old promise was: we'll teach you what the world will demand of you, and then you can make your way. That promise is broken. The world will demand something different tomorrow than today, and no curriculum can guarantee any longer what will still be scarce and valuable in ten years. What remains isn't a canon of knowledge but a stance - staying able to learn, being able to check, not believing everything, asking your own questions, developing your own taste, keeping relationships.
This stance has to be practised. It doesn't arise from click tutorials. It arises in the classroom, at the desk, in real tasks at which you fail and reorganise yourself. AI makes practice neither superfluous nor easier. It only changes the aim of the practice. Whoever has this aim clear gains - as a teacher, as a parent, as a business, as a society. Whoever doesn't have it clear practises in a void.
Who is currently shaping the global debate about AI you can read under Where we're heading. Why many don't get started despite their interest, in on AI paralysis. The factual basics on AI you'll find on wendwerk.de/en/wissen.
Curated by Johannes Hohls for wendwerk.