Listen to this post
The lost generation
AI lets us skip the slow, clumsy, error-prone work of becoming competent. The bill for that shortcut arrives in the future, when the people who were supposed to replace today's experts never built the ability to judge.
Those of us who still remember their college days or their first job experiences might occasionally think back in horror at how many stupid mistakes they made that were called out by teachers or superiors and probably landed them bad grades or got them into deep trouble. Those of us who can’t remember probably deal with students or new employees fresh out of school who make those embarrassing mistakes. It’s a schlep to deal with them, and managing them sometimes consumes more time than they save us compared to doing things ourselves.
With AI, this misery has an end. College students produce fluent output; job novices deliver polished content – if they are even needed. AI allows the experts to focus on what matters. The intern no longer spends three days assembling a market analysis; the model does it in ninety seconds. The junior legal associate no longer drafts the first version of the contract; the model does, and a partner cleans it up. Everyone moves faster. Everyone is more productive. What is not to like?
Here is what is not to like. Those embarrassing mistakes of students and newbies, the slow first attempts, the time a senior burns correcting them – we have been calling all of it a cost. It was never only a cost. It was the mechanism by which a novice turned into an expert. By stripping out these steps using AI, we keep the output while quietly deleting the apprenticeship. The efficiency is real and immediate. The bill is real and deferred – and it compounds.
Es beginnt in der Schule
Long before anyone reaches a job, the habit forms in school. Searching, structuring, filtering, framing – deciding what a question is actually asking, what counts as a good source, how to arrange an argument so it holds weight – these are not by-products of education. They are the education. And they are precisely the tasks a language model now does on request, much better than the average student.
And it does exactly that work for the average student. When Anthropic reviewed more than half a million student conversations, the dominant uses were not looking up facts, checking them or finding typos. The majority of users were asking AI to create and analyze – the higher-order tasks that are supposed to be the hard-won result of learning rather than something to be outsourced.
The consequence of using AI this way is that students never learn the cognitive skills to solve complex problems. Recent research on this confirms it. A 2025 study of 666 participants found that heavier reliance on AI tools was associated with measurably weaker critical-thinking performance. It is not surprising that handing the thinking to an external system has this impact. The effect was strongest in the youngest cohort, aged 17 to 25, who showed both the highest AI dependence and the lowest critical-thinking scores. A separate study of 580 university students found the same inverse relationship. What the students do here is called cognitive offloading, the delegation of core mental tasks to outside helpers. And this is only the beginning: the cohorts tested had completed only parts of their education using AI – what will happen once those students graduate who have been outsourcing all their writing to AI since high school?
The downside of adult offloading is that people get less sharp. The downside of adolescents growing up delegating their minds to AI is a generation that was never sharp to begin with.
Timothy Cook, M.Ed., Psychology Today, 2026
And teachers, at least in areas where the output is a final paper to be graded, are stuck. The times when you could detect AI output reliably are over, as various studies show. Plagiarism software capitulates when it comes to AI detection. Teachers now have two cohorts in their classrooms: those who use AI and those who don’t. Telling whether a good performance is the result of a brilliant mind or good AI use, probably even supported by savvy parents, is impossible with traditional grading methods.
Recently a friend told me a sobering story about their daughter’s performance in a high school project. She had spent three weeks on that project, researched everything herself, created a substantiated presentation about the subject, all without any help from AI. She landed 14 of 15 points, something to be very proud of. A few of her classmates walked out with a full 15 points, after spending a few hours with AI creating a polished presentation and then another day learning the necessary facts by heart, based on a cheat sheet provided by the same AI. Will she also use AI the next time?
Entry-level employees:
hard decisions for companies
For new job entrants, the job market is already beginning to tighten. In the first quarter of 2026, unemployment among recent US college graduates stood at 5.7 percent, well above the overall rate (New York Fed). A Stanford analysis estimated a 16 percent relative employment decline among 22- to 25-year-olds in the occupations most exposed to AI from 2022 to 2025, while experienced employees in the same occupations had far less to fear. The latest data shows that this gap continues to deepen. This is no surprise: AI does everything a newbie does, and better. The sad truth: with latest-generation models, experienced staff are much better at generating valuable output using Artificial Intelligence compared to handing work over to a junior. I was a partner in leading consulting firms and have been coding since I was fifteen. Today I can write and test code much faster with AI agents. I can produce a meaningful presentation within hours, instead of spending days briefing and reviewing with a junior consultant. And if juniors work on tasks with AI, this doesn’t return the same results. They lack the experience to guide AI and judge what it delivers. So as a senior, the hard question is: why explain your needs to a clumsy newcomer if you can talk to AI directly?
Für Unternehmen ist es auf kurze Sicht gewinnbringend, auf die Einstellung von Berufseinsteigern zu verzichten und stattdessen die erfahrenen Mitarbeiterinnen bei der KI-Nutzung zu unterstützen. Die Produktivität steigt, allerdings mit einem großen Nachteil: die nächste Generation erfahrener Arbeitskräfte fehlt. Diejenigen fehlen, die aus Fehlern und dem Input erfahrener Kollegen gelernt haben, und dank ihrer manchmal schmerzhaften Erfahrungen zu Könnern geworden sind. Und selbst die, die noch eingestellt werden, lernen es nicht, sie lernen nur, wie man KI dazu bringt, Dinge zu tun.
The short-term component is that the learning gap continues. After graduating successfully with AI, job seekers write confident applications and resumes using AI, and once they have – with much luck – landed their first job, they are expected to use AI, openly or behind their managers’ backs, because that is the only way they can compete against the others who do it. It is the same dynamic my friend’s daughter experienced in high school: anyone who isn’t using AI is performing at a lower level, delivering less output, taking more time, but learning much more for the future.
Die verfügbaren Daten sind eindeutig: dieser Druck, KI zu verwenden, verheißt nichts Gutes für die Entwicklung der kognitiven Fähigkeiten und den Aufbau von Berufserfahrung.
Das Bildungswesen braucht dringend eine (teure) Reform
Over the past decades, many university curricula have moved from written and oral exams toward grading papers. It is a way to cope with large student numbers and low student-to-faculty ratios. Reviewing and grading a stack of PDFs is a much more cost-effective way of pushing students through the system. In the past, this involved the risk of unwanted ghostwriting help and plagiarism, taking content from other people’s work, but the magnitude of this was limited and – in the case of plagiarism – controllable with software, for example from market leader Turnitin. The same company now sells an AI detector – and it is not working.
The numbers are sobering. Detectors catch raw AI text well – but only from older models they’ve been trained on. In a peer-reviewed 2026 test of 160 academic papers, three of the four major tools caught essentially none of the fully AI-written submissions (Vrije Universiteit Brussel, 2026), and lightly edited text cuts accuracy by more than half, from 39.5% to 17.4% (Perkins et al., 2024). A 2025 University of Chicago study found only one detector survived disguised text and warned of a “technical arms race” requiring constant re-auditing. Even worse, the errors hit the wrong people: in a Stanford study, detectors falsely flagged 61% of essays by non-native English speakers. By now, universities have done the math. Vanderbilt calculated that roughly 750 wrongful flags per year from Turnitin’s claimed 1% false-positive rate are too many and switched it off; Michigan State, Northwestern, and UT Austin followed.
The only path forward is one backward in time, and multiple universities have already moved that way. One version comes from the University of Sydney, where Danny Liu and Adam Bridgeman designed a “two-lane” assessment model that is now spreading internationally. Lane one drives the grades: secure, supervised in-room assessments. For lane two, open assessments, Sydney’s internal guidance suggests that they explicitly include the guided use of AI tools. The University of Bath adopts the same open/closed structure from 2026-27. The University of Surrey will embed discipline-specific AI teaching in every degree from September 2026 and transform assessment to safeguard genuine knowledge, skills, and independent thinking.
Clay Shirky, vice provost at NYU, argues that universities must shift toward in-class writing, oral examinations and required presence – assessments where knowledge is demonstrated in real time, in front of another human. He calls it what it is: not a loss of rigor, but a return to an older, more relational model of education. This includes Socratic dialogue, the oral defense, the teacher who knows the student’s mind because they have interacted with it.
Unfortunately, returning to a more direct hands-on teaching approach is expensive and will upset the economic model of many universities. It might not be feasible everywhere and might in the end distinguish institutions that create capable people from those who merely certify AI use. “Old-school grading” might become a quality label for higher education.
Die Antwort für Bildungseinrichtungen ist einfach, aber schwer umzusetzen. Die Nutzung von KI muss zum integralen Teil des Lehrplans werden; zugleich müssen die Ergebnisse eine klare Trennung zwischen KI-Output und eigener intellektueller Leistung ermöglichen. Dies zu erreichen wird bedeutet eine enorme Herausforderung für Schulen und Universitäten.
Aber letzten Endes ist es eine gesamtgesellschaftliche Aufgabe für uns alle: sicherzustellen, dass Lernen unsere Kinder dabei unterstützt, ihre eigenen Fähigkeiten zu entfalten.
Was für Möglichkeiten haben Unternehmen?
Unternehmen müssen gerade wichtige Zukunftsentscheidungen treffen. Die Firmen fallen dabei in eine von drei Kategorien. Diejenigen, die sich von KI fernhalten wollen oder sie ohne Mehrwert einsetzen. Diese bleiben bei der Produktivität im Heute stehen. Dann gibt es diejenigen, die KI umsichtig einführen und ihre Mitarbeitenden immer in der Verantwortung halten. Mit diesem Vorgehen erzielen sie gewisse Produktivitätsgewinne dank KI und können diese auch auf Dauer halten. Und dann gibt es die dritte Gruppe von Unternehmen, die erfolgreich auf KI setzen und kurzfristig deutlich an Produktivität gewinnen, dafür aber später teuer bezahlen.
The AI debt curveProfits matter, but the future matters too. Some companies have actively made a decision: IBM, for example, says it will triple its US entry-level hiring in 2026. Its chief human resources officer framed it explicitly as a pipeline investment: if the company stops hiring beginners now, there will be no experienced talent pool in three to five years. During training, a key approach is to separate tasks that merely consume time from activities that create expertise. The key idea is that repetitive “skill deserts” can be automated, focusing juniors on tasks that require framing a problem, defending assumptions, testing results and absorbing correction, even when AI could complete these tasks faster. This requires completely different onboarding and trainee programs that foster a culture that balances these aspects, teaches the use of AI and the skills to judge its outputs.
Firms that automate away the learning stage may improve their immediate margins but find themselves, a decade from now, without anyone who understands how their own AI-driven workflows actually behave.
Georgios Petropoulos, MIT Technology Review, 2026
This is the right choice to make, but it’s a costly one. Only time will tell how this commitment survives an economic downturn or increased competition from companies that rely on senior talent using AI efficiently. These questions have no easy answers, but they will need to be addressed soon to avoid walking into the trap of a short-lived productivity boost that is followed by a long-lasting trough. But ultimately, fostering apprenticeship will have to be viewed like any other investment, for example into R&D. The profits gained without investing are short-lived.
Referenzen und weitere Quellen
Forschung und Daten
Anthropic. (2025, April 8). Anthropic Education Report: How university students use Claude.
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O., & Mariman, R. (2025). Generative AI can harm learning. PNAS, 122(26).
Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab working paper.
Federal Reserve Bank of New York. (2026). The labor market for recent college graduates.
Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), Article 6. DOI: 10.3390/soc15010006.
Jabarian, B., & Imas, A. (2025). Artificial writing and automated detection. Becker Friedman Institute Working Paper No. 2025-116. DOI: 10.2139/ssrn.5407424.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. DOI: 10.1016/j.patter.2023.100779.
Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q. (2024). GenAI detection tools, adversarial techniques and implications for inclusivity in higher education. arXiv preprint arXiv:2403.19148.
Ranganath, S., & Ramesh, A. (2026). StealthRL: Reinforcement learning paraphrase attacks for multi-detector evasion of AI-text detectors. arXiv preprint arXiv:2602.08934.
Shen, J., & Tamkin, A. (2026). How AI impacts skill formation. Anthropic (preprint).
Tian, J., & Zhang, R. (2025). Learners’ AI dependence and critical thinking: The psychological mechanism of cognitive fatigue and the social buffering role of AI literacy. Acta Psychologica, 260, Article 105725. DOI: 10.1016/j.actpsy.2025.105725.
Who wrote this? Evaluating the reliability of AI detection tools in higher education. (2026). International Journal for Educational Integrity. DOI: 10.1007/s40979-026-00226-w.
Universitäten und Policies
Bridgeman, A., & Liu, D. (2024). Frequently asked questions about the two-lane approach to assessment in the age of generative AI. The University of Sydney.
Liu, D., & Bridgeman, A. (2023, July 12). What to do about assessments if we can’t out-design or out-run AI. The University of Sydney.
University of Bath. (2025). The two-lane approach to generative-AI assessment categorisation.
University of Surrey. (2026, April 27). AI to be embedded in every University of Surrey degree.
Vanderbilt University. (2023, August 16). Guidance on AI detection and why Turnitin’s AI detector was disabled.
Kommentare und Berichte zu Unternehmen
Cook, T. (2026, March 17). AI is quietly colonizing how you think. Psychology Today.
McConnon, A. (2026, March 2). The bottom rung returns as AI reshapes entry-level jobs. IBM Think.
McGregor, J. (2026, February 11). Why IBM is “tripling” entry-level hiring as AI reshapes jobs. Charter.
Petropoulos, G. (2026, May 26). It's time to address the looming crisis in entry-level work. MIT Technology Review.
Shirky, C. (2025, August 26). The university’s best weapon against A.I.: The 14th century. The New York Times.