If AI can generate work that matches, and in some cases exceeds, expert-level output in seconds, do university students really need to acquire foundational knowledge any more?

This question is being asked with increasing urgency and sits behind debates about essays, exams, coding assignments, legal research, medical training, business analysis and creative work. But it is very easy to answer.

The rise of AI straightforwardly makes foundational knowledge more, not less, important.

The student who benefits most from AI is one who can not only get AI to produce an answer but judge whether the answer is any good. The first part is easy, the second is anything but. For example, AI can draft a legal memo but a law student needs to have the judgement to know whether the argument would survive court. It can summarise clinical history but a medical student needs to be able to spot a dangerous omission and check against patient information.

É«ºÐÖ±²¥

AI makes shallow domain knowledge more dangerous than ever, because plausible answers can now be produced faster than they can be checked. Earlier this year, that courts have already sanctioned lawyers for filings containing AI-generated false citations. In medicine, showed that access to an LLM does not automatically improve physicians’ diagnostic reasoning. In fact, as both and recently reported, early evidence from a range of studies suggests that passive, answer-seeking AI use can improve immediate productivity while reducing cognitive engagement, persistence, motivation, verification or later unaided performance.

OpenAI’s own research on argues that standard training and evaluation systems often reward AI for guessing over acknowledging uncertainty. In knowledge work, the human task therefore shifts toward verification, integration and stewardship of the answer. All this requires acts of judgement.

É«ºÐÖ±²¥

But as the have long shown, people do not evaluate new information from nowhere. They interpret it through prior knowledge, organised concepts and their ability to connect ideas across contexts. That is, they use their deep domain knowledge, combined with lived experiences.

The notion that students simply need to learn how to question and prompt AI, then, is problematic. Teaching students how to ask good questions and write good prompts is useful but it can go only so far because asking good questions also depends on what you know – and so is judging the replies.

Students certainly need vertical foundations in their fields: law students need law, medical students need medicine, engineers need engineering. But they also need horizontal foundations – a shared core grammar – across and supporting the vertical domains. To use AI well, students must ask not only whether an answer is acceptable within a field but whether the evidence is reliable, whether uncertainty has been represented honestly, whether the reasoning follows, whether assumptions are hidden, whether numbers are being abused, and whether the proposal is justified.

Among the elements of this shared core grammar, two deserve particular attention: mathematics and philosophy.

Mathematics instils mind habits such as precision, structure, abstraction, proof, proportionality, probabilistic reasoning and comfort with uncertainty, among others. These habits are now essential because AI systems often speak in the grammar of certainty while operating in the machinery of probability.

Nor is the need to learn mathematics any more urgent for STEM students than it is for anyone else. The humanities and social sciences now sit inside a world saturated with numbers: polling data, demographic change, algorithmic recommendation, economic inequality, climate models, audience analytics, social networks, digital archives and machine-generated text. Graduates in these disciplines need enough mathematics to resist being intimidated by numbers and to recognise when those numbers are not being used appropriately – including by AI models. In the AI era, mathematical illiteracy becomes a practical vulnerability.

É«ºÐÖ±²¥

Likewise, philosophy is not a luxury for people who enjoy unanswerable questions. It is disciplined thinking about meaning, knowledge, value and action. It asks what counts as knowledge and evidence, what makes an argument valid, what justice requires, how freedom should be understood, what obligations we have to others, what a person is, and what kind of life is worth building. Many of the hardest debates created by AI are philosophical debates, including those on bias, privacy, authorship, automation, surveillance, academic integrity, human dignity and where responsibility for a decision lies when it is distributed across people and machines.

Whereas engineers, computer scientists, data scientists and biomedical researchers are often trained to ask whether something can be built, automated, optimised or scaled, philosophy forces the prior questions: Should it be built? What is the purpose? Who bears the risk? What assumptions are hidden in the metric? What kind of error is acceptable? What does fairness mean in this context?  AI makes philosophical questions practical and urgent.

É«ºÐÖ±²¥

Without philosophy, students might learn to use AI efficiently without asking whether they are using it well. They might optimise for speed while losing sight of truth. They might outsource writing and discover too late that they have outsourced thinking. They might treat ethics as a compliance checklist rather than a habit of moral attention.

This doesn’t mean that every student should become a mathematician and a philosopher. It does mean that every student should receive serious training in mathematical thinking and philosophical reasoning as the core grammar of preparing for life in the AI era. Calculus might not be directly used by every graduate but the habit of abstraction will be. Epistemology might not appear on a job description but the ability to distinguish knowledge from assertion is more important than ever.

Indeed, the labour market is beginning to recognise this. Witness Anthropic co-founder and literature graduate Daniela Amodei’s argument that the in an AI world, as well as ’s and magazine’s recent reports on how big tech is increasingly hiring many philosophers.

Employers will have plenty of applicants who can generate slides, write prompts and automate routine tasks. The scarcer graduate will be the one who combines technical adaptability with judgement: the one who can define the problem, interrogate the output, understand the domain, weigh trade-offs, communicate clearly and take responsibility for a decision.

In other words, employers will need those who can reason – not only with AI but, more importantly, against it.

is professor of learning sciences and higher education at ETH Zurich and director of the Singapore-ETH Centre in Singapore.

É«ºÐÖ±²¥