Scroll through Singapore discussions online in Nov 2025 (I started writing this around late 2025) and you will find the same anxious thread repeated in different forms. AI is destroying education. Students are using generative AI to write essays they never thought through. AI gives answers to homework questions without the struggle of figuring things out. This is not some imaginary concern, CNA’s May 2025 survey found that all 500 Singapore students surveyed had used AI for homework, with 84% of secondary-school respondents using it at least weekly. We hear teachers complain about submissions that feel like they came from another planet. Not just from another country, but almost as-if written assignments were from another cognitive species entirely.
I think about what it must feel like to be a Singapore teacher grading papers right now. You assign an essay on the economic history of Southeast Asia and get back something that reads like a McKinsey brief. Polished, thorough, with all the looks of an actual report, but completely soulless. And you have to figure out which parts the student actually wrote and which parts they just copied from an LLM prompt. Guess what? You cannot. Not reliably. And nobody wants to be an academic detective.
So the natural response is to ban the tool (or attempt to). Some Singapore educators have been trying that for specific assignments. The students disputed aspects of the findings. But this was not a university-wide ban: NTU itself clarified that students are generally allowed to use Gen AI, subject to declaration and academic-integrity requirements. I understand restrictions when the point is to test independent work. What concerns me is turning restrictions into a substitute for teaching students how to use these machines (I consider LLMs and agents to be only a form of software, contrary to some beliefs others may purport).
I understand the impulse, but banning AI from education echoes the arguments about banning calculators from math classrooms in the 1970s. People actually did this. They then changed the way that they educated to match the technology. The goal could move beyond computation to understanding what to do with the result. That is the shift we should be making now with LLMs in education, not shutting the doors and pretending the machines outside do not exist.
Wolfram Alpha got a similar treatment. It could solve complex mathematical problems and show the steps leading to solutions. The Chronicle of Higher Education reported in 2009 on professors debating whether students should be allowed to use it. And WIRED described a math teacher in 2017 discovering that students could obtain not just correct answers but apparently complete workings from Wolfram Alpha. That changes what a submitted solution tells you about the student. If your math exam can be done entirely by Wolfram Alpha, perhaps your exam was testing calculation procedures rather than mathematical thinking. Isn’t mathematical thinking what actually matters?
AI is the same pattern again, just faster and broader. And the speed matters here. A calculator sped up arithmetic. Wolfram Alpha sped up symbolic computation. An LLM speeds up the entire act of producing written work. That feels threatening in a way that the older tools did not, because writing is the thing we actually grade. We tend to grade output, not the thinking. And now the machine can produce the output for you.
That is the real issue. We built an education system that often evaluates students on what they can produce, not what they actually understand. And the machine can produce things better than many students can, at least on some kinds of assessments. In one University of Reading experiment reported by The Guardian in June 2024, AI-generated answers largely escaped detection and scored above the student average. So we can either change how we evaluate students or spend the rest of our teaching careers policing submissions. I think the first option is the right one.
But here is what I actually worry about. The students who could be left behind are the ones who never learn how to use these tools properly. Right now we have a generation of Singapore students growing up with LLMs in their pockets. Claude can draft an essay and Perplexity can produce a research summary, making the temptation to copy and submit a daily one. And if schools simply ban the tools instead of teaching students how to use them, who will guide the students in the right way to use these tools? And also, research discussed by KQED in May 2025 found that ChatGPT-assisted students improved their written work more, without a statistically significant advantage in knowledge gain or transfer. That cannot be the strategy either. We need to teach responsible use without letting the machine do all the thinking.
What we need instead is to teach students the things that LLMs cannot reliably do for them. People should realize this, the machine does not necessarily know if the answer it gives you is correct. It generates text that sounds correct, but not necessarily correct. Those are very different things. Nature reported in September 2025 that language models can confidently invent academic citations. So you need to teach students how to fact-check. How to spot a hallucination. How to cross-reference claims with actual sources. The risk of a plausible-sounding error is real. The output looks right until you actually know the material. And that is exactly why students need to know the material in the first place.
I think prompting is actually an interpersonal skill. You have to communicate clearly enough for the machine to understand what you want, which means understanding your own problem well enough to explain it. You need to think about your own thinking. CNA reported in February 2024 on teachers at Manjusri Secondary School using structured ChatGPT prompts to encourage students to question and revise their own writing. A student who can write a good prompt may be articulating the gap between what they know and what they need to know. That is not a trivial skill. That is actually quite hard, and worth teaching.
And the fundamentals of work are changing whether schools keep up or not. AI is already being integrated into everyday software. Over the next five years, I expect even more of the tools you use to have AI baked into them. Your word processor already suggests how to complete your sentence. Your spreadsheet already helps to analyse patterns. Your project management tool may already predict deadlines using historical data. If you graduate from a school that simply forbids these tools and never teaches you how to work with them, you will be at a real disadvantage.
Since I started writing this in late 2025, NTU has made its ambitions even clearer. CNA reported on 6 April 2026 that NTU planned to give all undergraduates access to premium Google AI tools from August 2026 and embed AI in 40% of courses by 2030, up from about 5% at the time. That is an ambitious plan, not a claim that it has already transformed every course. Then, in August 2026, CNA reported NTU’s decision to stop using its AI detector from 2027 because it considered the technology fundamentally unreliable. NTU instead emphasised assessments that reveal students’ learning processes. I think this is exactly the right call. When a university starts building AI into how students learn, instead of relying on surveillance alone, it has a chance to produce students who can use the tool like professionals. That is a very different outcome from simply banning the same thing.
NUS has also been moving in that direction. In January 2026, its Senior Vice Provost for undergraduate education described redesigning assessments so students could not succeed simply by outsourcing everything to AI. Then CNA reported in August 2026 that more than 7,700 incoming NUS students had begun a compulsory Applied Generative AI: From Prompting to Evaluation course in July. They were required to finish its online lessons, quizzes and workshop within their first two weeks of the semester. NUS also announced ChatGPT Edu access for all students, faculty and staff, with more advanced tools initially piloted in selected courses. I think that combination matters. Students get practical experience using the software, while assessments still have to establish what they understand. Learning to question an answer and take responsibility for the work you submit is a skill that should outlast whichever AI tool happens to be popular this year.
The shift for students is simple to describe and hard to execute. You need to move from only learning how to do something to asking what should be done and how to verify it was done well. Before AI, a major aspect was always the how. How do I structure this essay? How do I solve this equation? How do I write this report? Now the machine handles much of the how. Your job becomes what and whether. What should I even be asking? And how do I check that the answer I got is actually right?
That is a fundamentally different intellectual position. It requires more judgment and less production. You are directing the work, not grinding through it. That feels lighter in some ways and heavier in others. Lighter because the machine does the heavy lifting of formatting and structure. Heavier because you have to carry the responsibility of deciding what is right. But its akin to delegating the work to someone else, it’s not easy to delegate properly if you are not a good manager. And now, you are expected to be a good manager (of AI) from day 1 of graduation.
When you think about it, the machines are not the problem. Our refusal to adapt our teaching, and learning is. So here is a question readers who are fresh graduates should think about: if you graduated from school this year and everyone in your batch could use every AI tool available, what do you bring to the table that the machine is not able to do for you, and differentiates you from your peers?