Top News

HEC’s new rules on student AI use won’t work – Prism

There is no way to verify full disclosure of tools.

The Higher Education Commission has taken two big steps on artificial intelligence. It has notified a compulsory three-credit-hour AI course for every undergraduate and postgraduate degree in the country, starting from Fall 2026. It has also circulated a draft policy on the use of generative AI in our universities. Both steps are welcome.

But the rules the draft sets for students rest on two ideas that will not hold. It bans students from submitting AI-generated work as their own, and it asks them to declare which tools they used and why. These are honest instincts. They are also unenforceable. No teacher can verify either one, and the draft admits as much when it rules, correctly, that no student may be punished on the basis of an AI detection report alone.

The problem is not that these rules are wrong. It is that they aim at something we cannot see. We keep arguing about what a student may do at home, when the question we can actually answer is what we choose to grade. Start with one claim, because everything else follows from it. AI multiplies existing ability. It helps whoever already has some, and it does not create ability where none exists.

The strongest case against universities is not that students will cheat. It is that some of what we teach may no longer need teaching. Ask a modern coding assistant to clean a dataset or fit a forecasting model, and it will do the job in minutes. So why should anyone sit through a semester of statistics or machine learning? Half of that is true. Running such a model used to be a skill. Today it is a sentence typed into a chatbot.

But test the claim. On Kaggle, an online platform where thousands of data scientists compete on the same problem, everyone has the same data and the same AI assistants. If the tool were enough, every score would be roughly equal. It is not, and to my knowledge no live competition against strong human teams has yet been won by handing the problem to a model and walking away. These systems do not think. They execute, very well, in a direction a person has chosen. Someone still has to decide which question is worth asking, whether the data can be trusted, and whether an answer that looks excellent is quietly wrong. That is not clever prompting, which anyone learns in an afternoon. It is knowledge of the subject: knowing what the answer should look like before you see it.

So why teach anything a machine can already do? The usual defence is that machines cannot yet do the difficult part. I distrust that argument, because it means every course must justify itself again each time a better model appears.

The firmer ground is the reason we still teach children arithmetic when every phone has a calculator. Working through sums is how number sense is built, and none of that comes from watching someone else do it. Education builds capacity, and capacity is built by effort that cannot be outsourced. This does not mean teaching everything by hand. The test is whether the doing builds something that outlives the task, which is why we dropped log tables without loss and kept arithmetic. A student who writes an essay unaided is not really producing an essay. They are learning to hold an argument together. The essay is disposable. The habit is not. AI has not made our curriculum obsolete. It has changed the value of what students produce, not the reason we teach them.

The compulsory AI course is the largest curricular change our universities have made in years, and almost nothing has been said about what should go into it. Left to drift, it will become a course on prompting: how to phrase requests to a chatbot, plus a tour of the popular tools. That is an orientation session, not a course. It will be stale within a year, and the market will not pay for it, because everyone already has it. OpenAI said in February that nine hundred million people were using ChatGPT every week, and its rivals add many millions more. Whatever a semester might teach about typing requests into that box, the world has already taught itself for free.

What is scarce is the ability to build with these models rather than chat with them. Students should learn what an API is, how to ask a model for structured data instead of prose, and how to run a task across 50,000 records overnight. The basics are enough, because once a student knows what a system should do, the coding assistants will help write it. A graduate who chats with a model handles one document at a time. A graduate who can call an API handles the whole archive. In government, I built such pipelines, and what mattered was knowing the task could be automated at all.

Peter Steinberger, wrote a personal AI assistant for himself, leaning on OpenAI’s coding tool Codex for much of the work, and released it as open source under the name OpenClaw. Within months it was among the fastest-growing projects on GitHub, and by February OpenAI had hired him and agreed to fund the project as an independent foundation. Notice what produced that. Steinberger had a career in software behind him and an idea nobody else had had. The model supplied neither. But he could not have built at that speed without it either, and that is the whole point. Neither alone would have made OpenClaw. Multiplication yields something large only when both numbers are large, and forming one of those numbers is what a university is for.

If AI multiplies ability, the same tool must be treated differently at different stages. A doctoral student or a faculty member has expertise to multiply, and the draft is right to encourage their use of AI in research and teaching. A second semester undergraduate has little to multiply yet. The same tool that speeds up a researcher can quietly replace a student’s learning. This is not a contradiction. It is one principle applied at two stages.

That gives a workable rule. Separate learning from grading. Take home work becomes practice, with AI openly permitted and little weight attached, because its purpose is understanding. Grades move to settings we can see: work done in class, short oral defence of submitted work, viva, presentation. A student who used AI to think through an assignment will answer questions on it easily. A student who used AI to avoid thinking will not. Five minutes of questioning does what no software does reliably, and it accuses nobody.

The obvious objection is workload. Oral defence sounds fine in a class of twelve and impossible in a section of eighty. But the oral is not the main instrument here. The main instrument is supervised written work done in the room, timed and marked as it always has been. Every university in the country already runs examinations at scale. What changes is not the logistics but the questions: not recall of material anyone can now look up in seconds, but questions that only make sense to a student who did the work.

The oral is the second instrument, used where it adds most. On projects, theses and group submissions, put the question to one named member rather than to the group. Elsewhere, announce that any student may be asked to talk through any submission, and then sample, five minutes each on a rotation. This is not a hunt for cheats. It is simply that a student who does not know whether they will be asked has a reason to be able to answer.

published the results in the journal Patterns. On ninety-one TOEFL essays written by non-native English speakers, the detectors wrongly called the work AI-generated 61 per cent of the time, and on 18 of those essays all seven agreed on the false verdict. The same tools judged essays by American schoolchildren correctly about 95pc of the time. The cause was perplexity. These systems penalise plain and predictable vocabulary, which is exactly what careful writing in a second language often looks like.

Some vendors have since revised their models and report better results on that benchmark, so the picture varies by detector and by language. But no such system has been validated on Pakistani student writing, and until one is, a score is not evidence of anything. No student here should be judged on a detector report, and no university should buy one of these systems without first testing it on the writing of its own students.

Invigilation is a different matter. Proctored computer labs and supervised terminals accuse nobody. They set the conditions under which work is produced, which is what an examination hall has always done. For any task, the question is whether the task is the goal or the training. If it is the goal, use the tool. If it is the training, doing it yourself is the whole point. I cannot drive a car and claim the benefit of having run a mile.

One more thing no imported policy will tell us. The industry our graduates enter is at an earlier stage of adoption than the foreign case studies assume, and our class sizes are larger than those the guidance was written for. A policy lifted from a North American business school will not survive contact with these conditions.

I am not arguing for caution. I build these systems and I teach them. I am arguing that the order matters. We have started, as institutions usually do, with what to permit and what to buy. The prior question is what a degree certifies. If it certifies that a graduate can produce a document, we have already lost, because documents are now free. If it certifies that a person has been changed by the work, that they can frame a problem, judge whether an answer is any good, and stand behind it, then our task is simply to grade that, in the room, where it can be seen.

Education first, AI second. The tools will change again before this policy is finalised. What we believe about learning should not.

Source link