JUNE 1 — The breakneck pace at which large language models (LLMs) colonise our higher education system is a major reason many educators struggle to conceptualise their place within teaching and learning. So far, responses have been a mixed bag, perhaps even polarising. On one end, there is a Luddite-like rejection where some educators express a sense of personal offense when detecting some level of LLM use by students. On the other, many are simply throwing their hands in the air before ostensibly AI-generated student output. To be fair, academics are already beaten down by the daily realities of the Higher Education sector’s bloated bureaucracy, trend chasing, and, dare I say, feudalistic work environment.
So how should we grapple with the unmistakable cognitive erosion educators now observe in their classrooms? Most students are not using LLMs in the idealised scenario where the technology augments their thinking process. Instead, LLMs increasingly function as a de facto mental shortcut, trapping students in a feedback loop of cognitive offloading. The bypassing of mental effort weakens students’ ability to exercise their critical thinking muscle, precisely what university is supposed to cultivate. In time, the experience of ‘sitting with the pain’ of thinking through a concept, and building confidence to vulnerably articulate flawed ideas may become foreign within higher education itself.
Our most critical problem is the sluggish institutional response. Irresponsible LLM use is ultimately a form of cheating. James Lang’s Cheating Lessons is useful here because it reminds us that cheating is not merely a question of moral failure. The architecture of the learning environment itself shapes the conditions under which students cheat. We have been here before, in a different form, with the internet, which was also a massive enabler of plagiarism and copy-paste scholarship. Over time, we redesigned education around the internet instead of pretending it could be kept outside the classroom. The transition was imperfect, but it showed that education can survive technological rupture if it is willing to change its architecture.
Because of this policy lag, parts of our education system still rely on assessments designed for the past, despite knowing full well that the conditions of authorship and originality have radically changed. My armchair position is that the future of assessment lies less in policing AI use and more in redesigning assessments around forms of thinking that cannot be neatly outsourced. Ironically, this ‘turn to the future’ may require a partial return to the ol’ reliables of pen, paper, physical tests, and real-time discussion. Already, we’re seeing this rollback trend happening across educational systems we often regard as more advanced than ours.
At the same time, it is myopic to insist on the purity of educational tasks on the basis of ‘LLM independence’ alone. Instead, assessments should evolve to predicate some level of LLM use, and examine evidence of students’ reasoning development instead. In my own practice, I treated students’ chats with their chosen LLMs as part of the thinking process itself. I later developed a sandboxed LLM I dubbed the “Socratic Teacher”, which, like Socrates, responds only with guiding questions rather than direct answers until the student demonstrates a satisfactory level of understanding. The thing showed potential to work like a nicotine patch for the smoking addict! Students still interacted with an LLM, but one that helped them augment their own thinking instead of doing the thinking for them.
Much of the delusion surrounding LLM use stems from a basic misunderstanding of what these systems actually are. To this end, we cannot ignore the role of critical AI literacy in exposing the ‘magic’. Once the spell is broken, students may begin using LLMs with more care and critical suspicion without confusing fluency for intelligence. If we look at LLMs with eyes unclouded by hype (to borrow from Ghibli’s Miyazaki), we begin to see that these systems merely generate statistically probable text that imitates human language rather than functioning as truth-seeking systems. The phrase “stochastic parrots”, coined by critical AI scholar Emily Bender, captures this well because, much like parrots, LLMs are ultimately repeating and rearranging language patterns without actually understanding what is being said.
This is something I am currently trying to develop into a module, introducing students to topics such as the “ELIZA effect”, where humans tend to attribute intelligence to anything capable of conversational mimicry; how LLM training data reproduce knowledge biases; and how conceptual emptiness can hide behind polished prose. Students should also learn about the politico-historical setting that ushered in these technologies, including the power fantasies and techno-messianic culture surrounding figures such as Sam Altman and Elon Musk. Behind the clean interface sits outsourced content moderation work, often carried out by workers in poorer countries who absorb the psychological cost of cleaning these systems. Behind the promise of digital progress sits the familiar ecological exploitation of the Global South.
There is something to be said here about techno-optimism in Malaysia, especially our chronic tendency to hop onboard techy hype trains. I sorely remember the many buzzwordy assertions surrounding the Fourth Industrial Revolution, followed by claims about blockchain technology, and later the metaverse, revolutionising education. While LLMs, unlike the previously mentioned technologies, are already reshaping education, they similarly enter into the same culture of techno-mysticism where technology is treated as carrying its own inevitability, while our amnesiac response remains the same: adopt first and think later.
* Wan Saefullah is a researcher in climate politics and communication. He holds a PhD in Human Geography from King’s College London.
** This is the personal opinion of the writer or publication and does not necessarily represent the views of Malay Mail.
