Are Your Graduates Learning to Reason With AI, or to Copy From It?
The Nature editorial that reframes the higher-education AI debate, and what it means for the students arriving in South African workplaces from 2027.
Sign up for my Substack daily AI newsletter here.
See my AI Training course portfolio for corporate Business Leaders here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
Ask the question of your last three graduate hires and you may not like the answer. The students who will enter workplaces from 2027 are being taught to use artificial intelligence in four different ways across our universities, none of them yet standardised, and the corporate onboarding programmes receiving them are largely assuming that some form of critical engagement with the machine has been built in. Nature has just published an editorial that argues plainly it has not been, and that the fix is a discipline most universities are not yet teaching. The question of whether your graduates are learning to reason with AI or simply to copy from it is now a workforce readiness question, not a philosophical one.
CONTEXT AND BACKGROUND
The piece appeared as a Nature World View on 9 September 2026, and it is unusually direct about which framing the journal considers wrong. Good intellectual sparring partners, the editorial argues, are scarce. To have someone who will challenge your reasoning, hold you to your evidence and push back on your arguments in real time, you have historically needed a strong teacher, a sharp colleague or the right seminar audience. Most people never get the opportunity. Language models offer, for the first time, broad access to a capable counterpart that will reason with you on demand. The dominant story runs the other way, casting AI as a threat to human thought. That worry, Nature says, is aimed at the wrong object. The danger is not AI itself. It is passive use.
The evidence supporting that reframe is stronger than most South African readers will realise. A peer-reviewed field experiment published in the Proceedings of the National Academy of Sciences followed nearly one thousand high-school mathematics students across two conditions. Students given an unrestricted ChatGPT-style tutor performed better while they had access to it, and seventeen per cent worse than the control group once it was removed. Students given a tutor designed to withhold direct answers and instead offer teacher-crafted hints showed no such penalty.
A separate study, presented at the Human Factors in Computing Systems conference, surveyed three hundred and nineteen knowledge workers and found that higher trust in AI was associated with less critical thinking, while higher self-confidence was associated with more. The Bastani study establishes the effect in causal terms. The Lee study establishes the mechanism in behavioural terms. Nature has now written the editorial that draws the conclusion. The Russell Group of twenty-four research-intensive United Kingdom universities has moved in the same direction, adopting collective principles that oblige member institutions to teach AI literacy, adapt assessment to incorporate ethical AI use and prepare staff to develop these capabilities in students.
INSIGHT AND ANALYSIS
There is a specific discipline sitting behind the editorial’s argument, and it is worth naming because South African higher education does not currently teach it in the form the moment requires. Existing foundation-year critical thinking modules train students to reason against static texts, prepared arguments and cited sources. What students now need is the same discipline exercised against a system that produces plausible-sounding answers in real time, adjusts its position when challenged, and never signals its own uncertainty in a way an eighteen-year-old would notice. That is a different reasoning problem, and it is a teachable one. It means treating the machine’s output as a claim requiring warrant rather than as an answer requiring acceptance. It means asking, of any statement the machine produces, on what basis it is true, what evidence would falsify it, what a competent human in the same field would say instead, and what the machine would say if the prompt were framed differently.
The reframe in the Nature editorial sharpens the argument. A university that cannot teach the discipline of active engagement with a machine that will produce plausible answers to almost any question a student can ask it is not preparing that student for the working world they will enter. A student who accepts a language model’s output uncritically is not making a decision at all. The machine is. The student is a channel through which the machine’s output reaches the assessment. If we grade that student on the output, we are grading the machine. If we want to grade the student, we have to assess something the machine cannot do on the student’s behalf, and the honest form of that assessment is the student’s engagement with the machine, not their avoidance of it.
In practice, this means grading the record of the reasoning rather than the finished document. Requiring students to submit their prompts alongside their answers. Requiring them to show, in their own writing, where the model was wrong or biased, what evidence they used to establish that, and how their own knowledge of the subject pushed the model to a more defensible answer. It is a heavier assessment discipline than marking a submitted essay, and it is the assessment that actually measures what the student has learned. This is not a hypothetical exercise. Erasmus University Rotterdam has built exactly this into its teaching methodology under the name “The Sparring Partner,” a step-by-step model in which students first draft their own response to a question, then use AI to critique it, then defend or revise their answer against the machine’s pushback. What Nature has published as a principle, Erasmus has published as a lesson plan.
IMPLICATIONS
For South African universities and the Council on Higher Education, this reframes what quality assurance in AI-assisted teaching actually consists of. The current conversation, which I engaged with at the Quality Assurance 5.0 Summit in Durban this week, has been organised around detection, disclosure and integrity. Those things matter. But the Nature editorial is arguing that they are secondary. The primary question is whether the institution is teaching the discipline of active use, and whether it is assessing that discipline directly rather than assessing outputs that the student may or may not have produced themselves. Business schools internationally have started to make this shift explicitly, with the Chartered Association of Business Schools naming the move “from detection to development” and pointing to policies at Oxford and Queen Mary University of London as early examples. Neither the specific curriculum design question nor the specific assessment design question has been solved in South Africa, and both need to be, at pace.
For South African employers and boards, the implication arrives sooner than the university one. The graduates coming into your organisation from 2027 will have had four different experiences of AI in their degree, none of them yet standardised: unrestricted use, restricted use, nominal prohibition, and no policy at all. The corporate onboarding programme should not assume any of them can reason with a machine well. The corporate development budget should include the same active-engagement discipline the universities are being asked to teach. The alternative is a workforce that can produce AI-assisted work but cannot recognise when the work has gone wrong, which is the gap UNESCO’s own recent work on civil-service AI literacy has named at the level of governments. The board question the Nature editorial forces is a workforce readiness question, and it is more urgent than the current AI adoption conversation makes it appear.
CLOSING TAKEAWAY
The Nature editorial’s most important line is not the sparring-partner image, memorable as it is. It is the sentence that follows: the danger is passive use. Naming it that way puts the responsibility somewhere different from where the sector’s current conversation has placed it. It is not in the model. It is in the mode of use, and the mode of use is a teachable discipline. South African universities are not currently teaching it. South African corporates are not currently training for it. If the next twelve months bring a serious attempt to build both, this country will produce a generation of graduates and workers better equipped than most of their international peers to reason with these systems rather than accept their output. If they do not, we will produce a generation who can operate the tool but cannot argue with it, which is the definition of the problem Nature has just named.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments