AI Did Not Break University Assessments — It Broke the Illusion That They Were Working

The fluency fallacy was present long before ChatGPT. AI simply made it impossible to pretend otherwise — and the institutions responding with detection software are protecting the wrong thing.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/ai-did-not-break-university-assessments?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

When South African universities began reacting to AI with bans, detection software, and academic misconduct inquiries, they were responding to a real concern with the wrong instrument. The concern — that students were submitting work that was not their own — was legitimate. The instrument — a technological arms race between AI generation and AI detection — was misdiagnosed. It treated a symptom while leaving the underlying condition unexamined.

That underlying condition has a name. Writing in Daily Maverick, Professor Fulufhelo Nemavhola, Deputy Vice-Chancellor for Research, Innovation and Engagement at Durban University of Technology, calls it the fluency fallacy: the dangerous belief that coherent academic language is the same as understanding. A student may submit a beautifully written essay and still not understand the subject. A report may be grammatically elegant and intellectually empty. A policy analysis may sound sophisticated while evading the difficult work of judgment. AI did not create that gap between surface and substance. It made the gap undeniable — because AI can produce polished academic prose in seconds, making visible what was always true: that the assessment was measuring the surface and calling it the substance.

The institutions that responded to AI by protecting the surface are making the wrong investment. The institutions that understood what AI was exposing are making the right ones.

CONTEXT AND BACKGROUND

The fluency fallacy did not arrive with ChatGPT. It has been present in South African higher education, and in universities globally, for as long as assessment has been dominated by written submission formats — essays, reports, literature reviews, examination answers — whose primary signal of quality is the appearance of scholarly competence rather than evidence of actual understanding.

Nemavhola describes what this produces as assessment theatre: assessments that create the appearance of rigour through long assignments, formal rubrics, referencing rules, and polished submissions, while failing to test whether students can reason, apply, defend, create, or solve. The theatre performed its function adequately as long as the cost of performing it was high enough to serve as a reasonable proxy for the learning it was supposed to certify. A student who could produce a sophisticated essay in an examination hall, without AI assistance and without access to external sources, had at least demonstrated something about their knowledge of the subject. The certification was imperfect, but it was correlated enough with learning to serve its social purpose.

AI has broken that correlation entirely. A student who submits a sophisticated essay written by a large language model has demonstrated only that they know how to use a tool. The assessment can no longer distinguish between the two performances. And the response that treats this as a problem of academic dishonesty rather than a problem of assessment design is the response that preserves theatre at the expense of substance.

The employment evidence supports this diagnosis. I have previously written about the Federal Reserve Bank of Cleveland’s November 2025 finding that young university graduates are now finding work more slowly than their high school counterparts — a reversal not seen since the late 1970s. That finding is consistent with a decade of South African employer feedback describing a graduate capability gap — graduates who arrive with degrees and without the ability to make decisions under uncertainty, apply knowledge to unfamiliar problems, or take professional responsibility for their own judgment. The fluency fallacy is the mechanism that produces that gap. Assessment theatre is the system that rewards it.

INSIGHT AND ANALYSIS

Nemavhola’s distinction between submission-based and judgement-rich assessment is the most useful framework available for thinking about what universities should do in response to AI — and it extends well beyond the South African context into a question that every institution producing graduates needs to answer.

Submission-based assessment asks: can the student produce something that looks like the expected output? Judgement-rich assessment asks: can the student explain, defend, justify, critique, adapt, and apply what the output represents?

The first can be performed by AI. The second cannot — at least not in the specific, personalised, interactive context of a student being asked to defend their own submitted work in real time with someone who knows the subject.

The difference matters because judgment is not a soft skill. It is the capability that determines whether a graduate can function as a professional under real-world conditions — whether they can interpret ambiguous data, make a recommendation under uncertainty, identify the limits of their own knowledge, and take responsibility for the decision they have reached. These are precisely the capabilities that South Africa’s development challenges require, and precisely the ones that assessment theatre does not test.

Nemavhola is direct about the national stakes: a country confronting unemployment, infrastructure decay, water insecurity, energy instability, and weak public services cannot afford universities that certify fluent writing without testing real capability. Every weak assessment, he argues, eventually appears somewhere in society — in a weak bridge, a weak classroom, a weak clinic, a weak municipality, or a weak public institution. The assessment audit he proposes — asking of every assessment whether it tests capability or merely fluency — is not an academic exercise. It is a national capability investment.

I have previously written about the specific injustice that AI detection software inflicts on South African students who write in their third or fourth language — flagged at higher rates than native English speakers, penalised for the same characteristic writing patterns that made AI detection unreliable in the first place. The universities that invested in detection technology compounded the fluency fallacy rather than addressing it — doubling down on a surface measure while adding a layer of algorithmic bias to the injustice the surface measure was already producing. Judgement-rich assessment is not only more honest than submission-based assessment. In the South African multilingual context, it is also more equitable.

IMPLICATIONS

The practical implications of Nemavhola’s argument are specific enough to act on, and they extend beyond universities into the organisations that employ graduates and the boards and executives responsible for those organisations’ capability.

For universities, the prescription is clear. Every South African institution should conduct an assessment audit within the next academic year, asking one question about every major assessment: does this test capability or merely fluency? That audit should produce specific changes — process notes requiring students to explain how their work was developed, oral defence or viva-style verification for high-stakes submissions, capstone projects using real-world problems and local data, rubrics that reward judgement and application rather than structure and language alone.

Nemavhola’s most forward-looking proposal is the one that will be most contested: AI should be brought into assessment honestly, not smuggled in secretly. Students should be required to declare how they used AI, what outputs they accepted, what they rejected, what they verified, and how human judgment shaped the final product. The future graduate is not someone who never uses AI. It is someone who uses AI critically, ethically, and intelligently — who knows when AI is useful, when it fabricates, when it reproduces bias, and when human judgment must take over. That is the graduate that employers and South Africa’s development challenges actually require, and it is a graduate that can only be produced by assessments designed to test the judgment that sits above and around the AI tool.

For boards and executives, the implication is less obvious but no less important. The graduates entering South African organisations are the product of assessment systems that Nemavhola has characterised as largely testing fluency rather than capability. The professional development, mentorship, and performance management frameworks that organisations apply to those graduates are therefore compensating, at commercial cost, for the capability gap that higher education has been producing. The organisations that engage with this argument — that participate in defining what graduate capability should look like, that design graduate development programmes around judgement rather than credentials, and that tell universities what they actually need from graduates rather than accepting whatever assessment theatre produces — will build more capable workforces than those that continue to treat the degree as a reliable signal of professional readiness.

CLOSING TAKEAWAY

The AI moment in higher education is a mirror, not a crisis. What it reflects is a system that has been measuring the appearance of learning rather than learning itself, rewarding the performance of understanding rather than understanding, and certifying fluency without the capability that South Africa’s employers, public institutions, and development challenges require.

The institutions that respond by improving their detection software are protecting yesterday’s illusion. The institutions that respond by rebuilding assessment around judgement, application, defence, and real-world problem-solving are preparing graduates for tomorrow’s reality. The distinction is not only academic. It is national. And the urgency is not only because AI has arrived. It is because the gap between what South African universities certify and what South African society needs has been growing for longer than AI has existed.

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