‘Syncopathy’ and AI: African HE must choose verification
This article first appeared on University World News
Syncopation is not a missing beat. It is a beat the player declines to strike, deliberately, at the moment the listener is most certain it will arrive. The effect depends entirely on the expectation it disappoints: the ear supplies the stress, the music leans, and what ought to sound like an error sounds instead like style.
A language model is a syncopated instrument. It supplies everything about an answer except the operation that would make it one. The citation has the shape of a citation; the argument has the posture of an argument. What is not struck is the beat on which somebody opens the journal and checks.
Call it ‘syncopathy’. The pathology is not the skipped beat. It is that the music continues so smoothly that nobody reaches for the score.
This matters more for us than for most. What a university certifies is not that a thing was said. It is that somebody checked.
Three skipped beats
At the University of Pretoria, Professor Charne Lavery examined essays from 800 second-year students and found that 70% to 80% were, to some degree, machine-generated. Run through Turnitin, most came back clean. Her summary of the position is the most economical description of the problem I have read: the burden of proof lies with the academic, and there is really no way to discharge it.
The reflex is to make this a story about student character. It is not, and treating it as one has already cost us two years. Students did not acquire a new appetite for dishonesty in November 2022. They discovered that one step in a long process had become almost free to omit, and that nobody downstream could afford to detect the omission. That is not a moral event; it is a change in relative prices.
Hold on to that asymmetry, because everything that follows is the same asymmetry at larger scale. Skipping verification is cheap for whoever skips it and expensive for everyone obliged to establish that it was skipped.
Faced with work it could no longer vouch for, the sector then did something that deserves to be examined slowly. It did not build the capacity to check. It bought a machine and asked the machine to check on its behalf.
The machine was not very good at it. Pretoria reported approximately 53 AI-related disciplinary matters across 2024 and 2025, referred to its legal department, and now relies on triangulated evidence – drafts, version histories, a conversation in which a student is invited to explain their own argument – rather than on software. My own university discontinued Turnitin’s AI score with effect from 1 October 2025, following its Senate teaching and learning committee’s endorsement of a framework for AI in education.
The machines, meanwhile, went on issuing verdicts. At the University of Lagos in September 2025, candidates reported being flagged ineligible by an AI-assisted screening system for background noise, poor lighting and unreliable connectivity. One of them had scored 350 out of 400 in the national matriculation examination.
Note the direction in which that error travels. The detector bears nothing when it is wrong. The candidate in Lagos bears all of it, and must prove a negative to a system under no obligation to explain itself.
Then, this year, the pattern completed itself.
On 10 April, South Africa’s draft National Artificial Intelligence Policy was published in the Government Gazette. On 26 April the minister of communications and digital technologies, Solly Malatsi, announced its withdrawal – formally gazetted on 12 June – after at least six of its 67 academic citations were found to be fabricated. Some pointed to journals that do not exist. Others credited real journals, and real and respected scholars, with articles that were never published.
“This failure is not a mere technical issue,” the minister said, “but has compromised the integrity and credibility of the draft policy.”
I would not labour the irony; the facts are managing it unassisted. A national policy on artificial intelligence was produced by a process that skipped the very beat the policy existed to protect. At every tier of the system the reference list looked right. At every tier, nobody opened the journal.
Why verification loses
If this were a failure of character, we would expect it to be distributed like character: patchily, with conspicuous exceptions and a few heroes. It is not. It is distributed like an incentive.
Verification is expensive, slow, and invisible precisely when it works. A check that finds nothing produces no artefact, no number, no line in any report and no credit to anyone. Engagement is cheap, fast and visible; it arrives pre-quantified and on time. Any system under budget pressure will economise on the operation whose success cannot be observed. The model does this because its objective rewards plausibility. The university does it because its budget rewards throughput. One pressure, two systems, and nobody in the building acting in bad faith.
A university can absorb most irritations. It gives them a committee, a new name, a place in next year’s plan, and carries on. Unverified work is a rarer kind of pressure, because it reaches the legitimacy and the liquidity at the same time. And it compounds in a direction we cannot afford: every check we do not perform today is one that somebody – an examiner, an employer, a regulator – must perform later, on worse evidence, at higher cost, and with the institution’s word already in question.
What a check would cost
This is the point at which a list of governance rules would ordinarily appear. Three will do, and I have deliberately kept them off the classroom floor, because the classroom is not where this is being lost.
First, verification is a budget line or it is a wish. If an institution cannot say what proportion of submitted work was examined by a human being, and at what cost per unit, it is not verifying. It is hoping, and hoping is cheaper, which is precisely why it spreads. So publish the proportion. A faculty that cannot afford to check everything should declare what it does check, and declare it before the qualification is conferred rather than after it is doubted.
Second, whatever raises an allegation must carry the cost of being wrong. A detector risks nothing; the student risks a career. Any instrument permitted to accuse should be attached to a named human being who owns the error, and to an appeal that does not open with the accused proving a negative. If no one will put their name to the accusation, it is not evidence. It is weather.
Third, record the check, not only the verdict. The mark survives; the checking evaporates the moment it is done. Invert that. Make the account of how a judgment was reached the durable artefact, and the grade its by-product. Only then can anyone – an external examiner in Accra, an auditor in Kampala, a sceptical employer in Lusaka – establish years later that the beat was actually struck.
None of this requires a new technology, which is fortunate, because we have just been shown what happens when we wait for one.
The beat that is heard
In syncopation, the beat you decline to play is the one the listener hears most clearly. That is the whole trick: absence, placed precisely, carries further than presence.
A qualification works the same way. It is heard most distinctly at the point where somebody checked and could have said no. Skip enough of those, and what remains is a document with excellent rhythm and nothing underneath it – fluent, confident, properly formatted, and unable to say how it knows.
The drafters of a national AI policy did not set out to demonstrate the argument of this essay. That they did so anyway, in a gazette, under a minister’s signature, in a document about artificial intelligence, is the most useful thing about the episode. We have been taught the lesson early and cheaply. The fee for ignoring it will be neither.