There is no national survey of artificial intelligence use among South African students. Not from the Department of Higher Education and Training, not from the Council on Higher Education, not from Statistics South Africa. Every percentage you have seen attached to “South African students” traces back either to a survey conducted somewhere else, or to a single class at a single institution.
That is an uncomfortable answer for anyone who needs a citable figure for a literature review or a policy paper. It is also the correct one, and using it well is more defensible than borrowing a number that was never measured here.
What actually exists, and what it measures
| Study | Where the fieldwork happened | Sample | Source and year |
|---|---|---|---|
| Undergraduates’ ChatGPT dependence as a tool for learning | One university of technology, KwaZulu-Natal | 50 first-year students | Business Excellence and Management, 2024 (Mangosuthu University of Technology) |
| Engineering students’ perceptions and use of generative AI | One South African university | Undergraduate and postgraduate engineering students; the abstract does not state the number | Computer Applications in Engineering Education, 2025 (Stellenbosch University) |
| Student perceptions of generative AI in programming education | Kenya, Nigeria and South Africa | 322 university students across all three countries | Computers and Education Open, 2025 (University of Johannesburg co-author) |
Three studies. One with 50 respondents, one whose sample is not stated in its abstract, and one whose 322 respondents are divided across three countries so the South African component is a fraction of that. None of them was designed to produce a national estimate, and none of their authors claims one.
What they do report is worth knowing. The Mangosuthu study found that most of its 50 first-year respondents were familiar with ChatGPT, used it across all their modules, and expressed trust in it — a finding specifically about students from linguistically and socio-economically constrained backgrounds, which is a population most international surveys do not reach. The Stellenbosch engineering study reports widespread adoption, students who support AI as a learning aid while being uneasy about it in formal assessment, a preference for discipline-specific guidance over blanket institutional policy, and only 1% supporting a complete ban on generative AI use.
Where the numbers you have seen actually come from
The figures in general circulation — that almost all students now use AI, that the proportion has climbed from roughly half to nearly universal in two years — are real, well-sampled and repeatedly measured. They are also British.
| Finding | Figure | Source and year |
|---|---|---|
| Students using AI in at least one way | 95% | HEPI, Student Generative AI Survey 2026 (Report 199), UK, fieldwork Dec 2025 |
| Students using generative AI to help with assessed work | 94% | HEPI, Student Generative AI Survey 2026, UK |
| Students directly including AI-generated text in assessed work | 12% (from 8% in 2025 and 3% in 2024) | HEPI, Student Generative AI Survey 2026, UK |
| Students saying assessment has changed significantly in response to AI | 65% | HEPI, Student Generative AI Survey 2026, UK |
| Students who feel their institution encourages AI use | 36% | HEPI, Student Generative AI Survey 2026, UK |
| Students provided with AI tools by their institution | 38% | HEPI, Student Generative AI Survey 2026, UK |
| Students who believe AI skills are essential | 68% | HEPI, Student Generative AI Survey 2026, UK |
| Students who feel teaching staff help them develop AI skills | 48% | HEPI, Student Generative AI Survey 2026, UK |
| Students using AI for assessments, previous wave | 88% (from 53% the year before) | HEPI, Student Generative AI Survey 2025 (Policy Note 61), UK |
The 2026 edition was conducted by Savanta in December 2025 among 1 054 full-time UK undergraduates; the 2025 edition surveyed 1 041. That is what a national measurement looks like: a defined population, a professional fieldwork provider, a stated sample and three waves so change can be read rather than guessed.
Nothing of that description exists for South Africa. Quoting the 95% figure in a South African context is not a small approximation — it imports a country with near-universal fixed broadband, institutional device provision and a very different fee and access structure, into a system where the constraint on tool use is frequently the cost of the data to reach the tool. Our figures on student internet and data access in South Africa set out how large that difference is.
The indexing trap: studies that are not as South African as they look
Search a bibliographic database for South African research on student AI use and the result overstates what exists. Databases attribute a study to a country by the affiliation of its authors, not by where the fieldwork happened. Three examples from a single search of works with a South African affiliation:
- A survey of 59 library and information science students in Zimbabwe, indexed to South Africa because an author is affiliated to the University of South Africa.
- A structural equation model of 176 undergraduates at a public university in south-western Nigeria, indexed to South Africa through a University of Zululand affiliation.
- A mixed-methods study of 402 students at the University of Ibadan, Nigeria, indexed partly to South Africa through a University of Johannesburg co-author.
All three are legitimate studies. None of them measures South African students. If you build a literature review by filtering on country and counting hits, you will report an evidence base roughly twice the size of the real one.
The check takes thirty seconds: open the abstract and find the sentence naming the participants and where they were recruited. If the abstract does not say, the study cannot be counted as evidence about a specific country until you have read the methods.
Where we looked for a national figure, and what we found instead
| Source checked | What it publishes | Does it measure student AI use? |
|---|---|---|
| Department of Higher Education and Training / HEMIS | Enrolment, graduations, staffing, institutional returns | No |
| Council on Higher Education | VitalStats participation and throughput; qualification frameworks; institutional audits | No |
| Statistics South Africa, General Household Survey | Household internet access, device ownership, connectivity | No — access, not application use |
| National Research Foundation annual reporting | Funding allocations and grant-holder statistics | No |
This is a documented absence rather than a claim that no such data could exist anywhere. Individual universities run internal surveys and some report them at conferences; those are not published national statistics and should not be cited as though they were.
What to write in your dissertation instead
You do not need a national figure to justify studying this. You need an accurate statement of what is known, which is a stronger opening than a borrowed percentage.
- State the absence, and cite the search. “No national survey of generative AI use among South African students has been published; the available South African evidence consists of single-institution studies with samples of 50 to a few hundred.” That sentence is defensible and an examiner can verify it.
- Cite the international benchmark as international. “In the United Kingdom, HEPI’s 2026 survey of 1 054 full-time undergraduates found that 95% report using AI in at least one way.” Naming the country and the sample is what makes the citation legitimate.
- Cite the South African studies for what they measure. Attitudes and use within a named cohort at a named institution — never as national prevalence.
- Turn the gap into your rationale. An absent national figure is the clearest justification a South African study of this topic can have. Say so explicitly in your problem statement.
The same care applies to the policy side. What your university permits is a separate question from how many students do it, and it is answered by your institution’s own rules — our explainer on whether AI is allowed for dissertations at South African universities covers what UCT, Wits, Stellenbosch and Unisa actually say, and our guide to what an AI dissertation writer should and should never do covers declaration.
How to cite the figures on this page
Give the survey, the country, the sample size and the fieldwork date every time. “95% of students use AI (HEPI Student Generative AI Survey 2026, n = 1 054 full-time UK undergraduates, fieldwork December 2025)” is a citation an examiner can check. “95% of students use AI” is not, because the reader cannot tell which students.
For context on the population these figures would apply to if they existed, our national figures on South African postgraduate enrolment give the denominator: how many master’s and doctoral students there actually are.
Tesify keeps each figure attached to the source you took it from, with its year and its sample, so the qualification travels with the number instead of being reconstructed the week before submission — which is precisely how a British statistic ends up in a South African sentence.
Build your literature review in Tesify
Frequently asked questions
How many South African students use AI?
No national survey has measured this. The published South African evidence consists of single-institution studies — one of 50 first-year students at a KwaZulu-Natal university of technology, one of engineering students at Stellenbosch, and one multi-country study of 322 students across Kenya, Nigeria and South Africa.
Is it true that 90% of students use AI?
In the United Kingdom, yes: HEPI’s 2026 survey of 1 054 full-time undergraduates found 95% using AI in at least one way and 94% using it for assessed work. That figure has not been measured in South Africa and should not be reported as if it had.
Why is there no South African figure?
Because no national instrument collects it. DHET and HEMIS collect enrolment and graduation data, the CHE publishes participation and throughput statistics, and Stats SA measures internet and device access rather than what students do with it.
Can I cite the UK survey in a South African dissertation?
Yes, provided you name it as United Kingdom data with its sample size and fieldwork date, and do not present it as a South African prevalence estimate. Using it as an international benchmark is legitimate; using it as a local figure is not.
What is the largest South African study on this?
Of the studies we located, the multi-country programming-education study with 322 respondents is the largest, but its sample is divided across Kenya, Nigeria and South Africa, so the South African component is smaller still.
Why do database searches return more South African studies than this?
Because bibliographic databases attribute works by author affiliation, not by fieldwork location. Several studies indexed to South Africa surveyed students in Zimbabwe or Nigeria. Read the methods section before counting a study as national evidence.
Do South African students use AI less than students elsewhere?
Nobody knows, and any confident answer is invented. There are plausible reasons to expect differences in either direction — mobile data cost and device access on one side, high mobile penetration and large distance-learning cohorts on the other — but no measurement to settle it.
What did the Stellenbosch engineering study find?
Widespread adoption of generative AI applications, general support for it as a learning aid alongside concern about its use in formal assessment, a preference for discipline-specific guidance over institutional policy, and only 1% of respondents supporting a complete ban.
What did the Mangosuthu study find?
Among 50 first-year students surveyed on mobile devices, most were familiar with ChatGPT, used it across their modules and expressed trust in it. The authors frame it as evidence about students from linguistically and socio-economically constrained backgrounds specifically.
Is a small single-institution study worth citing?
Yes, cited accurately. A study of 50 students at one institution is evidence about those students. It becomes a problem only when it is quoted as a national rate, or when several such studies are added together as though they formed a representative sample.
How should I phrase the gap in my problem statement?
Directly: no national survey of generative AI use among South African students has been published, the available evidence is confined to small single-institution samples, and the figures in general circulation were measured in other countries. That is a rationale, not an apology.
Will a national survey be published?
We cannot say. If one appears from DHET, the CHE or a national body with a stated sampling frame and fieldwork date, it will supersede everything on this page — and that is the standard to hold any new figure to before you cite it.