South African universities are adopting increasingly cautious positions on artificial intelligence detection tools. Some have discontinued their use, one has never adopted them, and many emphasise that AI-generated content detectors cannot be treated as conclusive evidence of academic misconduct.
Stellenbosch University (SU), Unisa, Vaal University of Technology (VUT), the University of Johannesburg (UJ), Nelson Mandela University (NMU), North-West University (NWU), the University of the Witwatersrand (Wits), the University of Cape Town (UCT) and the University of the Free State (UFS) spoke to ITWeb about their use of AI detection tools.
AI detection tools gained prominence amid growing concerns over the use of generative AI platforms such as ChatGPT and Gemini in academic work. However, studies have raised concerns about their accuracy, with researchers warning that the systems are susceptible to both false positives and false negatives.
UCT, SU and UFS have all moved away from AI detection tools, citing reliability concerns and placing greater emphasis on assessment redesign, academic judgment and AI literacy.
UCT disabled its AI detector in October 2025. SU discontinued Turnitin's AI text detection functionality at the end of 2025. UFS announced in May that it would discontinue AI detection software from 1 July, although Turnitin's standard similarity and plagiarism-detection tools will remain in use.
Instead of relying on automated detection, the institutions said they are focusing on assessment redesign, clear guidelines around acceptable AI use, invigilated assessments, oral presentations and process-based approaches. These allow lecturers to evaluate students' understanding and progression over time.
Dr Hanelie Adendorff, senior advisor at SU's Centre for Teaching and Learning, said reliance on AI detectors risks creating “an enforcement illusion: the belief that AI use can be controlled through detection when, in practice, such control is limited.
“A more sustainable response is to design assessment environments in which students understand the purpose of the task, know what forms of AI use are appropriate, and can be held accountable for the knowledge, skills and judgment they are expected to demonstrate.”
Prof Francois Strydom, senior director of the Centre for Teaching and Learning at UFS, said: “Because an allegation of academic misconduct carries severe psychological, academic and disciplinary consequences, UFS cannot and will not compromise the rights of its staff and students by outsourcing academic judgment to inaccurate, automated systems.”
UCT said removing AI detection tools had shifted attention towards learning rather than surveillance.
“The most significant impact has been a productive shift in focus," said Sukaina Walji, director of UCT's Centre for Innovation in Learning and Teaching. "Removing AI detection has allowed staff and students to move past the adversarial framing of detection and towards a more honest conversation about learning, expectations and the appropriate use of AI.”
Wits said it does not use AI detection tools and never has.
“Detection tools were known from the very start to produce false positives and negatives and should never have been used,” said Prof Nicole De Wet-Billings, senior director of academic affairs at Wits.
She added that the university would revisit the matter if sufficiently accurate tools became available.
Other institutions, including NMU, VUT and NWU, said AI detection reports are used as supporting information and are not treated as conclusive evidence of misconduct.
NMU said: "AI tools were never used as the sole mechanism to determine academic misconduct. Tools such as the Turnitin AI Similarity Indicator may provide a broad indication, but they are not regarded as definitive evidence.”
Similarly, VUT said it “does not rely on AI detection software as a primary mechanism for academic integrity investigations or assessment decisions”, adding that “no disciplinary action is based solely on an AI detection report”.
Unisa said it uses a “multipronged approach adopted from guidelines, policies, technology-related solutions to workshop training interventions”. Academics are required to review outputs before cases are referred to College Academic Integrity Committees, which validate reports before recommending them to the institutional Student Disciplinary Office.
NWU similarly said AI detection reports are used only as indicators for further investigation, although the university expressed confidence in the technology when used appropriately.
“These tools do not provide definitive proof. Lecturers and subject specialists must verify whether AI was used unethically – that is, whether a student presented AI-generated work as their own.”
The university added: “While it is not definitive, its accuracy rate of over 98% (with a false positive rate below 1%) makes it a useful initial indicator of possible unethical AI use.”
UJ said it has adopted a hybrid approach that combines pedagogical redesign with the selective use of similarity detection and AI detection tools.
“These tools are integrated into the learning management system and assist staff in flagging unusual patterns, paraphrased AI content and probable AI-generated segments, all subject to human review and interpretation,” said Prof Sehaam Khan, deputy vice-chancellor for academic affairs at UJ.
The university added that it is “moving beyond a narrow ‘ban or detect’ paradigm towards a more mature model that integrates AI literacy, assessment redesign, governance and equitable access”.
The responses from South African universities reflect a broader shift taking place internationally.
AI thought leader and AIforBusiness founder Johan Steyn said universities around the world are increasingly re-evaluating the role of AI detection tools.
“Evidence against them has become impossible to ignore. The tools do not work as advertised, and the students they fail most are precisely the ones institutions are meant to protect.
"UCLA, UC San Diego, Vanderbilt, Cornell and others deactivated their detectors between 2024 and 2025 over accuracy and equity concerns. This is no longer a fringe position. It is a growing institutional consensus that detection-based enforcement has failed its students.”
Steyn said the unreliability of AI detectors is structural, not a teething problem.
“These tools work by measuring something called perplexity – broadly, how predictable and simple the writing appears. Predictable, simple writing is exactly what a human student produces when writing in a second, third or fourth language. A 2023 Stanford study found that seven leading detectors flagged 61% of authentic, human-written essays by non-native English speakers as AI-generated, and 97% were flagged by at least one tool. The bias is baked into the architecture. You cannot patch your way out of a design that mistakes linguistic disadvantage for dishonesty.”
Steyn warned that over-reliance on AI detectors could expose universities to false accusations, reputational risks and unnecessary costs.
“First, falsely accusing students – disproportionately the most linguistically vulnerable – of misconduct can derail a degree, and in some countries even a visa. Second, legal and reputational liability for those false accusations.
“Third, financial waste: US institutions were paying anywhere from under $3 000 to over $110 000 a year for tools, many of them later switched off. For South African universities operating on constrained budgets, that is money not spent on the assessment redesign and academic development that actually work.”
Steyn said alternatives to AI detection tools include “authentic assessment – work that is hard to fake because it is designed to surface genuine thinking. Oral defences, reflective journals, iterative drafts, in-class writing components and assessments that reward the development of a student’s own voice and reasoning.
“These are not soft options; they demand more staff time and are harder to scale than a Turnitin submission, which is exactly why they require real investment in staff development and assessment-design expertise. The UFS framing is the right one: stop asking whether a student used AI, and start designing assessments that reveal whether they learned.”
Share
ITWeb proudly displays the 'FAIR' stamp of the Press Council of South Africa, indicating our commitment to adhere to the Code of Ethics for Print and online media which prescribes that our reportage is truthful, accurate and fair. Should you wish to lodge a complaint about our news coverage, please lodge a complaint on the Press Council’s website, www.presscouncil.org.za or email the complaint to enquiries@ombudsman.org.za. Contact the Press Council on 011 4843612.
Leave a Reply