After the University of the Free State's detectors wrongly flagged students' own essays, its student government reframed the issue: the training data itself is the integrity problem.
The University of the Free State retired its AI-text detection software at the start of the second semester on 1 July 2026, joining a small but growing list of institutions that have judged the tools unreliable enough to pull them out of the classroom.
According to OFM, a central South African broadcaster, students at the University of the Free State, a public university there, were being flagged by the detectors for essays they had written themselves. Siphiwayinkhosi Dlamini, the university's Student Representative Council (SRC) member for policy and transformation, led the SRC push to remove the tools, pairing the false-positive complaint with a deeper ethical objection about how the AI itself was made.
The argument that carried the day is not the usual "detectors don't work" complaint. In remarks to OFM, Dlamini put it this way: the integrity problem is not that the detectors are flawed. It is that generative AI models are trained on people's writing, artwork, and other creative works without consent or compensation. Those models replicate others' work as their own, he said, and amount to intellectual theft. That shifts the dispute past tool accuracy into how the AI was built.
A detector-failure complaint ends with "buy better software." An integrity-at-the-training-data argument ends with teaching, evaluation, and disclosure in place of policing. Dlamini's remarks to OFM explicitly stopped short of discouraging AI use and called instead for a dedicated AI ethics module. The move, in other words, is from running a detection tool to teaching the reasoning the tool was supposed to enforce.
What UFS actually changed is narrower than a wholesale ban. The policy applies across all faculties and took effect at the start of the second semester, NovaNews reported. Turnitin's similarity-checking feature, which compares a submission against existing published material rather than against what AI output "looks like," is staying. UFS published its own institutional explanation in a May 2026 campus notice, framing the change as an academic-integrity update rather than a technology ban. The Times of South Africa and The Witness both reported the move in May and June. The 7 July OFM confirmation closes the loop on the rollout, recording that the change is now enforced rather than proposed.
The SRC's additional integrity proposals are only partially on the public record. OFM describes further academic-integrity steps beyond the existing Turnitin similarity check, but the specifics of any replacement regime have not been fully disclosed in the sources reviewed here. The shape of what a UFS student now faces when submitting assessed work is, in part, still an open question.
UFS sits inside a wider pattern of detection-tool skepticism. Detection Drama's running 2026 list tracks other institutions that have reached similar conclusions. Paper-Checker's 2026 comparison of GPTZero, Turnitin's AI writing module, and Copyleaks documents the accuracy disputes that have powered those withdrawals. The academic record has been carrying the same message for years. Brandeis University's published guidance on the limits of AI detection tools describes detector scores as probabilistic and hard to interpret, with elevated false-positive rates against non-native English writers and against edited AI output. That technical ceiling is the floor on which any "we will catch cheaters with software" claim has to stand.
Two things to watch. First, the full set of replacement measures at UFS, beyond the Turnitin similarity check, when UFS or the SRC publish the detail. Second, whether other South African universities cite UFS as precedent, or whether the next institution to act is a private university with a different exposure to public-sector procurement and disclosure rules. One student government has now written the position that "the integrity problem is the training data, not the detector" into an institutional policy. The next test is whether it travels.