Key Takeaways
Stellenbosch University rejected a single institution-wide AI rule, instead letting individual academics prohibit, restrict, permit, encourage, or require generative AI use based on each assessment's learning outcomes.
Deputy Vice-Chancellor Liezl van Dyk framed the new guidance as evidence-informed support for instructors rather than a fixed rulebook to enforce.
The model treats a programming assignment, a laboratory report, and a literature review as needing different AI permissions rather than one campus-wide standard.
The approach still anchors every instructor's decision to institutional principles of authenticity, fairness, accountability, and transparency.
Enterprise leaders weighing centralized versus team-level AI governance can read Stellenbosch's bet as evidence that outcome-based judgment can scale, provided training keeps decision-makers consistent.
No single rule for every assessment
Most university AI policy debates settle into one of two camps, banning generative AI in coursework outright or accepting it broadly with only light restriction. Stellenbosch University, in South Africa, has chosen a third path that sidesteps both extremes. Rather than issuing one campus-wide rule, the university published guidance that puts the decision in the hands of individual academics, assessment by assessment, course by course. Depending on what a given task is meant to measure, an instructor may prohibit AI entirely, restrict it to specific uses, permit it freely, actively encourage it, or in some cases require it as part of the assignment itself.
The underlying logic is that a single university-wide standard cannot account for how differently AI use affects a coding assignment, a laboratory report, and a literature review, three tasks that measure fundamentally different skills. A blanket ban treats all three the same, even though only some of them genuinely depend on unassisted work to test the skill they claim to measure. A blanket permission runs into the identical problem in reverse, applying uniform latitude to tasks that need very different levels of scrutiny.
Learning outcomes as the deciding factor
Liezl van Dyk, Stellenbosch's Deputy Vice-Chancellor for Academic Affairs, described the new resource as offering "practical, evidence-informed guidance for academics, from quick-reference resources to more in-depth material on AI in teaching, learning, assessment, research and scholarship." The framing is deliberate, positioning the guidance as support for a judgment call each academic makes themselves, backed by evidence and institutional precedent, rather than a rule they simply apply on autopilot without engaging with the specifics of their own course, their own students, or the particular skill a given assignment is designed to build.
Instructors are asked to start from what a task is actually assessing before deciding on AI permissions. A programming assignment testing whether a student can structure a working solution may reasonably allow AI as a coding partner, since the graded skill is architecture and problem decomposition. A lab report testing whether a student can observe, measure, and interpret their own data calls for a much stricter line. A literature review testing synthesis and argument construction sits somewhere between those two poles, with the right answer depending on the specific skill the assignment is built to measure.
The tradeoff nobody says out loud
This approach is considerably harder to implement than a blanket rule, and Stellenbosch's own guidance implicitly acknowledges as much by investing in detailed reference material rather than shipping a one-page policy memo and calling it done. Every academic now needs enough fluency in both their assessment's actual purpose and AI's real capabilities to make a defensible call, and then needs to make that same call consistently across every section of the same course, semester after semester, as new AI tools and student behaviors continue to shift underneath them.
That consistency requirement carries a real operational cost that is easy to underestimate at the policy-writing stage. A blanket ban or blanket permission requires almost no training beyond communicating the rule clearly once. A distributed judgment model requires ongoing calibration across what could be thousands of instructors campus-wide, and it opens a genuine door to inconsistency between two sections of the same course taught by different academics with meaningfully different comfort levels around AI tools and their limitations.
Why this is a governance question beyond campus
Stellenbosch's bet is that outcome-based judgment, anchored to explicit principles of authenticity, fairness, accountability, and transparency, produces better decisions over time than a single rule designed in advance to cover every possible case. That is a governance problem that extends well past academia. Any organization writing AI usage policy for teams doing genuinely different kinds of work, engineering compared with legal compared with marketing, faces the same underlying choice between one enforceable standard applied everywhere and many context-specific judgment calls made closer to the actual work.
The universities and companies choosing distributed judgment are betting that their people are equipped to make sound calls once given the right guardrails, treating judgment as an asset worth investing training dollars into rather than a liability to design around. That bet only pays off when the guardrails are detailed enough, and the training behind them consistent enough, to keep the model from collapsing into whatever each individual decision-maker happens to prefer on a given day, which is the real risk hiding inside any distributed governance approach.
The decision this puts on your desk
If your organization's AI policy currently reads as one rule applied uniformly across very different teams, Stellenbosch's approach is worth studying closely as a working alternative model rather than something to copy wholesale overnight. The core question it forces is whether your existing policy is actually calibrated to what each type of work is meant to produce, or whether it is simply a single standard chosen for how easy it was to write and communicate rather than for how well it fits the work it governs.
The honest tradeoff sits between training investment and enforcement simplicity. A distributed model demands considerably more upfront work to get decision-makers aligned and steady ongoing effort to keep them that way as tools and roles evolve, but it produces policy that fits the work rather than work that gets bent awkwardly to fit the policy. Stellenbosch has placed its bet on that tradeoff paying off. Whether it holds up over time will depend entirely on how consistently the university invests in keeping thousands of individual judgment calls pointed in the same direction.
