A Four-Tier Label on Every Assessment
The University of Manchester approved an AI in Teaching and Learning Policy that requires every summative assessment to carry one of four AI categories, according to reporting on 23 July 2026. The categories run from AI Prohibited, where any use is forbidden, through AI Minimal, which allows limited help such as copy editing, to AI Permitted, which names specific approved uses, and AI Required, where students must use AI to complete the work. Each category has to be stated in the assessment brief and repeated on Canvas so students know the expectation before they begin.
The university expects most assignments to land in the two middle bands, where AI is allowed within defined limits. That distribution matters because it moves the default away from a blanket ban that staff cannot enforce and away from an open door that undermines the point of the assessment. By forcing a per-task decision, the policy makes AI use a design choice attached to each learning objective. Unit leads have to think about what a specific assessment is meant to measure and whether a model helps or hollows out that goal.
Why Clarity Became the Priority
The policy responds to years of inconsistent guidance that left students guessing. Since ChatGPT arrived in November 2022, different courses and lecturers sent conflicting signals about whether AI was cheating or a required skill. Dr Mark Carrigan, Senior Lecturer in Education and an AI Fellow at the university's Institute for Teaching and Learning, described the problem directly. "In the absence of agreement we confront students with a cacophony of messages," he said, pointing to the gap between staff who condemn AI and staff who insist students must learn to use it.
Professor Sarah Dyer, Associate Vice-President for Teaching Excellence and Innovation, framed the goal as reducing that confusion. She said the framework was designed to provide clarity for staff and students while acknowledging the genuine differences of opinion surrounding AI in higher education. That framing is deliberate. The policy does not try to settle whether AI belongs in education. It gives everyone a shared vocabulary so the argument happens once, at the level of each assessment, and produces a clear instruction for the student rather than an unresolved standoff between departments.
Decisions Stay With the People Who Own the Work
A defining feature of the Manchester approach is where authority sits. The choice of category for any assessment is treated as a disciplinary decision, made by academic staff who understand what the assignment is meant to develop and test. The policy supplies a shared language and a set of definitions, and it leaves the classification to the people closest to the learning. Carrigan noted the framework cannot resolve the debate on its own, though he said it can contribute to making these conversations a routine feature of the work.
That design decision is the transferable part. Central teams that try to write one AI rule for every situation tend to produce guidance that is either too loose to matter or too rigid to fit real tasks. Manchester instead standardized the format of the decision and delegated the content of it. Every assessment gets a category, the category is chosen by a subject expert, and the whole policy is reviewed annually. The result is consistency in how the rule is expressed paired with judgment in how it is applied.
The Governance Gap This Exposes
Manchester is moving while most institutions are still improvising. A survey from D2L and Tyton Partners found that 71 percent of administrators, 52 percent of instructors, and 61 percent of students now use AI at least weekly. Adoption has clearly outrun governance. The same research found that only 32 percent of institutions report having a centralized AI policy, and just 22 percent of faculty at those institutions consider the policies effective. Weekly use is close to universal while formal rules remain the exception.
That gap will look familiar to any technology leader. Employees have already wired AI assistants into their daily work, often faster than the organization can write policy, and the guidance that does exist is frequently ignored because it feels disconnected from actual tasks. Manchester's contribution is a workable pattern for closing the gap without pretending to have every answer. Attach a clear, task-level label to the work, put the classification in the hands of the people who own the task, and make the label visible at the moment the work begins.
A Template Enterprise L&D Can Reuse
Corporate learning teams face the same question Manchester answered, with higher stakes around client data and compliance. Employees complete assessments, certifications, and skills checks inside an LMS, and a single blanket policy on AI use rarely survives contact with the variety of that work. A compliance exam and a creative brief demand different rules. The four-tier model offers a ready structure: mark each course, assessment, or task as AI Prohibited, Minimal, Permitted, or Required, and state the rule where the learner sees it.
The deeper lesson is about how to write AI rules that people follow. Governance holds when it is specific to the task, owned by someone with authority over that task, and surfaced at the point of use. For L&D leaders under pressure to certify an AI-ready workforce while proving that assessments still mean something, Manchester supplies a defensible middle path. It lets people use AI where it builds capability and blocks it where it would fake the very skill the training exists to verify.
What It Means for the Roadmap
The practical takeaway is that AI governance works best as metadata on the work itself, not as a standalone document nobody reads. Leaders planning learning and knowledge programs for the next year should consider tagging content and assessments with a clear AI category, wiring that label into the LMS, and reviewing the scheme on a set cadence as tools and norms shift. The engineering lift is small. The behavioral payoff is a rule employees encounter exactly when they need it.
The wider implication reaches past training into how organizations classify AI use across all knowledge work. Manchester has effectively piloted a labeling system for when AI is off limits, assistive, permitted, or expected, and it did so in an environment with strong incentives to get integrity right. Enterprises weighing how to let staff use AI without eroding standards now have a tested reference model. Adopt the format, delegate the judgment, and make the category visible, and AI policy stops being a poster on the wall and starts shaping day-to-day work.
