University of Manchester adopts four-tier AI policy for student assessments

Unit leads must define permitted AI use for individual assignments under a framework intended to replace inconsistent guidance with ongoing discussion among staff and students.

The University of Manchester’s policy establishes four categories for defining acceptable AI use in student assessments

The University of Manchester has adopted its first AI in Teaching and Learning Policy, requiring unit leads to classify assignments according to four levels of permitted artificial intelligence use.

Approved by the university’s Senate in April 2026, the framework ranges from assessments in which AI is entirely prohibited to those where it is required. Most assignments are expected to fall within the two middle categories, covering either minimal AI use, such as copy editing, or specific uses communicated to students through assessment guidance.

The policy was developed by a working group convened in February by Professor Sarah Dyer, Associate Vice-President for Teaching Excellence and Innovation. Academic staff, professional services colleagues and student representatives were involved, bringing different positions on the role of AI in higher education into the process.

Reflecting on the work on LinkedIn, Dyer said the group had sought to “provide clarity for staff and students while acknowledging the genuine differences of opinion surrounding AI in higher education.”

The policy’s development and underlying approach were examined in a University of Manchester blog published by Dr. Mark Carrigan, Senior Lecturer in Education. Carrigan co-leads the Digital Education Manchester group and serves as an AI Fellow at the Institute for Teaching and Learning.

In the blog, titled “Building a working consensus on AI in teaching and learning,” Carrigan presented the policy as an attempt to establish shared expectations in an area where university staff do not agree on whether AI should be excluded, accommodated or actively incorporated into teaching.

“Most academics perceive the presence of AI in their classrooms and their student’s work. There is much less agreement however about how we ought to respond to it,” he observed.

Carrigan described positions ranging from colleagues who believed protecting the values of higher education required universities to keep AI out, to those who saw little realistic prospect of doing so. Others, he noted, believed institutions needed to adapt to prepare students for a future in which the technology would be widespread or explore its potential to support more engaging and personalized teaching.

Addressing inconsistent AI guidance

The absence of a shared position had practical consequences for students, Carrigan argued. He said policy needed to do more than secure compliance if it was to influence teaching practice and establish institutional norms.

“In the absence of agreement we confront students with a cacophony of messages, ranging from condemnations of AI through to claims they must learn how to use it or risk being left behind,” he stated. “This means that students are effectively left to figure it out for themselves when presented with inconsistent guidance.”

The university had struggled to offer a clear institutional position following the release of ChatGPT in November 2022, according to Carrigan. The working group concluded that students urgently needed greater clarity about what counted as appropriate AI use in assessments.

However, the group did not attempt to produce a fixed answer for every subject or assignment. Carrigan pointed to the speed of technological change, varying levels of AI literacy among staff and students, and the different requirements of disciplines across a large university.

“Attempting to be too prescriptive would inevitably be counterproductive but equally there was an urgent need for a path forward,” he explained.

Responsibility for classifying individual assignments now sits with unit leads. Teaching teams are expected to provide students with guidance explaining the boundaries attached to the selected category.

Moving beyond a traffic light model

The framework departs from the red, amber and green traffic light systems used elsewhere by introducing a fourth category. Carrigan acknowledged that the additional distinction created its own potential ambiguity, particularly around the boundary between AI Minimal and AI Permitted.

Rather than presenting that difficulty as resolved, he argued that requiring unit leads to draw the distinction should prompt more detailed consideration of how AI relates to the learning and assessment objectives of a particular unit.

“The problem with the traffic light system tends to be that most assignments get categorised as amber, which means the question of clarity gets displaced,” he said.

An amber classification could appear to offer a safe middle option, Carrigan explained, but it provided limited help if students were not also told which uses were acceptable. The university’s four-tier system is intended to move those decisions into discussions among unit leads, teaching teams and programs.

“The advantage of this approach is that it encourages us to draw these distinctions,” he argued. “It invites reflection by unit leads and conversation within teaching teams and programmes about these boundaries and what they mean for assessment.”

The distinction between the two middle categories will therefore depend on the guidance attached to each assignment. The category alone does not establish which tools, tasks or forms of assistance students may use.

Policy treated as an evolving approach

Carrigan cautioned against presenting the classification system as a permanent resolution to AI use in assessment.

“It is an ‘approach’ though, not a solution. AI in assessment is a classic example of a wicked problem. There are only better or worse options, involving trade offs,” he warned.

His blog identified coding agents, wearable computing and AI companions among the developments that could create further challenges for universities. Against that background, the policy focuses on establishing a process through which expectations can be discussed and revised.

Carrigan placed the framework in the context of Russell Group principles that called for regular dialogue between academic staff and students to establish a shared understanding of appropriate generative AI use.

The Manchester policy seeks to support that dialogue in two ways. First, unit leads must examine their assignments and identify the expectations that need to be communicated to students. Carrigan suggested this could also expose areas where acceptable use had not been agreed within teaching teams.

Second, the university will review the policy annually, allowing it to be revised as technology and institutional practice develop. Carrigan characterized the review as a governance mechanism rather than a substitute for substantive discussion, but said it could make conversations about what was and was not working a more routine part of university teaching.

“The policy will not bring that about in itself but it can contribute to making these conversations a routine feature of our work together within the university,” he concluded.