The Four Boxes Every Assessment Now Needs

A senior lecturer in sociology at the University of Manchester opened her spreadsheet on 3 September 2026 and counted 14 assessments she had set across three units. Each one now needed a letter: A, B, C, or D. The university's new AI in Teaching and Learning Policy took effect on 1 September 2026, requiring every summative assessment to be assigned to one of four AI-use categories, and unit leads in the Faculty of Humanities had until 22 September to classify their assessments and post the result on Canvas.

The policy is not a recommendation. Every assessment must be assigned to one of four categories, and students must be told which one. The four letters map to four levels of permitted AI use, and the category assigned to an assessment determines the information that must be included in the assessment brief.

She had taught two of these units for six years. The assessments worked. Students passed them, struggled with them in predictable ways, and occasionally surprised her. But the policy does not care whether an assessment worked last year. It asks a different question: what is a student allowed to do with a model in order to submit this piece of work?

The four categories do not represent institutional positions on pedagogy. Decisions about which category applies stay with the academics who set each assessment. The policy standardised the labels, not the thinking. What it did do is make a question that used to live as an uneasy silence in a syllabus into a mandatory, published decision that sits next to every grade-carrying task a student receives.

1. Category A: AI Prohibited

Category A is AI Prohibited. A student may not use generative AI for any part of the work. She looked at her third-year research methods essay, a 3,000-word piece analysing a qualitative dataset she provides. She could put it in A. The question is whether she can defend it.

The policy asks unit leads to articulate their expectations around the use of AI by students and to ensure these are communicated in a clear and effective manner, which does not mean changing existing assessment practice but rather ensuring that what you ask of students in relation to AI use is made entirely clear to them. A complete prohibition is clear. It is also the category that carries the highest enforcement cost if a student contests an allegation.

Students should know when AI touched their work and have a route to human review, and never lose an appeal to a detector score, and the Financial Times reports students are now bringing lawyers to misconduct cases. If she puts the essay in Category A and a submission feels wrong, she will need to document why without relying on a probability score from a commercial tool. The policy says failure to document use of generative AI in assessed work could be a form of plagiarism under the university's Academic Malpractice Procedure, but proving non-documentation when the work reads plausibly is a different problem from proving copy-paste.

She moved the essay to Category B.

2. Category B: AI Minimal

Category B is AI Minimal, for presentation such as grammar and formatting. A student may use AI to polish language, fix citations, check spelling. They may not use it to generate arguments, structure sections, or write sentences that carry analytical content.

The boundary sounds clean until she tries to write it into an assessment brief. A student pastes a paragraph into Claude and asks it to "make this clearer". The model rephrases three sentences, tightens two claims, and suggests a topic sentence the student did not write. Is that polishing or drafting?

A 2026 report from Tyton Partners shows 47% of faculty are modifying assessment design because of AI. The modifications are not all in the same direction. One group, labeled Defenders, is responding to AI by returning to more controlled formats such as in-class exams and blue books, while a third group, Integrators, is redesigning assessments around AI use through project-based work, iterative assignments, AI disclosure, and student critique of AI output. Category B tries to hold a middle ground between them, and the ground keeps moving.

She had assumed B would be the safe default. It turned out to be the category that required the most precise language in the brief, because every verb in the instruction now needed a paired statement about which parts of that verb could touch a model. "Analyse the interview data" becomes "Analyse the interview data (you may use AI to correct grammar in your final draft but not to code themes, generate quotations, or write interpretive claims)".

Three of her 14 assessments went into B. She spent 90 minutes writing the expanded briefs and then read them back. They sounded like legal disclaimers.

3. Category C: AI Permitted

Category C is AI Permitted, with acceptable use set out in the brief. Anything which extends into the substance of the assessment belongs in category C, where the requirement is to specify what is acceptable rather than leaving it open.

This is where the redesign happens, not as a pedagogy workshop but as a compliance task with a 22 September deadline. She looked at her second-year group project, a media analysis that had always produced uneven quality because some groups split the work and some actually collaborated. She could rewrite it to require students to use a model as a research assistant, log every query, and submit the log alongside a reflective essay explaining what the model got wrong.

Integrators assign an average of 22% of the course grade to AI-incorporated assessments and report fewer challenges with attendance and cheating than Defenders because they are implementing AI policies that redesign the work itself so students demonstrate reasoning, application, collaboration, or revision. The trade is clear: less policing, more grading of process evidence. She would need to assess 11 logs and 11 essays instead of 11 reports, and the rubric she had used for six years would need to be rewritten to evaluate something it was never built to measure.

Educators are being asked to police AI use, redesign assessments, and keep pace with rapidly evolving tools, often without dedicated support or protected time to do any of it properly. Her workload model allocates 40 hours to the unit, a figure set in 2019 and unchanged since. The policy does not come with time; it comes with a date.

She put five assessments into C and noted that she would need to rewrite all five briefs and three rubrics before the teaching block started in two weeks.

4. Category D: AI-Integrated

Category D is AI-Integrated, where AI use is necessary and students cannot opt out. Students may choose not to use AI in any assessment except those in Category D.

She had one assessment she thought belonged here: a new third-year elective on computational social research where the learning outcome explicitly included "use AI tools to clean, explore, and visualise social data". The brief already required model use. Labeling it D formalised what was already true.

But D also creates a new problem. Every assessment brief must clearly state its AI category and explain what students may and may not use AI for, and where AI has contributed to an assessment, students are expected to include an AI Reflection explaining how AI was used, how outputs were reviewed or revised, and what their own contribution was. The AI Reflection is not a checkbox; it is a separately assessed piece of metacognitive writing that has to be scaffolded, explained, and graded.

Professors in three University of Virginia departments described their experience with AI use in their own classrooms where large language models pushed them to rewrite assignments, restructure exams and in some cases, rebuild entire courses. Rebuild is the right word. A Category D assessment is not an existing task with AI added to it. It is a task redesigned so that competent AI use is the thing being measured, and everything else in the brief has to serve that measurement.

She put one assessment in D and made a note to talk to her head of department about whether the workload model recognised the assessment type that did not exist when the model was written.

What the letters make visible

The sociology lecturer submitted her classifications on 19 September, three days before the deadline. Twelve of her 14 assessments now carried a letter that had not been there in August. Two assessments, both in-class tests, stayed in Category A because she could supervise them. Everything else moved.

The policy standardises how AI rules are communicated. It does not standardise the rules themselves, and it cannot. A law seminar putting oral advocacy into Category A is making a different pedagogical bet from a statistics course putting the same task into D because the statistician wants students to use a model to generate the counter-argument they will then dismantle. Both are defensible. Neither is portable.

What changed on 1 September was not the university's position on AI in assessment. The policy introduces a common framework for communicating expectations to students of how AI can be used in assessed work and sets out a clear, proportionate ask of teaching colleagues. The change is that the decision moved from implicit to explicit, from something a student might guess at by reading a syllabus to something that must be declared in four published places: the brief, the Canvas page, the marking rubric, and the academic misconduct procedure.

57% of students say they lack adequate AI guidance for assessments, only 29% believe their instructors can guide them on AI use, and just 31% of faculty say their institution involves them meaningfully in shaping AI policy. Manchester's policy does not close that gap by telling faculty what to think. It closes it by forcing them to write the answer down and show it to the student before the work is submitted.

The lecturer I spoke to said the hardest part was not picking the letter. It was realising how many of her assessments had survived six years because nobody, including her, had asked what they actually measured. A 3,000-word essay can measure research skill, argument construction, source evaluation, or adherence to a house style. When a model can do three of those four, the brief has to be rewritten around the one thing left, and sometimes the thing left is not the thing the degree programme said it was assessing.

She has 14 letters on a spreadsheet now.Next year she will have to justify them.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.