Your AI Policy Has One Setting. Every Assignment Needs Four.

Somewhere in your organization there is a document that says what people may do with AI. It is probably one page. It probably applies to everyone, on every task, at once. And it is probably the reason nobody follows it.

This week in Times Higher Education, Kaśka Porayska-Pomsta, who directs the UCL Knowledge Lab and studies AI in higher education for a living, said universities have “moved too quickly and at the same time not quickly enough.” Read that twice. It is a diagnosis, and it describes the failure mode most institutions are sitting in right now.

Too fast and too slow, in the same building

Too fast, because institutions are buying and mandating tools ahead of any evidence they work. Porayska-Pomsta names it: they are “adopting technology that’s not really backed with evidence and best practices.” Vendors are happy to help. As she notes, “it’s in their interest.”

Too slow, because the hard work never starts. Nobody sits with a specific task and asks what the tool does to that task. So the rollout goes wide and shallow, and lands as what she calls implicit pressure “that everyone should use generative AI regardless of context.”

What she asks for is nerve. Educators should “have the courage to pause and to think about the practices which will allow for a more considered case.” Pausing looks like falling behind. It is the only way to stop being both too fast and too slow at once.

Manchester stopped writing one rule and started setting four values

In April 2026 the University of Manchester Senate approved something narrower and far more useful than a policy statement. Professor Sarah Dyer, Associate Vice President for Teaching Excellence and Innovation, had convened a working group in February. What it produced was a four value scale, applied per assessment.

Look at the fourth value. A policy that only ever restricts cannot contain it. Manchester built a scale that can say yes on purpose, which is what makes the no credible when it comes.

Dr Mark Carrigan described the mess it replaced: “students are effectively left to figure it out for themselves when presented with inconsistent guidance.” That line is the real cost of the one page policy. Students were not confused about the rule. They were confused because the rule said the same thing about a lab report and a reflective essay, so it told them nothing about either, and they were left guessing per class anyway.

Carrigan is equally clear about the other ditch: “attempting to be too prescriptive would inevitably be counterproductive but equally there was an urgent need for a path forward.” Four values is not much of a rulebook. That is the design. The scale carries the structure, and the person who knows the task carries the judgment. Unit leads classify their own assessments and explain the boundary to students in the assignment itself.

Stellenbosch refused to write the rule at all

Stellenbosch University reached the same conclusion and went further. It declined to issue a single institution wide rule, on the grounds that “a programming task, laboratory report, literature review, design project or professional placement may each require a different approach.”

Read that list again as a test of your own policy. If one sentence governs all five of those, it is either so loose it permits everything or so tight it breaks the programming task. There is no third option.

Stellenbosch gave academics four principles to decide with instead: authenticity, fairness, accountability, transparency. Then three moves available on any task. Allow or encourage AI. Prohibit it where the pedagogy justifies prohibiting it. Require it as part of the work.

The deciding question is the one worth stealing whether or not you teach: does the finished work still amount to “a valid representation of what the student knows, understands and can do.” Swap in employee, or analyst, or yourself. The question survives the swap intact.

You do not need a Senate to run this

Both of these are university stories, and the mechanism underneath is not. Any team that shipped an AI policy this year shipped the same shape of mistake: one rule, aimed at an average task that does not exist, written by people who were not going to be doing the task.

The manager who bans AI in client deliverables and permits it everywhere else has drawn the line in the only place that is easy to draw it, not the place it belongs. A first draft of a routine status update is Required. A discovery call summary somebody will act on without checking is Prohibited, and the ban has nothing to do with clients. Same policy, two settings, and the one page version cannot express either.

This is the whole SeedStacking argument in institutional clothing. You do not get fluent by adopting a tool. You get fluent by taking one real task and working out what the tool does to it. Manchester and Stellenbosch just proved that scales.

Set the dial on one task this week

Pick one task you assign, or one you do. Not the whole catalog. One.

1. Name what the task actually proves. Write the sentence. If the finished product cannot show it, the problem started before AI arrived.

2. Set the value. Prohibited, Minimal, Permitted, Required. Pick one and say why in a sentence. If you cannot say why, you picked from habit.

3. Write the line people will read. Not the policy. The one or two sentences that go on the task itself, where the person is standing when the question comes up.

4. Do a second task and compare. Two settings side by side is where you can see your own reasoning, and where a colleague can pick it up. One is an opinion. Two is a method.

Four sittings. About an hour and three quarters. At the end you have something a blanket policy cannot produce at any length: a defensible answer for one real task, and a method that survives contact with the next one.