How Schools Write an AI Marking Policy
Most schools wrote an academic integrity policy the moment generative AI arrived — a document about what students may and may not do. Far fewer have written the other half: a policy on how staff use AI to mark. As AI marking tools move from a few enthusiastic early adopters to whole departments, the absence of a clear institutional position starts to show. Teachers want to know whether they are allowed to use these tools and how. Data protection officers want to know what happens to student work. Governors want assurance that standards and accountability have not quietly been handed to a machine. An AI marking policy for schools answers all three before they become a problem, and writing one is more straightforward than it sounds.
This guide is a practical framework for senior leaders drafting that policy. It is written for the institutional view — the deputy head, the assistant head for teaching and learning, or the SLT lead who owns assessment — and it focuses on the decisions a policy actually has to make rather than on boilerplate. It deliberately covers different ground from a student-facing academic integrity policy: this one governs your staff, not your pupils.
Define Scope: What the Policy Covers
The first job of the policy is to say plainly what it applies to, because "AI marking" can mean very different things. A useful policy distinguishes between AI that produces a draft mark and feedback against a defined rubric — the case this document is really about — and the much looser practice of a teacher pasting work into a general-purpose chatbot. The two carry completely different data-protection and reliability profiles, and a school is far safer steering staff toward a purpose-built tool with a clear contract than tacitly tolerating ad-hoc use of consumer chatbots.
Scope should also name where AI marking is and is not appropriate. Many schools land on permitting it for formative assessment, mocks, low-stakes class tests, and feedback generation, while keeping a human-only line around anything that feeds a formal reported grade or a statutory record without teacher sign-off. Setting that boundary in the policy, rather than leaving each teacher to guess it, is what lets staff use the tools confidently. It also gives you a single answer when a parent or governor asks "do you let computers grade my child's work?" — the honest answer being that a teacher always signs off the grade, and the policy says so.
The Non-Negotiable: Teacher Judgment Stays in Charge
The clause that does the most work in an AI marking policy is the one that keeps the teacher accountable for every mark. AI marking is assistive: it produces a draft against criteria the teacher has set, and the teacher reviews, adjusts, and owns the final result. A policy that states this clearly protects the school on every front at once — it preserves professional accountability, it reassures parents and governors that standards have not been outsourced, and it sets the right expectation with staff that the tool saves time on the mechanical bulk of marking without removing their responsibility for the outcome.
It is worth writing this as a positive requirement rather than a caveat. The policy should say that a draft mark is never released or recorded without a teacher's review, that teachers remain responsible for the accuracy and fairness of marks, and that the tool's role is to speed up and standardise, not to decide. Framed that way, the judgment clause is not a brake on adoption; it is what makes confident adoption defensible. Tools built for schools support this directly — GradeOrbit puts every AI-generated mark in front of the teacher to review before it counts, so "teacher signs off" is how the workflow already behaves, not a rule bolted on top of it.
Data Protection: The Section Your DPO Will Read First
The clause your data protection officer cares about most is what happens to student work. AI tools attract extra scrutiny because the default assumption is that they hoard data to train on, so the policy should state your school's position explicitly and require any tool in use to meet it. The questions to pin down are simple: is student work stored, and if so where and for how long; is it used to train AI models; and is there a signed agreement covering all of it.
Schools using GradeOrbit's institutional tier sign a Data Processing Agreement at onboarding, which is the document your DPO will want on file. Under the school journey, student work is persisted deliberately — so teachers can return to a student's marked work and results, and so the school retains its own data — but it is held within a defined retention window, wiped on a schedule, and removed in full when a school offboards. Customer work is not used to train AI models. The policy should require that any AI marking tool the school adopts comes with a DPA, a clear retention position, and a no-training guarantee, and should name the approved tool or tools so staff are not free-styling with whatever consumer service they happen to like. That single section is usually the shortest third-party review a DPO will run all year, provided the tool's answers are specific.
Equity, Consistency, and the Standards Case
A good AI marking policy also makes the positive case, because used well these tools improve consistency rather than threaten it. Marking drifts when it is spread across tired evenings and snatched free periods; the same rubric applied uniformly to every script removes that drift. A policy can legitimately frame AI marking as a standardisation aid — a way to mark a whole cohort to one consistent standard before a teacher reviews — which is a genuine benefit for moderation and for fairness to students.
The policy should set the conditions that keep that benefit real. It should require that teachers, not the tool, define the rubric and exam-board criteria; that results are reviewed before use; and that the school monitors how the tools are being used across departments so practice stays even. Where a school is building this into its wider workload and assessment strategy, it sits naturally alongside the planning that SLT do when costing tools for the year ahead — the considerations in our guide on planning the 2026-27 budget map closely onto the policy decisions here.
Keep It Auditable and Reviewed
Finally, a policy that no one can evidence is a policy that does not really exist. Keep the document short, dated, and owned by a named SLT member, and give it a review cycle — annually is sensible given how fast the tools change. Specify which tools are approved, how staff request access, and how the school records that the teacher-sign-off requirement is being met in practice. For governors and for an Ofsted conversation about workload and assessment, being able to point to a current, owned, reviewed AI marking policy is worth far more than the policy's exact wording.
None of this needs to be elaborate. A two-page document that defines scope, fixes teacher judgment as non-negotiable, sets the data-protection bar, names the approved tool, and commits to an annual review covers everything that actually matters. The schools that handle AI marking well are not the ones with the longest policies; they are the ones whose staff know clearly what they are allowed to do and trust that the institution has thought it through.
Build Your Policy on a Tool That Already Fits It
The easiest AI marking policy to write is one whose requirements your chosen tool already satisfies. GradeOrbit keeps every AI-generated mark under teacher review, marks against your departments' own criteria, scans handwritten and physical work, runs school accounts on a single shared credit pool with a clean invoice for every payment, and ships with a standard Data Processing Agreement you can put in front of your DPO this week. That turns most of the policy from an aspiration into a description of how the tool already behaves.
If you are the SLT lead drafting an AI marking policy and want to see how the school journey handles teacher sign-off, data protection, and retention in practice, visit our homepage and we will walk you through it properly — straight answers on judgment, data, and accountability, not a sales deck.