Before You Sign: What AI Marking Really Asks of School Leaders
Photo: WorldTraveler4, CC BY 4.0, via Wikimedia Commons
The pitch is always the same, and it is always seductive: upload the work, receive differentiated comments in ninety seconds, save each teacher several hours a week. In a term where recruitment is tight and workload is the thing governors ask about first, that pitch lands.
But AI-assisted marking is not a stationery purchase. It touches pupil data, professional judgement, exam integrity and the relationship between a child and the person who reads their work. None of those things can be delegated to a vendor's terms and conditions. So before a school signs anything, here is what we would put under the microscope.
Start with the job, not the product
Leadership teams that get this right tend to begin with a question about practice rather than procurement: which marking is genuinely low-value, and which is the thing that makes us a good school?
Most teachers will happily hand over the spelling corrections on a set of Year 8 explanations. Almost none want to hand over the comment that notices a pupil has finally understood fractions. A tool that is clear about which of those it is doing — and stays in its lane — is far easier to defend to staff, parents and governors than one that claims to do both.
Write down the specific task before you see a demo. If a supplier cannot describe how their product does that task, you have your answer.
The data question nobody wants to ask
Under UK GDPR, pupil work is personal data, and often special category data if it reveals anything about health, religion or SEN. That means a written data processing agreement, a documented lawful basis, and a data protection impact assessment for anything at scale — not a PDF of a privacy notice stapled to an order form.
Ask explicitly:
- Where is the data processed, and who are the sub-processors? "The cloud" is not an answer.
- Is pupil work used to train or improve models? Get a contractual no, in writing, with a deletion schedule.
- How long is anything retained, and can we export it? Your exit plan matters as much as your onboarding.
- What happens on termination? Certificates of deletion, not reassurances.
- Does the tool make automated decisions? If a grade or a report is generated without meaningful human review, you are in Article 22 territory and your DPO needs to be in the room.
If a supplier's DPO cannot speak to those points in a single call, that is not a red flag on its own — but it is a reason to slow down.
Bias is a marking problem before it is a tech problem
Markers have always been inconsistent. The difference is that human inconsistency can be moderated, appealed and discussed in a department meeting. A model's inconsistency arrives wearing the costume of objectivity.
Ask to test it. Take twenty scripts already marked by your own staff, including a couple where two teachers disagreed, and run them through. Look for patterns: does the tool penalise shorter answers, certain dialects, EAL phrasing, or handwriting-to-text conversion errors? Does it reward verbosity? Does it detect when a pupil has answered a different question entirely?
Two teachers disagreeing about a script is a professional conversation. A model systematically marking down a group of pupils is a safeguarding and equality issue, and you need evidence before you scale.
Keep the teacher visibly in the loop
Students are sharper than we give them credit for. They will spot generic feedback, and they will trust it less. The schools handling this well tend to enforce a few rules: no AI-generated comment goes home unreviewed; children know when a machine has drafted something and why; and pupils can always ask a human.
There is also the exam question. For coursework and non-exam assessment, the current JCQ guidance on authenticity and AI use applies as much to a teacher's marking tool as to a pupil's draft. Make sure your exams officer has read the contract too.
A short procurement checklist
- A completed DPIA, signed off by the DPO, before any pupil work is uploaded.
- A DPA naming sub-processors and confirming no training on your data.
- Evidence from a trial on your own scripts, including edge cases.
- A documented human-review step written into the marking policy.
- A named owner in school — normally a deputy or curriculum lead, not an IT technician.
- Training that covers what the tool is bad at, not just what it is good at.
- A break clause and a data export path.
- Cost modelled per pupil, not per teacher seat, so you can see what happens if you scale.
What good looks like six months in
Success is not "we saved 400 hours". It is quieter: teachers say feedback is more consistent across parallel classes; pupils can explain what they need to do next; the marking policy still describes something recognisable; and nobody in the staffroom is quietly re-marking everything the tool produced.
Failures tend to look the same too. Pilots that were approved by IT alone. Training that was a webinar. A tool bought because a neighbouring MAT had it. That is not a critique of the technology — it is a reminder that AI marking is a leadership decision about trust, and trust is not in the licence.