A Potential Copyright Checklist for Choosing AI Assessment Tools (4/4)

This final post is deliberately practical. It’s a short checklist you can use when you’re buying or reviewing AI‑enabled assessment tools. It won’t turn you into a copyright lawyer, but it should help you spot the difference between a responsible posture and a risky one.

The new baseline for buyers

You don’t need to memorise every paragraph of the recent government reports to understand what’s changing. For education, four ideas are enough.

First, the UK has stepped back from the idea that AI developers should get a broad, automatic exception to train on any copyrighted material they can see. Instead, the emphasis is moving toward licensing and negotiated access, especially where valuable, human‑created works are concerned. Second, there is growing recognition that creators – including students – should retain meaningful control over whether their work is used in training, and on what terms. Third, transparency is no longer a “nice to have”: there is serious talk of labelling AI‑generated content and improving insight into what data has trained which models. Finally, there is a renewed focus on the value of human creativity. Purely machine‑generated content is unlikely to gain special protection, while AI‑assisted human work remains firmly within the scope of copyright.

For schools, trusts and exam boards, that adds up to a simple message. You need to start treating student work, and the judgements professionals make about it, as assets that deserve the same care and respect as any other copyrighted material. When you choose AI‑enabled tools, you are not only choosing features; you are choosing a copyright posture.

Ten questions to ask any AI assessment vendor

What follows is not a legal standard. It’s a set of questions we think are worth asking today. In each case, you are not just listening to the answer, but to how easy it was for the vendor to give you a straight one.

1. What do you use student work for, beyond running the assessment?
   Ask for a clear, plain‑language description. Do they use scripts solely to support marking, moderation and reporting, or do they also feed them into training pipelines, research projects or product development? If the answer is “we may reuse anonymised work for improvement”, press for specifics. Assessment use and training use are different categories and should be described separately.
2. Do you claim any right to use student work for training by default?
   Look at the contract and privacy wording. Is there a blanket licence that lets the vendor reuse student work for “AI development” or “benchmarking” without further agreement, or is any training use carved out and subject to a separate opt‑in? A conservative stance will treat training on student work as an exceptional, explicitly agreed‑upon step, not the default.
3. Who owns the assessment outputs and any models trained on our data?
   Clarify ownership of raw scripts, judgement data, rankings, analytics and any derived models. A healthy answer will be that learners own their work, institutions own the judgement graph and assessment outputs, and the vendor owns its platform code – with any model trained on your data clearly governed by explicit terms. Be wary of arrangements where everything that passes through the system is treated as the vendor’s IP.
4. Which models and jurisdictions see our student work?
   If the tool uses external AI services, ask which providers are involved, where they are hosted, and what those providers are allowed to do with the data. “We use a big US model but don’t really know how it handles training” is very different from “Here are the providers, regions and retention periods, and here is our commitment that your data is not used to train them”.
5. How do you separate assessment use from training use in practice?
   Even if the vendor promises not to train on student work, ask them how they enforce that distinction technically and organisationally. Do they have separate environments and pipelines? Can they show you where assessment data flows and where it doesn’t? Vague assurances are less reassuring than a simple, inspectable architecture.
6. Can we exit with our full assessment graph – and what happens to any training you’ve done?
   Imagine you stop using the tool. Can you take your scripts, judgements, scores and analytics with you in usable form? If the vendor has trained models using your data, what are your rights in relation to those models after exit? A sovereignty‑friendly answer will offer full export and avoid “IP tails” – situations where your data has permanently enriched a vendor model you have no say over.
7. How do you inform students and parents about AI and training use?
   Ask to see the vendor’s suggested wording for privacy notices, consent forms or parental communications. Do they acknowledge that student work is creative, explain clearly what AI is doing, and distinguish between assessment and training use? Or do they rely on institutions to absorb all the communication risk while they sit in the background?
8. How do you validate AI outputs against human standards?
   This is partly a quality question and partly a rights question. A responsible design will keep human judgement as the anchor: AI may help with speed or triage, but consensus among professionals – for example, through comparative judgement – remains the ground truth. Some systems formalise this as a “validation layer”, where human‑generated standards are used to check and correct AI behaviour. That approach honours the value of human expertise and makes it easier to argue that models are serving professional judgement, not replacing it.
9. How do you log and label AI involvement?
   Over time, you are likely to be asked not only “what did this AI train on?” but also “where did AI touch this piece of work?” Ask whether the system records when AI has generated, rewritten or scored content, and whether that information is visible to teachers, students or moderators. Silent substitution – AI writing or marking without any trace – is hard to square with emerging expectations of transparency.
10. Can we choose where and how our “assessment intelligence” lives?
    Finally, ask about deployment flexibility. Can you run key models in environments you control – on‑device, in your own cloud, within national or sector‑owned platforms – or are you tied to a single vendor endpoint? Systems that let institutions repurpose their own judgement data into things like rulers, and then decide where those rulers are deployed, are better aligned with a world in which assessment intelligence is seen as an institutional or sector asset.

How we’re trying to answer these questions in RM Compare

Throughout this series we’ve tried not to pretend we have everything figured out, and that applies here too. Our answers will evolve as the legal and policy picture evolves. But it may help to say, briefly, how we are thinking about these ten questions today.

Our starting point is that student work comes into RM Compare to be judged. It is processed to support assessment and moderation, not to feed a general‑purpose training pipeline. Learners remain the creators of their work; institutions retain control over the judgement graphs, rankings and analytics built on top of it. Our tenancy model and export capabilities are designed so that those assessment assets can be kept, moved and reused by licence holders, rather than disappearing into a black box.

When it comes to reuse, our first instinct is not to train opaque models, but to turn human judgements into reusable standards. Ranks become rulers that belong to the licence holders who created them. Through the wider ecosystem – 💻|Studio, 🔗|Hub, 📳|Live – those rulers can be governed and deployed across an institution’s own landscape. In that sense, we are trying to make “your work builds your intelligence” concrete, without claiming ownership of the underlying creative work.

We already use AI in places, and expect to use it more, but we try to keep human judgement as the validation layer. AI can help with speed, triage or explanation; comparative judgement and professional consensus remain the anchor. We keep assessment and training use conceptually and technically separate, and any shift toward training on student work – whether at institutional or sector level – would, in our view, need new governance, explicit agreements and clear benefit back to the people whose work is being used.

None of this is final. As the law develops, as DfE and others refine their guidance, and as the sector gains more experience with real AI deployments, we will need to adjust. But we think the direction is right: treat student work as creative property; avoid hidden training; keep institutions in control of the intelligence their judgements generate.

Where this leaves you

If there’s one thing to take away from this series, it’s that copyright and AI in assessment are not abstract legal curiosities. They are about who gets to decide how student work is used, who benefits from the models built on top of that work, and how easy it is to change course if we realise we’ve taken a wrong turn.

You don’t need to wait for perfect legislation to start asking better questions. The next time you look at an AI‑driven assessment product – including ours – ask what it does with student work, who owns the results, and how easy it would be to walk away. The answers will tell you a lot about whose future the system is really designed to serve.

Finally, we should acknowledge that these questions land on our wider family too. RM Assessment’s AI marking initiatives are subject to the same scrutiny we’re encouraging you to apply to any vendor. Inside RM, we are using similar checklists ourselves: challenging our own projects to justify how they use student work, to separate assessment and training use, and to design for exit and sovereignty from the start.