AI principles in education: how Inclusion Studio uses AI

AI can save staff time on writing, but learner support records are personal and sensitive. These AI principles for learner support in education set the limits we build to. They apply to the production application and to anything we pilot.

Do I have to use AI to use Inclusion Studio?

No. The full support journey, from hearing the learner to agreeing a plan, recording delivery and reviewing what helped, works without AI. An organisation can switch AI off and lose nothing essential.

What is AI used for?

AI is used to suggest drafts, for example a first draft of a review summary from notes a member of staff has already written. Every draft:

• is clearly labelled as an AI draft;
• is shown to a person, who edits, accepts or discards it;
• is saved only when a person chooses to save it;
• is built only from records the person is already allowed to see.

See AI drafts for how this appears in the product.

What will AI never do?

AI in Inclusion Studio will never:

• diagnose a learner or suggest a diagnosis;
• refuse, reduce or ration support;
• make decisions about a learner;
• share information outside the organisation;
• invent evidence that support was delivered;
• act as a therapeutic chatbot, or an open-ended chat, for children or learners;
• read or label a learner’s emotions, or infer a disability;
• be trained on case records.

Decisions about support, permissions and recorded actions are made by people and by ordinary, predictable software rules, not by an AI model.

Has an AI provider been chosen?

Yes, for testing. Gemini on Google Cloud Vertex AI is the selected AI provider. It has been exercised with synthetic data only, in a preview environment, at an EU multi-region endpoint accepted for preview use. Where real learner data would be processed is not decided, the live application uses a mock, and no real learner data has been sent to any AI provider. The prototype uses a fixed sample draft and calls no external AI service.

Before any real record reaches an AI model, we will decide and publish:

• which models are used;
• where prompts and responses are processed;
• how long the provider keeps them, and whether they are ever used for training;
• what happens if the provider is unavailable.

There will be no silent fallback to a different provider or region. If the approved route is unavailable, the AI feature stops and the rest of the product carries on. Requests go directly to Vertex AI rather than through a hosted AI gateway. Vercel’s AI Gateway, for example, is a separate hosted service: its documentation says requests are routed globally unless a region is pinned, and the regions offered are currently US and EU. Using any gateway would be a separate, documented decision.

How is AI kept inside the permission model?

AI retrieval uses the same server-side permission checks as the rest of the product. If a tutor cannot open a document, AI cannot read it on their behalf. Uploaded files stay unavailable to AI until they have passed a malware scan. Each AI feature will have a kill switch so an organisation, or we, can turn it off quickly.

How will AI quality be checked?

Before AI is offered on real data, we will test drafts against synthetic examples for accuracy, tone, respectful language and whether they add anything not in the source notes. We will publish what we tested and the limits we found. Staff remain responsible for what they save, so drafts should be read with the same care as anything else in a learner’s record.

Where can I learn more about AI principles in education?

Our guide to using AI safely in learner support covers questions to ask any supplier. The ICO’s guidance on AI and data protection and the Department for Education’s guidance on generative AI in education are linked in the sources below. For data locations, see where data is processed.