Using AI safely in education and learner support

Generative AI is already in FE. Staff use it to draft emails, summarise meetings and rewrite documents in plain English. In learner support, the time savings are tempting: a support plan drafted from conversation notes, a review summary in minutes, an easy-read version of a letter. But learner support records are among the most sensitive data a provider holds, and the people they describe are often young, disabled or vulnerable. The risks are real too.

This guide sets out where AI can help, where it should never be used, and the safeguards that make the difference. It draws on DfE guidance, the DfE product safety standards and ICO guidance. It is not legal advice, and the ICO notes that its AI guidance is under review following the Data (Use and Access) Act 2025.

What does DfE guidance say about using AI safely in education?

DfE’s guidance on generative AI in education applies to schools and colleges in England. The points most relevant to learner support:

• Personal data must be protected in line with data protection law. The guidance recommends keeping personal data out of generative AI tools, and where it is strictly necessary to use it, taking all steps to protect it.
• Professional judgement is required. Staff must use their professional judgement, and responsibility for final documents stays with the professional and their organisation, whatever tools were used.
• Outputs must be checked. Content needs critical judgement for appropriateness and accuracy, because AI can produce false information presented as fact.
• Learners’ work must not train AI without permission or a copyright exception.
• Safeguards must be in place where students use AI.

DfE also publishes product safety standards for edtech developers and suppliers, which colleges can use to assess products. They cover filtering, monitoring and reporting, security, privacy and data protection, intellectual property, design and testing, governance, and, since January 2026, cognitive development, emotional and social development, mental health and manipulation. Among other things, they say products should not use personal data for commercial purposes and should store only the minimum data necessary.

Where can AI help in learner support?

AI is most useful for low-risk drafting and formatting tasks where a knowledgeable person checks the result.

Where should AI never be used?

Draw firm lines and write them into policy.

1. No AI decisions about support. AI should not decide, approve or refuse support. Under the UK GDPR as amended by the Data (Use and Access) Act 2025, a significant decision based wholly or partly on special category data cannot be taken solely by automated processing except in limited circumstances (Article 22B). A decision is “solely” automated where there is “no meaningful human involvement” (Article 22A). Beyond the law, a learner deserves a person who is accountable for decisions about them.
2. No diagnosis. AI must never suggest or imply a diagnosis. Learner support responds to need, and diagnosis is for qualified professionals.
3. No safeguarding judgements. Use your safeguarding procedures and trained staff.
4. No sharing outside the organisation by AI. A person decides what is shared, with whom, and on what basis.
5. No unapproved tools. Staff should not paste learner information into personal or consumer AI accounts.

What makes human review meaningful?

Meaningful review is more than a click. A reviewer should:

• have the knowledge and authority to change or reject the draft
• compare the draft with the source, especially the learner’s own words
• check for invented details, such as an adjustment nobody agreed or a condition nobody mentioned
• check tone: respectful, person-first or identity-first as the learner prefers, no deficit language
• edit the draft, rather than accept it as is by default
• be the named author of what is saved

Watch for automation bias, the tendency to trust a confident-sounding output. It grows when staff are busy. Building review into the workflow, with the AI draft clearly labelled until a person saves it, helps.

Fictional example. A learning support coordinator records a conversation with Aisha, 17, who says she finds group work stressful and wants to know the topic of each session in advance. An AI tool drafts a support plan section. The draft says Aisha “has social anxiety” and recommends “excusing her from group work”. The coordinator notices that Aisha never mentioned a diagnosis, and that she asked for advance notice, not to be excused. She rewrites the section in Aisha’s words and saves it as her own work. The draft saved time on structure; the human review stopped two errors reaching Aisha’s record.

How do you keep learner data safe when using AI?

Treat an AI feature like any other processing of special category data, with extra attention to where the data goes.

• Do a DPIA first. The ICO lists innovative technology such as AI among the factors that point to a DPIA. See DPIA for learner support software.
• Know the AI provider. Which company processes the data, where, and under what contract?
• No training on your data. Contracts should prevent the supplier and AI provider using learner data to train models.
• Minimise inputs. Send only what the task needs. Remove names and identifiers where possible.
• Retention. Prompts and outputs should not be kept by the AI provider longer than necessary.
• Off by default. The ICO’s 2026 edtech audits found AI features sometimes introduced without safeguards or switched on by default. Choose tools where AI is optional and switched on deliberately.
• Transparency. Tell learners in your privacy information where AI is used.

How do you check AI for bias?

Ask how outputs differ across groups of learners, and test it yourself. Bias can appear in language, such as describing some learners in more negative terms, in suggestions, such as recommending different support for similar needs, or in omissions. Practical checks:

• run the same scenario with different names, ages or backgrounds and compare outputs
• ask disabled learners and staff to review sample outputs
• monitor a sample of saved records for patterns over time
• give staff an easy way to report problematic outputs

The ICO’s guidance on AI and data protection has a chapter on fairness that covers bias and discrimination.

How should you introduce AI to a learner support team?

Start small, measure honestly and be ready to stop. A sensible approach:

1. Pick one low-risk task. For example, rewriting standard letters in plain English, using no personal data.
2. Agree what success means. Time saved, quality of output, staff confidence, and learner feedback. Decide in advance what would make you stop.
3. Complete the DPIA before any personal data is used. Involve your DPO from the start, not at the end.
4. Train the people involved. Cover what the tool can and cannot do, how to review, automation bias, and how to report problems.
5. Ask learners. A learner panel can tell you how they would feel about AI being used to draft their plans, and what would reassure them.
6. Review after a set period. Look at a sample of outputs and saved records. Did review catch errors? Were any missed? Did anything get worse?
7. Expand only with evidence. Move to a medium-risk task only when the first one is working and the safeguards are proven.

Expect some staff to be enthusiastic and others wary. Both are useful. The enthusiasts find uses; the sceptics find problems.

What should an AI policy for learner support include?

Keep it short and practical. A one-page policy for learner support teams might cover:

1. approved tools, and a ban on unapproved ones for learner data
2. permitted tasks and prohibited uses, based on the tables above
3. review rules: who can save AI-drafted content and what they must check
4. labelling: how AI-drafted content is marked before review
5. data rules: what can and cannot be entered
6. training for staff, including automation bias
7. how learners are told, and how they can ask questions or object
8. how to report a problem, and who reviews the policy

Link it to your existing data protection, safeguarding and acceptable use policies rather than duplicating them.

How does Inclusion Studio support this process?

Inclusion Studio is learner support software being built for FE colleges and apprenticeship providers in England, currently pre-pilot with a prototype using fictional learners. Its AI drafts feature is optional and the full support journey works without it. Drafts are labelled, and a person reviews before anything is saved. AI never diagnoses, refuses support, makes decisions, shares externally, or is trained on case records. The prototype uses a sample draft. 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. Our AI principles set out these commitments in full.