The AI Revision Assistant Under Scrutiny: What It Can and Cannot Do for Adult Learners

The question worth asking

Every few months a new study assistant arrives with a confident promise: a personal tutor, available at midnight, patient with repetition, indifferent to embarrassment. For adult learners returning to study after years away, the appeal is obvious. There is no waiting list, no hourly rate and no need to admit in front of another person that a concept has not yet landed.

So the argument is no longer about whether AI belongs in open learning. It plainly does. The argument concerns where the boundary sits between useful assistance and misplaced dependence - and who is drawing that boundary.

Our view, having watched how open-learning students actually use these tools, is that they are excellent at some parts of studying, mediocre at others and unsuitable for a few. That distinction matters more for learners working outside an institution than for campus students with a tutor, a library and a welfare team within walking distance.

The case in favour deserves a fair hearing

Scepticism about AI in education is often framed as a defence of rigour. Sometimes it is. Sometimes it is a defence of the status quo. The plain fact is that many adult learners go weeks without anyone reading their work. A tool that responds within seconds, at any hour and without judgement fills a genuine gap.

It also lowers the cost of being wrong. A learner who fears looking foolish in front of a tutor will ask fewer questions; a learner rehearsing an idea with a machine will ask more. That asymmetry alone is worth something.

Where the tools genuinely earn their place

Three uses stand out as reliable.

Planning. Converting a module description into a twelve-week schedule that accounts for a job, a commute and caring responsibilities is exactly the kind of administrative work a language model does well. It will not know that your bus is unreliable on Fridays, but it will produce a workable first draft in a minute.

Retrieval practice. Testing yourself is among the better-evidenced study techniques, and writing questions is tedious. A tool that turns your notes into twenty short-answer prompts - questions rather than answers, so that you perform the recall - is a legitimate efficiency gain.

First-pass feedback on structure. Asking whether an argument follows, whether each paragraph earns its place, whether the introduction promises something the conclusion delivers: this is useful commentary at a stage when no tutor would thank you for a draft.

The accountability problem

What no assistant supplies is obligation. There is no one to notice that you have gone quiet for three weeks, no one who will raise an eyebrow at a missed deadline, nobody to whom you owe an explanation. In open learning, attrition is rarely intellectual; it is structural, and it is frequently loneliness dressed up as busyness.

There is a second, subtler issue. A model is agreeable by default. Ask whether your argument works and you will usually be told that it does, followed by suggestions for refinement. That is pleasant and occasionally useful, but it is not the same as being told the premise is weak and the evidence does not support the claim. Real tutors disappoint people. Tools rarely do.

Depth, error and the confidence trap

Language models generate plausible text. Plausibility and accuracy are different properties, and the gap widens in proportion to how much you already know. If you can spot an invented reference, you are relatively safe. If you cannot, you are exposed.

The practical defence is asymmetry management: rely on the tool only where you can verify, and verify against primary sources - legislation.gov.uk for statute, GOV.UK for guidance, the awarding body's own specification for qualification rules, a peer-reviewed paper for a research claim. Numerical work, legal and regulatory detail, medical content and anything with a safety dimension should always be checked outside the tool.

A useful habit is to request sources and then confirm that at least two exist and genuinely say what was claimed. Citation fabrication remains one of the most common failure modes, and it is easily detected.

Accessibility: a tool, not an adjustment

For learners with dyslexia, attention difficulties or English as an additional language, the benefits are real: plain-English rewording, summarising, structuring notes and generating practice questions from a dense PDF.

But an AI assistant is not a reasonable adjustment. Under the Equality Act 2010, the duty to make adjustments rests with the education provider or employer, not with whichever app the learner happens to own. A learner relying on a personal subscription to compensate for provision that should have been made is absorbing a cost - and a risk - that belongs elsewhere. The tool can supplement an adjustment; it cannot discharge the duty.

There is a further caution. Tools built on narrow training data can serve some learners worse than others: regional accents, Welsh-language study, particular disability-related communication patterns. If a tool consistently misreads you, that is a reason to escalate rather than to blame yourself.

Data, privacy and safeguarding

Anything typed into a consumer AI tool should be treated as potentially retained and reviewed. Do not paste medical details, other people's personal information, confidential employer material or anything covered by a non-disclosure agreement. Under UK GDPR, personal data requires a lawful basis, and casual use rarely has one.

Where study is funded or supported by an employer, check the organisation's acceptable-use policy before deploying an assistant for work-adjacent tasks. Where learning involves children or vulnerable adults, no AI tool substitutes for a safeguarding process, and the provider's safeguarding lead remains the correct route for any concern.

Four tests before you trust an AI study assistant

The calibration test. Ask about something you already understand well within the subject. Does the answer correct you, or merely agree with you?

The citation test. Request sources for three claims and check that two exist and support what is said. Fabricated references are the clearest early warning sign.

The boundary test. Ask the tool directly what it cannot reliably help with, and whether it would recommend a human for that part. A good answer names its limits.

The accountability test. Does your workflow create a record, a deadline or an obligation? If not, keep a person or a system in that role, because the tool will not do it for you.

A division of labour that holds up

Use the assistant for drafting plans, generating practice questions, rewording dense text, checking structure and rehearsing explanations. Reserve for humans the judgements that carry consequence: assessment decisions, accreditation and credit advice, welfare and wellbeing, subject depth at the edges of your knowledge, and safeguarding.

The verdict

The study buddy does not replace the tutor. It replaces the blank page - and that is a meaningful contribution. Where open learners come unstuck is in mistaking fluency for authority and availability for accountability. Used with a clear division of labour and four simple tests, an AI assistant is a genuine asset on a self-directed pathway. Used as a substitute for the human structures that hold learning together, it is a very articulate way of studying alone.

That is our position: adopt deliberately, verify habitually, and keep a person somewhere in the loop.