AI in Education: Opportunities, Limits and Responsible Use

Use AI in education for a bounded task: draft a question, suggest a variation or propose an explanation, then check it against the learning goal and an approved source. Start with material that contains no personal student data and keep a manual alternative. This guide explains the benefits and limits, works through a numerical error, and shows how to differentiate an exercise without quietly changing what it assesses.

What AI in education actually brings

AI can produce a first draft of comprehension questions, alternative wording or practice items. The useful time saving is the difference between drafting manually and checking the generated version, including any repair. If checking an invented example takes longer than writing a correct one, generating more of it does not help. Try one bounded activity before building a recurring workflow around the tool.

A second use is proposing differentiated practice: shorter wording, a worked scaffold or an extension. The teacher still needs to check whether the variation changes the learning goal. Removing a time conversion might help a learner practise division, for example, while making that version unsuitable for assessing independent conversion. The worked comparison below makes that trade-off explicit.

AI can also suggest contexts and question sequences when a teacher needs alternatives. Ask for a small set tied to the taught material, then discard ideas that require unsupported facts or unfamiliar vocabulary. The useful output is a reviewable choice of drafts, rather than a claim that the activity is ready for the class.

Tools in the edtech space increasingly bake this in. In Skolina, for example, AI can draft Quiz questions from a topic, a course document or a PDF, which a teacher then reviews and edits before anything reaches a Student. If you want a practical, step-by-step walkthrough of that specific workflow, our guide on how to generate a quiz with AI covers it end to end. Review the complete question, key and feedback before sharing the draft.

Where AI falls short

For every opportunity, there is a real limit, and pretending otherwise does students a disservice.

Hallucinations. Generative AI produces fluent, confident text that can be simply wrong: an invented date, a misattributed quote, a plausible-sounding explanation that collapses under scrutiny. In an educational setting, where the whole point is accuracy, this is not a minor quirk. It is the single strongest argument for human review of anything AI produces before it reaches learners.

Bias. Models learn from large bodies of existing text, and they reproduce the assumptions and gaps in that material. That can show up as skewed examples, narrow cultural references, or uneven treatment of a topic. A teacher’s judgement is what catches this.

Data questions. Many AI features run on third-party infrastructure. In Skolina, account data is hosted in the EU, but the AI generation itself runs on OpenAI’s models, which means that particular processing happens outside the EU. That is a meaningful distinction worth being honest about, especially in a school context. Anyone weighing up these tools should understand where pupil information travels; our overview of student data privacy and GDPR for quiz tools digs into what to check before adopting any platform.

The cheating question. If AI can write an essay or answer a Quiz, students can use it to shortcut the very thinking an assessment is meant to measure. There is no purely technical fix for this. Some platforms offer passive signals, such as detecting when a student switches browser tab or pastes a large block of text during an online Quiz. These are honest, lightweight hints, not surveillance: there is no webcam monitoring and no forced full-screen lockdown, and they should never be sold as such. The deeper answer lies in assessment design, not in policing.

Worked decision: keep, correct or stop an AI draft

Imagine a fictional teacher who wants a short question about average speed from an approved, name-free course note. The intended task is to convert 20 minutes to hours and find the speed for 8 km in 20 minutes. A constructed faulty draft says “8 ÷ 20 = 0.4 km/h”. This is an example written for this article, not a reported test of a model.

Check the units against the lesson: 20 minutes is 1/3 hour, so the correct speed is 8 ÷ (1/3) = 24 km/h. The draft’s 0.4 is kilometres per minute. Rewrite the key and explanation, then test a parallel item yourself. If you cannot verify the source, arithmetic or target reading level, discard that draft and write a simpler question manually.

UNESCO’s guidance on generative AI in education places human agency, validation and privacy at the centre of responsible use. That supports checking the activity; it does not endorse this particular product or prove the lesson will improve. After the quiz, ask students to explain the conversion rather than treating a correct click as the whole outcome.

Differentiate an AI exercise without changing the learning goal

A simpler-looking version is not necessarily an equivalent assessment. Continue the fictional speed lesson above: the learning goal combines choosing distance ÷ time with converting minutes to hours. Before requesting “an easier version”, identify which of those demands you intend to keep. Otherwise the generator may remove the very step you need to observe.

Read the draft versions side by side. Check that each retains the units, constant-speed condition and intended evidence. Reject a “reading support” version that quietly supplies the answer, adds an unfamiliar concept or changes the context so much that the calculation becomes harder. A scaffold can be useful instruction while still being unsuitable for comparing independent performance with the original question.

After learners try the supported version, remove only the support you want to check and use a new example. Ask for one conversion step as well as the number. Record whether the remaining difficulty concerns reading, conversion or division; do not infer a permanent ability level from these four constructed prompts. This gives AI a bounded role: proposing alternatives that the teacher checks against a stable objective.

Principles for responsible use

If AI is here to stay, the useful question is not whether to use it but how. A few principles travel well across subjects and age groups.

Keep a human in the loop

Treat every AI output as a draft that an expert checks. This is not a temporary precaution until the technology improves; it is the operating model. The teacher remains accountable for accuracy, tone and appropriateness, and that responsibility cannot be delegated to a model.

Be transparent

Tell students and colleagues when and how AI was used. Explain which parts were drafted with AI, which sources were checked and which decisions remain the teacher’s responsibility. Follow the school’s rules on permitted use and disclosure. Transparency also extends to data: knowing which tools process information where, and choosing accordingly.

Make reasoning visible in AI-assisted assessment

Rather than treating AI as a threat to be detected, redesign tasks so that thinking is visible and hard to outsource. Oral defences, in-class work, process-focused assignments, personalised reflection and well-built formative checks all shift the emphasis from a final answer to the reasoning behind it. This is where pedagogy does the heavy lifting; our piece on formative assessment that works explores concrete ways to build checks that reveal understanding rather than recall alone.

Anchor it in what we know about learning

Choose a learning activity before choosing an AI feature. Retrieval practice asks the learner to recall; spaced review revisits an objective later; Bloom’s taxonomy helps describe the intended cognitive demand. These are different design choices. A generator may help prepare prompts for them, but it does not establish that the prompts are accurate or that the resulting activity improves learning.

Before adopting an AI activity, use a five-question check:

• Purpose: what learning problem does the tool solve?
• Evidence: how will you know the activity improved understanding?
• Accuracy: who checks facts, answer keys and reading level?
• Data: what will be sent to the provider, and is any of it personal or sensitive?
• Fallback: can the lesson still work if the tool is unavailable or unsuitable for a Student?

This is also a useful way to compare products. Look past the headline feature list, read the data information and check the current Skolina plans and limits before deciding whether the workflow fits your setting.

Evidence and guidance behind this article

UNESCO’s guidance for generative AI in education and research calls for a human-centred approach, data protection and pedagogical validation. Its AI competency framework for Students likewise treats ethical judgement and human agency as core competencies. These documents support a measured approach; they do not prove that every AI use improves learning, and effects still depend on the activity, learner and implementation.

Choose one AI task and define the stop condition

Start with one lesson and one decision: draft alternatives for the speed question, then keep, repair or discard each version against the checked method. Record how much preparation and verification the activity actually takes. For the learner, ask for the conversion step and test it in a new problem; a fluent generated explanation or a correct click does not answer that question by itself.

Stop the AI workflow when the source cannot be verified, the activity changes the intended skill without making that clear, or the approved data conditions are not met. A manually written question using the same objective is a complete fallback, not a failed lesson. Review the decision after the first use before expanding the activity to other classes.

If you would like to put these ideas into practice, you can create a free Skolina account and try building a Quiz, with or without AI assistance, while keeping that essential human review at the centre. Keep the checked key and a note explaining which supports you changed so the next version can be reviewed.