Where can AI help in teaching, and where should a person decide?

Generative AI is good at first drafts of teaching work: questions written from your own materials, a suggested mark for a short written answer, a plain summary of a month’s results. A person should decide everything that reaches a learner or changes their record: which questions are published, which mark stands, what a report says and which level a learner is on. UNESCO’s guidance and Malaysia’s national AI guidelines both place that responsibility on people, not on software.

What does generative AI do well for teachers and trainers?

It removes the blank page. Three jobs come up in tuition centres, schools and training teams alike.

Drafting questions. Given a chapter, a slide deck or a procedure manual, a model can propose practice questions, answer options and short explanations in seconds. UNESCO’s Guidance for generative AI in education and research (2023) lists this among possible uses: AI “may also help teachers prepare tests and exams by offering examples of questions and rubrics for evaluation”. It files the idea under “potential but unproven uses”, and adds that human curriculum designers “need to verify the factual knowledge”.

Suggesting a mark. Given the question, a marking guide and a learner’s written answer, a model can suggest a score and a reason. The teacher starts from a suggestion, not from nothing.

Summarising progress. A model can turn a month of figures (scores, lessons completed, quizzes passed) into a few readable sentences for a teacher or trainer to correct.

In each case the output is a draft: the work moves from writing to checking, and the checking still has to happen.

Where does AI go wrong?

Wrong answers, stated confidently. A model writes fluent text whether or not it is right. The UNESCO guidance says that AI outputs “cannot be relied upon without critical evaluation” and that these models “can frequently produce inaccurate or unreliable text”. In teaching, one wrong answer key is not one mistake: it marks every learner who answered correctly as wrong.

Bias. UNESCO warns that generative AI “reproduces dominant worldviews in its outputs” and asks institutions to test tools for bias, gender bias in particular. In a lesson this can show up as examples or assumptions that suit one group of learners and not another, or as language that is “culturally insensitive or contextually inaccurate”, in the guidance’s words.

Privacy. UNESCO notes that using generative AI “almost always involves users sharing their data with the GenAI provider”. If a learner’s name, contact details or a teacher’s private notes go into a prompt, they have left your organisation.

Over-reliance. UNESCO’s AI competency framework for teachers (2024) warns that over-reliance on AI “could lead to the atrophy of teachers’ essential competencies”. A teacher who stops reading the drafts stops noticing the errors.

What does the guidance say?

UNESCO, 2023: keep people accountable

The guidance takes a human-centred approach. Its recommendations include protecting data privacy, checking tools for their ethical and educational fit before adopting them, and telling learners what data AI tools may collect from them and how it is used. One line is blunt: “prevent ceding human accountability to GenAI systems when making high-stakes decisions”. For AI assistants that answer learners, it says institutions “need to guarantee human supervision” of the answers. It proposes 13 as the minimum age for a learner’s independent conversations with AI chat platforms. UNESCO also publishes the guidance in Bahasa Melayu and in Chinese.

UNESCO, 2024: the teacher stays responsible

The framework sets out 15 competencies across five dimensions (a human-centred mindset, the ethics of AI, AI foundations and applications, AI pedagogy, and AI for professional learning) at three levels: Acquire, Deepen and Create. On responsibility it is direct: “Teachers should remain accountable for pedagogical decisions in the use of AI”, and AI tools “should never be designed to replace the legitimate accountability of teachers in education”. The same principle fits a corporate trainer or an assessor.

Malaysia, 2024: seven principles, with people in the loop

Malaysia’s National Guidelines on AI Governance & Ethics, published by the Ministry of Science, Technology and Innovation and launched in September 2024, are voluntary. They set seven principles: fairness; reliability, safety and control; privacy and security; inclusiveness; transparency; accountability; and the pursuit of human benefit and happiness. Three of them bear directly on teaching:

• Transparency includes full disclosure that an AI system is being used in decision making, and the ability to challenge decisions an AI system made.
• Privacy and security asks developers to obtain consent, where necessary, before using or disclosing personal data for AI, and to build privacy in from the start.
• Human benefit and happiness describes human oversight, including “human-in-the-loop”: a person able to step in at every decision the system makes.

What rules can an organisation adopt?

These six rules follow from the guidance above.

1. AI drafts, a person approves. Nothing a model wrote reaches a learner, a mark or a report until a named teacher or trainer has read it and accepted it.
2. Keep the AI’s suggestion beside the person’s decision. Store both, and never overwrite the AI’s version with the correction. You can then see how often people change what the AI proposed, and explain any mark a learner challenges.
3. Never send a learner’s personal details to a model. To suggest a mark, a model needs the question, the marking guide and the answer. It does not need a name, an identity number, a phone number, a date of birth or a teacher’s notes.
4. Tell learners, and parents or managers, when AI was involved. Say which tasks use AI, what it receives and who checks its work, in your privacy notice and in plain words.
5. Keep high-stakes results human. Placement, moving up a level, certificates and anything that affects a job or a school place are decided by a person. AI can prepare; it does not decide.
6. Test before you adopt, then keep testing. Try an AI feature on your own materials, in every language you teach, and look for wrong answers and bias. Afterwards, sample its drafts regularly, even when it seems reliable.

If you teach children, add a seventh: no child under 13 chats with a general AI tool on their own, in line with UNESCO’s suggested minimum.

How does AI Teacher handle this?

AI drafts questions from the teacher’s own uploaded materials, and by default every draft stays unpublished until a teacher reads and publishes it. The teacher’s edits are kept beside the AI’s original. From 1 November 2026 an organisation can opt in to software publishing some practice questions after automated checks; those never appear in a placement test, a unit quiz, a level-up test or an assessment behind a certificate.

Where an organisation switches AI marking on, AI suggests a mark for each short written answer. The answer stays unmarked until the teacher confirms it or gives their own, the AI’s suggestion is kept beside the teacher’s decision, and learners see only the teacher’s mark. Levels and skill scores are calculated by fixed rules, not by a model. A monthly report reaches the learner, or their parent or manager, only once the teacher has written its comment; an AI summary, where it is switched on, is a draft for the teacher and is not on the report they read.

The AI model providers are outside Malaysia, and each job is sent only what it needs. Apart from what a learner typed in an answer, no job is sent a learner’s name or any identifier, an email address, a phone number, a date of birth, an address, a guardian’s contact details, a photograph, or anything a teacher wrote about a learner. The features page shows where AI is used, role by role.