How AI Can Support Teachers Without Replacing Their Professional Judgement

Ask an AI tool for a Basic 6 Science worksheet on the water cycle and you will have one in less than a minute. It will be neatly organised, confidently written and, very often, mostly right. That last phrase is the problem. Teaching is a long chain of judgements: what to teach next, how to explain it to this class, which learner needs extra help, whether an answer shows understanding or lucky guessing. AI can produce material, but it cannot make those judgements, because it does not know your learners, your scheme of learning or your community. This article is about keeping those judgements where they belong.

What professional judgement means in teaching

Ghana's National Teachers' Standards describe the teacher's role in three domains: Professional Values and Attitudes, Professional Knowledge and Professional Practice (National Teaching Council, 2017). Professional judgement draws on all three at once. It combines knowledge of the subject and the curriculum, knowledge of particular learners, and values about fairness and care.

An AI tool brings only a version of the first. It has absorbed a great deal of general text, but it has never met your class. It does not know that half of them struggled with fractions last term, that one learner has just lost a parent, or that the example it has chosen means nothing in your community.

This view is increasingly shared in Ghanaian teacher education. At the National Conference of Principals of Colleges of Education, reported in September 2026, a speaker argued that integrating AI into teacher education must align with the National Teachers' Standards and the National Teacher Education Curriculum Framework, warning against letting technology drive pedagogy (Rainbow Radio, 2026). Internationally, UNESCO's AI competency framework for teachers begins with a "human-centred mindset" as one of its five dimensions (Miao and Cukurova, 2024).

What AI can usefully do for teachers, such as drafting lesson ideas, questions and adapted texts, is covered in an earlier article on AI in Ghanaian basic schools. This article is about what happens next: checking the output.

Where AI output goes wrong

Confident errors

Generative AI produces fluent text by predicting likely words, not by checking facts. It can state something false in the same confident tone as something true. Researchers call this "hallucination".

A clear illustration comes from a 2023 study that asked ChatGPT to write short literature reviews and then checked the 636 references it produced. Of the references produced by GPT-3.5, 55% were fabricated: the works did not exist. For GPT-4 the figure was 18%. Even among the real references, many contained substantive errors (Walters and Wilder, 2023). Newer models perform better on many tasks, but the lesson stands: fluency is not evidence of accuracy.

In a classroom, confident errors take familiar forms: an answer key with the wrong option marked, a calculation that does not add up, a scientific explanation that is almost right, or a date or name attached to the wrong event. Because the material comes from the teacher, learners will trust it.

Missing context

AI tools learn mostly from text produced outside Ghana. Their default examples may involve snow, dollars, baseball or names that Ghanaian learners never encounter. Prices may be in the wrong currency. The language may be pitched at native English speakers rather than at Basic 4 learners for whom English is a second or third language. Content may be drawn from another country's curriculum, at a different level, or in a different order from the scheme of learning.

None of these is a disaster on its own, but each makes the material a little harder to learn from. Together, they can make a lesson feel foreign to the very learners it is meant to serve.

Fairness

AI systems reproduce patterns in the text they learned from, including stereotypes. A worksheet might always make the doctor a man and the nurse a woman, give every successful character a name from one ethnic group, or describe some regions only in terms of poverty. These patterns are easy to miss in a single worksheet and harmful when they repeat across a term.

Material that looks right but teaches the wrong thing

An AI-generated activity can be attractive and still miss the point. A question may test recall when the indicator asks learners to apply or explain. An activity may be fun but unconnected to the content standard. Only someone who knows the curriculum, and the class, can see the gap.

Gradual over-reliance

The subtlest risk is to the teacher. Writing good questions, explaining ideas clearly and anticipating misconceptions are skills that grow with practice. A teacher who hands all of that work to a tool may find, over time, that the skills have weakened, along with the ability to spot the tool's mistakes.

Five questions before anything reaches learners

Before using any AI-generated material, run through five questions:

1. Is it correct? Check every fact, answer and calculation yourself. Do the questions before your learners do.

2. Is it aligned? Does it serve the indicator in your scheme of learning, at the right level of thinking?

3. Is it right for this class? Is the language accessible? Are the examples familiar? Does it build on what your learners already know?

4. Is it fair? Look at names, roles, places and pictures. Is anyone stereotyped or left out?

5. Would I put my name to it? If a parent, head teacher or circuit supervisor asked about it, could you defend every part?

If the answer to any question is no, edit the material or discard it. Keep a note of what you changed. Those changes are your professional judgement made visible, and over time they show you what the tool does well and badly.

Decisions that should stay with the teacher

Some decisions should not be delegated to AI at all, even as a first draft:

• Final marks, grades and promotion decisions about individual learners.

• Comments on a learner's character, behaviour or ability in reports. AI might suggest general phrasing, but the judgement about a particular child must be the teacher's.

• Identifying special educational needs or labelling learners in any way.

• Disciplinary decisions.

• Sensitive communication with families, such as concerns about a learner's welfare.

These decisions depend on knowing the learner, carry real consequences for a child, and involve personal information that should not be typed into a public AI tool. A later article in this series looks at protecting school data in more detail.

Adopting AI responsibly as a school

Individual care matters, but a shared approach is stronger.

• Agree on what AI may and may not be used for, using the list above as a starting point.

• Review AI-generated materials together in staff professional learning sessions. Colleagues spot different problems, and the whole staff learns which tools and requests work.

• Be open about use. Teachers should feel able to say "I drafted this with AI and then checked it" without embarrassment, and school leaders should value the checking as much as the speed.

• Teach learners to question AI. Older learners who meet these tools outside school need to learn that AI can be wrong, and how to check it.

• Keep it tied to the standards. The question for any new tool is whether it helps teachers meet the National Teachers' Standards better, not whether it is new.

Conclusion

AI does not replace a teacher's judgement. What it can do is give that judgement more to work with, more quickly. A teacher who uses AI well is not someone who accepts its output, nor someone who refuses to touch it. It is someone who treats every AI draft the way an experienced head of department treats a new colleague's first worksheet: with interest, with care, and with a red pen ready.

Try this: The next time you use AI to prepare material, run the five questions above and write down every change you make before using it. After a month, read your notes. They will tell you more about the tool, and about your own expertise, than any advertisement.

References

• Miao, F. and Cukurova, M. (2024). AI competency framework for teachers. UNESCO. https://doi.org/10.54675/ZJTE2084

• National Teaching Council (2017). National Teachers' Standards for Ghana: Guidelines. Ministry of Education, Ghana. PDF

• Rainbow Radio (2 September 2026). Ghanaian institutions urged to integrate AI skills directly into teacher training curricula. rainbowradioonline.com

• Walters, W. H. and Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5