Module 1 · Start safe
AI does not decide grades or a student's fate
In short
AI may help prepare questions, rubrics and draft feedback. Grades, promotion, sanctions and recommendations are decided by the teacher who reads the student's work. For students under 13, UNESCO advises against independent use of AI tools in class.
After this lesson you can
- Tell AI help that prepares material apart from decisions that stay with the teacher
- Explain why grades, promotion and sanctions must not be decided by AI
- Apply age limits and supervision when students use AI tools in class
Terms in this lesson
Understand
Marking forty essays over a weekend is exhausting, and the temptation to let a machine grade them is strong. But a grade is not just a number. It decides remedial work, rankings, report cards, sometimes scholarships or promotion. It is also the teacher's message to the student about what they have mastered. A decision like that is called an automated decision when only a machine makes it without a person weighing it, and for children the impact is large.
A language model grades in a way you cannot justify to a student or parent. It can give the same essay different grades when asked twice. It is swayed by length, polished style or certain words, not only content. It can make up convincing reasons for a grade that is really random. And it does not know what happens in your class: who worked hard, who just moved schools, who was asked to write in a particular way.
Besides, grading means reading student work. If essays with their authors' names are pasted into an AI tool, you have broken the rule of the previous lesson.
So the line is this. AI may help before and after assessment: drafting rubrics, writing sample answers at each level, tidying the language of feedback whose content you have decided, or summarising error patterns from your own notes. Reading the work, giving the grade and deciding next steps stay your job. You know this principle from Level 1 as human in the loop.
What about students using AI themselves? UNESCO's guidance on generative AI in education and research, published on 7 September 2023, proposes an age limit of 13 for the use of AI tools in the classroom and calls for teacher training. Many providers also set age limits in their terms of service. For younger students, use AI as a teacher's tool: you type, you read the output, then you choose what to show the class. For older students, use is still supervised, with clear rules on what data they may type.
Another decision often forgotten is the recommendation: scholarship letters, advice on study tracks, reports to parents on a child's progress. Their content is a judgement about a person's ability. You may ask AI to tidy the sentence structure of points you wrote yourself, but the points must come from your own observation. A recommendation written by a machine from grades alone sounds generic and can include praise or weaknesses you never saw.
Before
The prompt
Here are 35 class 10 student essays on climate change. Give each a score of 0-100 and the reasons.
What came back
Bad example output: the model invented these figures and references, do not use them: Essay 1: 82, strong argument. Essay 2: 64, lacks data. Essay 3: 91, very good... (generic reasons, no reference to any criteria, and scores change when asked again)
After
The prompt
Goal: I teach class 10 science and need a draft rubric for an argumentative essay on climate change. Reader: myself and colleagues who help grade. Material: learning goals: state a claim, use evidence from class readings, explain how the evidence supports the claim, write a conclusion. Limits: do not grade any essay; rubric only. Do not assess neatness or length. Shape: a table of four criteria by three levels, with a short evidence example for each cell. Check: flag criteria that are hard to tell apart between levels.
What came back
Draft rubric for the teacher to review: Claim, beginning: claim present but vague; proficient: clear claim; advanced: clear, qualified claim. Evidence, beginning: no source; proficient: one piece of evidence from the readings; advanced: two or more, compared... Check note: the evidence-to-claim criterion may be hard to separate between proficient and advanced; add examples.
What changed
- The first request hands grading to a machine and sends student work. Its output cannot be justified.
- The second request asks for an assessment aid, not grades. The teacher still reads and grades.
- The limit on neatness and length keeps the rubric on the learning goals.
Common mistake
- Mistake
- Using an AI grade as a provisional grade with the intention of checking it later.
- Why it hurts
- Provisional grades quickly become final: they go into the list, students see them, they decide remedial work. Checking forty grades afterwards is almost as heavy as grading, so it often does not happen.
- Do this instead
- Use AI for the rubric and sample answers before grading. When grading, you read and give the number.
Try
Sorting card
Decide whether AI may help with this task or the teacher must decide it.
Writing the content of a scholarship recommendation
Drafting a rubric from the learning goals
Imposing a sanction for suspected cheating
Tidying the language of feedback whose content you wrote
Deciding who is not promoted
Writing sample answers for each rubric level
6 cards left
A grade 4 class is learning about weather. You want to use AI. Which option fits best?
Take home
The decision line in your class
- I set grades and report cards after reading the students' work myself
- I use AI only for rubrics, sample answers and tidying language
- Promotion, sanctions and recommendations are decided by authorised people
- Students under 13 do not use AI tools on their own
- Older students' AI use is supervised with clear data rules
What went wrong for us
A product claim checked against the code
- Context
- Lessons and cases for several professional tracks were written with examples from products we run.
- What happened
- The reviewer read the code of the products mentioned and found case texts saying a product did something its code does not do.
- Impact
- Had it been released, the material would have taught something untrue about our own products.
- Root cause
- The drafts described products from names and summaries, not from the source.
- What we did
- Each false sentence was rewritten to say what the code really does, and every other product claim in those tracks was checked against the code.
- The guard we keep now
- Our rule: every claim about what a product does is checked against its source before it goes into material, or written generically without a claim.
The lessonA convincing sentence still needs proof from its source.
What it means for your workWhen AI writes facts about a tool, a rule or your school, check the original document before passing them to students or parents.
Sources
- UNESCO: Governments must quickly regulate Generative AI in schools (7 September 2023) (accessed 26 October 2026)