Fair to every student

Lesson material shapes how students see themselves. Word problems that always show fathers at work and mothers in the kitchen, scientist examples who all come from one region, or images showing only one skin colour all teach something beyond the subject. Students who never see themselves in the examples conclude that the subject is not for them.

AI adds to this risk in two ways. First, model bias: models learn from huge amounts of human text, including its prejudices, so unasked they tend to give the names, jobs and roles most common in their training data. Ask for ten word problems and the doctor is likely a man, the nurse a woman, and every character lives in a big city. Second, speed: a teacher who writes five questions may notice the pattern; a teacher who receives fifty at once from a machine rarely reads them all closely.

What is often missed are stand-ins, details that do not name a protected trait but carry it. A question about holidays abroad or expensive gadgets assumes every student is well off. A question using regional terms makes it harder for students from elsewhere, when what is measured is maths. Feedback that judges accent or handwriting judges what was not taught.

Fairness applies to language too. Students with visual impairments or dyslexia need material that assistive tools can read, short sentences and clear fonts. AI can help simplify language, but you make sure the content does not shrink for that student.

Checking is not complicated. Read the AI output with three questions. Who appears in the examples, and who never does? Is any question easier for students from a particular background, when that is unrelated to the learning goal? Does any sentence judge students on something from the protected traits list? If unsure, ask a colleague from a different background to read the first ten questions.

AI-made pictures and illustrations need the same care. Ask for a picture of a teacher, scientist or athlete, and an image model tends to produce the faces, clothes and bodies it has seen most. If you use illustrations on worksheets or slides, look at the whole set together, not one by one. Ask whether students with a headscarf, a wheelchair or a different skin colour ever appear, and whether any character is drawn as funny or foolish because of a physical trait.

Watch example names too. Names hint at religion, ethnicity and origin. Models often use names from one region only, or give certain names to the character who makes the mistake. Replace names with roles or use a mix from many regions, and never use names of students in your class as characters, even meaning it warmly. A student whose name is used for the character who cannot count may be teased by classmates for days.

Finally, remember that fair does not mean every student gets exactly the same material. Students learning Indonesian as a second language, or who need large print, may get an adapted version, as long as the learning goal is the same and they are assessed on the same criteria.