Bias in the Gradebook

Every school has historical data — years of grades, discipline records, placement decisions. AI tools love historical data. That's the problem.

If past placement decisions favored certain groups of students — and in many districts, they did — then an AI trained on that history learns the pattern, not the justice. It will recommend advanced classes for the same kinds of kids who always got recommended, and it will do it with mathematical confidence. Bias laundered through software looks like objectivity.

You don't need a data science team to check for this. You need one habit: disaggregate. When an AI tool makes recommendations — placements, reading levels, intervention flags — break the results down by student group. If one group is consistently sorted downward, you have a question to investigate, not a tool to trust blindly.

Then ask the vendor the uncomfortable question: what data was this trained on, and did you test it for disparate impact across groups? Get the answer in writing. "Our model is fair" is marketing. Test results are evidence.

None of this means rejecting AI in schools. It means remembering that a tool trained on the past will reproduce the past unless someone deliberately corrects it. Fairness isn't a feature you buy. It's a practice you maintain — one disaggregated report at a time.

I explain frameworks; I don't give legal counsel.