Algorithmic bias, and why we don't rank children

Education technology products often ship a familiar feature: dashboards that rank children on some metric. It is technically easy, gives administrators a feeling of measurability and demos well. In early childhood it is a serious mistake — and not only for ethical reasons, but measurement ones.

The measurement problem

Ranking assumes what is measured is genuinely comparable. Yet early childhood data is gathered in natural settings by different teachers. A class with many observation records may not hold more advanced children; it may hold a teacher who records more. Ranking cannot see that difference — it mistakes noise for signal.

The second problem is representation bias. Verbally expressive children generate more records; quiet, cooperative children are both observed less and placed lower. The system then measures visibility, not development.

The real risk: the changed gaze

The gravest effect is not in the list but in the adult who reads it. Once a child is seen as “near the bottom”, later observations are read through that expectation. Expectation changes what is seen; changed seeing changes the record. The loop feeds itself, and the child is fixed to a number.

• Institution-level pattern questions are legitimate: “in which domain are records sparse?”
• Child-level comparison is not: “which child is behind?”
• The meaningful comparison runs one way: the child against themselves three months ago.
• An indicator that cannot be explained is not an indicator; it should prompt a teacher's question, never a verdict.

None of this means abandoning data. On the contrary: institution-level patterns are valuable — where observation thins out, which class drops records in which period. These are read to direct support, not to evaluate teachers.

What a product does not contain is often as meaningful as what it does. Not ranking children is not a technical gap; it is the field's own principle, written into the code.