Student context can change an AI assessment
Six language models changed scores, feedback or answers when prompts included explicit demographic descriptions or prior conversation histories. Removing a demographic field alone may therefore leave contextual signals in an assessment workflow.
Why it matters for leaders
Before an assessment pilot, have the assessment lead compare identical work with and without student context. Keep teachers responsible for grades, document disagreements and provide an appeal route.
The evidence
Controlled model tests used 40 sampled tasks from each of two educational datasets and 200 user profiles. Essay scoring, formative feedback and English-language questions were examined. The September 15 arXiv release identifies the work as accepted for EMNLP Findings.
Who was studied
Existing essay and English-language question datasets; no classroom intervention.
What the evidence cannot tell us
These are model-output experiments, not measured discrimination rates in a deployed district. Automated readability and sentiment measures do not establish educational harm. Results depend on the tested models and prompts.
Read the original sources
Summary and leadership analysis from the Week ending 2026-09-18 brief. The edition’s selection notes explain source access and evidence limitations.