Can universities personalize support without predictive risk scoring?

Two very different models of personalization

Assistance-based personalization begins with an explicit student task. The system uses the minimum relevant context to make that task easier, shows the student what it used, and allows correction. Predictive personalization begins with an institutional outcome, infers risk, and may trigger treatment the student did not request or understand.

The distinction matters because the same technical capability can support or constrain agency. A recommendation can broaden options and explain tradeoffs. A risk label can narrow how staff see a student, travel across systems, and become difficult to challenge even when it is wrong.

Design rules for assistance without surveillance

Collect context at the moment it becomes useful. Make optional fields genuinely optional. Separate a student's planning workspace from administrative monitoring. Do not train models on student content by default. Retain data for a stated purpose and period, and make correction and deletion understandable rather than merely possible.

Sensitive inferences need a higher bar than ordinary recommendations. If a system recognizes potential distress, the institution needs a carefully reviewed response that minimizes overreach while making real help reachable. The model's signal is not a diagnosis, and it should not become one in downstream records.

• Purpose limitation: every field has a named use.
• Data visibility: students can see and correct their context.
• No silent repurposing for model training or unrelated analytics.
• No adverse action based solely on an automated inference.
• Separate support signals from permanent student labels.

Questions procurement should ask

Institutions should ask vendors to diagram data flows, subprocessors, retention, model-provider settings, access controls, deletion behavior, audit logs, and incident response. They should also ask the product question that security questionnaires miss: what judgments does the system make about a student, and who can act on them?

A strong answer is testable. Contract language, product behavior, administrative controls, and the student-facing explanation should agree. If the system cannot show why a recommendation appeared or which source supported it, the institution cannot meaningfully govern it.