Overview
Generative AI can raise the quality of a submitted essay, calculation, or program while weakening the inference that the student owns the underlying capability. This series shows how to redesign assessment around an explicit human learning claim, authentic task conditions, and the smallest set of product, process, explanation, and transfer evidence needed to support it. It examines the false-mastery mechanism, Evidence-Centred Design, a structured conversational exam, the staffing and fairness conditions for scale, and different roles for AI as tutor, feedback partner, or assessment tool. The overall conclusion is conditional: AI need not be excluded, but grades become more defensible when students must attempt, verify, explain, adapt, and demonstrate what remains after assistance is removed. Evidence remains task-specific and limited; no single format solves assessment.
Learning outcomes
- A polished AI-assisted product is not sufficient evidence of independent learning when the system performed the target reasoning.
- Assessment validity improves when the human capability, task conditions, observations, and inference are explicitly connected.
- Scaffolded AI can support learning when students still attempt, evaluate, explain, and transfer the central operation.
- Structured human observation can work at moderate scale, but staffing, fairness, privacy, and evidence limits remain part of validity.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why can AI-assisted assignments overstate student learning?
They can overstate learning when AI performs the reasoning the assessment intends to observe, raising task performance while leaving independent capability untested.
What evidence should an AI-era assessment collect?
It should collect product evidence plus only the relevant process, explanation, or transfer observations needed to support the stated human capability and its boundary.
Can students use AI without undermining assessment validity?
Yes, when the assessment makes the permitted role explicit and still observes the student’s attempt, verification, judgment, and ability to apply or explain the result.
Does a conversational exam scale for large classes?
The reported case covered 58 students in ten half-hour sessions, showing moderate-class feasibility rather than proving linear scale; preparation, staffing, accommodations, and comparable questions remain necessary.