Generative AI can now produce assessment questions in seconds. For educators, that is attractive because writing good multiple-choice questions takes time.
But assessment is not only about producing questions. It is about making defensible decisions about learning. A polished AI-generated item is not automatically a valid item.
The problem with polished questions
A sophisticated-looking question can still contain factual errors, ambiguous wording, cueing, weak distractors or a mismatch with the learning outcome. It may reward recognition rather than reasoning.
What recent evidence suggests
A 2026 systematic review in the Postgraduate Medical Journal found that large language models can be useful drafting tools for medical multiple-choice questions, but current evidence does not support unsupervised use in summative assessment.
What expert review still needs to check
Educators still need to review alignment, clinical accuracy, distractor quality, cueing, difficulty, fairness and consequences for learners.
A better workflow
The educator should begin with the blueprint and learning outcome. AI can generate draft options, but the educator reviews, edits, checks and decides.