AI-Assisted Assessment Item Generation

AI-assisted assessment item generation uses a model to propose questions, answer options, marking guidance, or variations of an existing task. The useful unit is a reviewed assessment item, not a fluent question alone. An item must elicit evidence about a specified learning goal, have a defensible answer, and work for the intended learners. Generating many alternatives quickly can help drafting while also multiplying errors that reviewers must detect.

Begin with an assessment blueprint: the concept, cognitive demand, response format, relevant prior knowledge, and evidence a correct response should provide. A prompt can then request a question with a worked solution and an explanation of each distractor. These requests make review easier, but the model's account of its own question is another output to verify. A label such as 'application' does not establish that the item actually requires applying knowledge.

Suppose the target is reasoning about a fraction of a quantity. A draft asks: 'A tank is three-quarters full. Two-thirds of its water is removed. What fraction of the tank's capacity remains?' The removed amount is one-half of capacity, leaving one-quarter. Distractors might include one-half, the amount removed; one-twelfth, obtained by directly subtracting the two given fractions; and two-thirds, copied without interpreting the situation. A teacher checks both the arithmetic and whether those explanations represent plausible errors.

The wording still needs attention. Does 'its water' clearly refer to the water currently in the tank? Could a diagram reveal the answer unintentionally? Does unnecessary language make the question harder for reasons unrelated to fractions? Are all options expressed in the same form? Independent solution and peer review can catch mistakes that asking the same model to approve its own draft may miss. The NBME item-writing guidance illustrates why clarity, appropriate content sampling, and functioning options matter, although its examples come from health sciences.

Research on human–language model collaboration for mathematics questions explores how educators can shape and review generated items. Such studies support examining a collaborative workflow; they do not establish that every generated item is ready for consequential use. Content review is followed by trial responses where feasible. Observed difficulty, discrimination, and distractor choices reveal how an item behaves with learners. A model's predicted difficulty is a hypothesis, not a calibrated measurement.

Near-duplicate variants also deserve scrutiny. Changing names or numbers may leave the same reasoning and reveal little additional evidence. More substantive variation can accidentally change the construct or introduce an extra prerequisite. Maintain a record of the final item, source material, solution, reviewers, revisions, and any trial evidence. Check copyright and avoid including identifiable student responses in an external service without an authorized purpose and appropriate safeguards.

For classroom practice, teachers can select a small number of drafts, solve them independently, and compare learner explanations after use. High-stakes examinations require a more rigorous development and security process. The aim is to use generation to widen useful drafting possibilities while preserving professional responsibility for what the item measures and what conclusions its responses can support.