Turn an interest into a research direction

Interest, question and contribution are separate decisions

An interest names a field; a question identifies an unresolved relationship in a setting; a contribution explains what the answer would change. Write what you want to understand, then what existing accounts explain and leave unresolved. That gap needs traceable literature; until then, label it a candidate gap.

Use constraints instead of asking AI for the hottest topic

List time, skills, material access and supervision before comparing questions. Feasibility is not a summed score: unavailable essential data can rule out a design. Ask AI to identify evidence bottlenecks, alternative explanations and narrower variants. Choose the uncertainty yourself. Textual interpretation, qualitative inquiry and quantitative causal questions need different kinds of evidence.

Design evidence that distinguishes explanations

Suppose students score higher after using AI. Understanding, fluent phrasing and extra practice are competing explanations. Immediate short answers may not distinguish them. Equal-time comparisons, delayed testing and transfer to unfamiliar material supply different information at different costs. State an observation that would change your preferred explanation.

Use AI as a question editor and critical reader

Draft first. Ask AI to identify ambiguous terms, missing populations and evidence gaps; then revise yourself. Without accessible source texts, use it for search terms and hypotheses to check, not assertions that nobody has studied a topic. Preserve your tradeoffs and rejected alternatives in the final card.

Worked example · fictional

From a problem to inspectable work

Fictional example: I want to know whether AI improves university learning. I have six weeks, public papers and no participant-recruitment arrangement.

1. Separate satisfaction, immediate correctness, retention and transfer; they are different outcomes.
2. Compare a new intervention study with examining transfer measures in existing studies. The former needs unavailable recruitment and review arrangements.
3. Choose a feasible literature analysis: check existing reviews, bound the setting and outcome, and record search scope. Not finding a study does not prove absence.

Candidate card: examine how retrieved university-course studies operationalize transfer, comparing task distance, test timing and scoring. Pilot extraction on three accessible full texts. If measures are incompatible, map definitions rather than pooling effects. The question's value still requires literature checking.

Failure case and repair

Ask AI to generate novel topics from popularity.

Interests, gaps and accessible evidence are disconnected.

Compare two explanations and one feasible test.