Teaching · Practical guide
Use AI to Develop Research Questions and Verify Sources
Turn a broad interest into a searchable question, keep a search log and evaluate evidence rather than treating generated references as research.
Separate question development from finding evidence
AI can help you identify possible terms, perspectives and subquestions. It cannot make an unsupported claim become evidence. Begin with the decision or uncertainty you want to investigate, then define the population, context and phenomenon closely enough that you can recognise a relevant source.
For a novice exercise, replace “AI improves learning” with “What evidence would help compare two feedback approaches for adult beginners learning a calculation?” This asks about evidence without assuming the conclusion. Record what would count as an outcome, such as an independently completed follow-up task, and what would not, such as the attractiveness of the feedback text.
Build a small search plan
List concept groups and possible synonyms. For the example, one group concerns feedback, another concerns adult or higher education, and another concerns learning or performance. Ask AI to suggest alternative terminology, then check those terms in real subject databases or library resources. Database syntax differs, so verify how phrases, truncation and Boolean operators work in the service you use.
Keep a log containing the database, date, search expression, filters and useful results. ACRL treats research and searching as activities requiring judgement and iteration. In this exercise, the practical purpose of the log is reproducibility: you should be able to explain how you found a source and why you changed your search.
Worked example: narrow an untestable question
Broad claim: “AI feedback is better.” Problems include an undefined comparator, learner group, task and meaning of “better”. A more focused illustrative question is “For adult beginners practising percentage calculations, how could we compare AI-drafted, teacher-checked comments with a standard worked solution on a later independent problem?” This is a proposed question, not a report of an experiment.
Possible evidence includes the study’s participant context, the exact feedback received, the task used to measure learning, the comparison conditions and limitations. Satisfaction alone would not answer a question about independent calculation. Faster completion might be relevant but would need interpretation alongside correctness.
A search could begin with concept groups for formative feedback, generative AI and higher education. If it returns only opinion pieces, keep those separately as context rather than treating them as experimental evidence. If it misses relevant work using “automated feedback”, revise the terminology and record that change.
Verify existence, relevance and support separately
For each promising reference, locate the original publisher or repository record. Compare title, authors, year and identifier. Then read the relevant text. A real DOI proves neither that an accompanying citation is accurate nor that the paper supports the claim you want to make.
Create three separate decisions: does the work exist, is it relevant to the question, and what does it actually support? A study about school pupils may inform a higher-education question, but the population difference belongs in your interpretation. Do not erase that difference because a generated summary says the source is “highly relevant”.
Use a claim-to-evidence table
A useful table contains proposed claim, source location, evidence, limitation and decision. Write the evidence in your own words after reading. Keep direct quotations short and accurate, with the required citation. Mark “not reported” when a detail is absent rather than asking the model to fill a gap.
For a fictional source card stating only that participants preferred shorter comments, an acceptable claim is about reported preference in that context. “Short comments improved exam performance” exceeds the card. The correct decision is to narrow the claim or find additional evidence, not to make the sentence sound more academic.
Practice: a 20-minute research-planning exercise
Spend five minutes narrowing a broad topic and defining two key terms. Spend five minutes drafting a search expression and recording where you will run it. Use five minutes to inspect one real primary source or a supplied source card. Spend five minutes writing one supported claim and one unresolved question.
Example answer for “technology helps students”: choose a specific tool-assisted task, learner context and observable outcome. A source about time saved can support a qualified statement about time under its conditions; it cannot by itself support deeper understanding. Write that boundary in the evidence table.
Self-review: is the question open to an answer that challenges your expectation? Can someone repeat your search from the log? Does each retained claim match what you actually read? This activity produces a defensible starting point, not a complete literature review or a claim that the evidence base has been exhausted.
Reusable prompt
Select and copy this template, or customise your own using the free builder.
Help refine this research interest: [topic]. Suggest bounded questions, key concepts, alternative search terms and evidence needed to answer each question. Do not invent references or claim a search was performed. Given these verified source notes [notes], separate supported claims from assumptions and gaps. Preserve population, method and limitation details.Open the prompt builder →
Pause · Think · Check
Test your understanding
Choose one answer per question, then check your reasoning. This is practice; scores are not saved.