03 · What You Need to Know
The Research-Question Decisions That Require Human Scholarly Judgment
1. Deciding Which Research Problem Deserves Investigation
Perhaps the most consequential decision comes before the research question is written: determining what problem is worth investigating.
Generative AI can identify frequently discussed topics, suggest possible knowledge gaps, and generate questions that resemble those found in scholarly literature. However, the frequency with which a topic appears in academic writing does not establish its scientific importance.
A research problem may deserve attention because existing explanations are inadequate, evidence is contradictory, a theoretically important relationship remains uncertain, or a practical or policy decision requires better evidence.
Determining which of these concerns matters requires engagement with the relevant scholarship and research context.
For example, AI might recommend studying university students' acceptance of generative AI because technology acceptance is a familiar research area. Yet the researcher may recognize that the more consequential problem concerns how universities evaluate learning when students routinely use AI assistance.
Both topics could support research. The decision about which problem warrants investigation cannot be settled by the model's preference for one.
AI can help convert a topic into candidate research questions, but the researcher must establish why a particular question deserves to be asked.
2. Determining the Intended Contribution
A research question should be connected to a contribution the researcher can articulate and defend. That contribution might involve developing or challenging theoretical understanding, addressing uncertainty in existing evidence, improving methodological approaches, or informing decisions in practice and policy.
AI can propose statements about significance and originality. It may even describe a question as groundbreaking or claim that little research has examined the issue.
Those statements are not evidence of contribution.
The researcher must examine what is already known, what remains uncertain, and how answering the proposed question could advance understanding.
Importantly, a contribution does not always require investigating an entirely unexplored topic. Replication, contextual extension, methodological scrutiny, and the examination of conflicting findings can also be valuable when appropriately justified.
The decision about what contribution matters should emerge from scholarly reasoning rather than an AI-generated novelty claim.
3. Choosing What the Research Question Actually Asks
One research topic can support many different questions. The researcher must decide which inquiry reflects the intended intellectual purpose.
Consider the topic of generative AI in academic writing. Possible questions include how students use AI, why instructors permit certain forms of assistance, whether AI use is associated with writing confidence, or whether an AI-assisted intervention improves assessed writing performance.
These questions are not simply alternative phrasings. They concern different phenomena and require different evidence.
AI can help draft and compare research questions, but the researcher must choose the inquiry that corresponds to the justified research problem.
A fluent question is not necessarily the right question.
4. Defining the Population, Context, and Scope
Every research question establishes boundaries, whether explicitly or implicitly. These boundaries determine which participants, settings, phenomena, relationships, or periods the study will investigate.
AI can suggest narrowing a question to a particular educational level, institution, discipline, or population. Such suggestions may improve feasibility, but they may also introduce arbitrary restrictions.
For example, limiting a study of instructors' AI-related assessment decisions to one institution may be reasonable because of access or contextual relevance. It should not be presented as theoretically necessary unless that claim can be defended.
The researcher must decide which boundaries are justified by the problem and which limitations are imposed by practical circumstances.
When using AI to refine an overly broad research question, the aim should be a defensible inquiry rather than the narrowest sentence the model can generate.
5. Selecting the Theoretical or Conceptual Perspective
Theories and conceptual frameworks influence which phenomena researchers notice, how constructs are defined, and what relationships are considered meaningful.
AI may recommend a familiar theory because its terminology resembles the proposed research topic. In educational technology, for example, technology-acceptance frameworks may appear whenever a question concerns AI adoption.
But studying AI adoption does not automatically mean that a particular acceptance model is the most appropriate theoretical foundation.
The researcher must evaluate whether a framework addresses the actual phenomenon, what assumptions it makes, how those assumptions fit the research context, and whether competing perspectives offer a more convincing account.
Sometimes an inquiry may require multiple complementary perspectives. Other studies may be exploratory or theoretically generative rather than organized around a single predetermined framework.
AI can explain candidate theories and compare their assumptions. The researcher must justify the perspective adopted.
6. Defining the Meaning of Key Constructs
Research questions often contain concepts whose meanings are not interchangeable.
Writing confidence, writing self-efficacy, writing competence, and writing performance may be related, but they represent different conceptual and potentially operational commitments.
AI may substitute one term for another because the words frequently appear in similar contexts. Such changes can alter the research question even when the revised sentence sounds more academic.
The researcher must determine which constructs are theoretically relevant, how they are defined, and what evidence could represent them appropriately.
This is particularly important when AI is asked to improve existing wording. Researchers should prevent revisions from changing the intended meaning unless a substantive change has been deliberately approved.
7. Deciding Whether the Study Makes Descriptive, Relational, Predictive, or Causal Claims
The kind of question a researcher asks determines what forms of inference may be required.
A descriptive question asks what is occurring. A relational question concerns associations. A predictive question concerns the ability to anticipate outcomes. A causal question concerns what would happen under different exposures or interventions, subject to an appropriate causal framework.
AI sometimes strengthens wording by replacing "associated with" with "influences" or "causes." This may appear to improve the question's significance while introducing a much more demanding claim.
The researcher must decide which kind of inference is intended and whether the study can support it.
Statistical prediction does not automatically establish causation, and observing an association does not establish that changing one variable would change another.
The choice of inferential target is a substantive research decision, not an editing preference.
8. Deciding Which Variables and Relationships Belong in the Study
Generative AI can suggest predictors, outcomes, confounders, mediators, moderators, and other variables. It can also produce convincing explanations for relationships among them.
However, familiar educational constructs do not automatically belong in every model.
For example, motivation may be relevant to students' AI use and writing performance. Yet its role could differ depending on the research question, timing, and assumed causal structure.
The researcher must determine whether each proposed variable is necessary and whether its theoretical or causal role can be justified.
This becomes especially consequential when evaluating AI-suggested causal variables. Incorrect assumptions about confounding, mediation, or adjustment can distort the intended estimate.
AI can generate possible relationships. The researcher must decide which ones deserve inclusion and which should remain speculative.
9. Choosing Whether Hypotheses Are Appropriate
Not every research question requires a formal hypothesis. Some studies aim to test theoretically justified expectations, while others investigate experiences, meanings, processes, or phenomena without predetermined predictions.
AI may generate hypotheses automatically because they commonly appear in research proposals. Their presence, however, does not establish methodological rigor.
The researcher must decide whether hypotheses serve the inquiry and, when they do, which expectations are justified by theory and evidence.
Directional predictions require particular care. A hypothesis should not predict a positive relationship merely because AI considers that direction plausible.
Researchers should also distinguish hypotheses specified before examining outcomes from those developed through exploratory analysis.
10. Selecting a Research Design Capable of Answering the Question
AI can recommend surveys, interviews, experiments, quasi-experimental designs, case studies, mixed methods, and analytical procedures. These suggestions may be useful, but method selection requires more than matching familiar research-question words to familiar methodological labels.
The researcher must decide what evidence would constitute an answer and which design could generate that evidence.
For example, a one-time survey may be appropriate for estimating associations under suitable assumptions, but it would generally be insufficient to establish that AI use caused improvements in writing performance.
Likewise, a qualitative study investigating instructors' decision-making should use methods appropriate to the phenomenon and selected methodological tradition rather than automatically imposing quantitative variables and hypotheses.
The researcher must also evaluate feasibility, ethical constraints, sampling, measurement quality, and the limitations of the intended inference.
AI can help identify potential inconsistencies between questions and methods, but it cannot independently validate the research design.
11. Deciding Which Ethical Trade-Offs Are Acceptable
Research-question decisions can have ethical consequences long before data collection begins.
Choosing to investigate vulnerable populations, sensitive experiences, institutional practices, or identifiable records may introduce risks involving privacy, consent, confidentiality, fairness, or potential harm.
AI may recommend a population or data source without knowing the institutional permissions, legal obligations, or ethical sensitivities involved.
The researcher must consider whether the question can be investigated responsibly and whether its anticipated contribution justifies the burdens or risks associated with the proposed study.
Formal ethics review may be required depending on the institution, jurisdiction, funding arrangements, and research activities. AI-generated assurances that a study is ethically acceptable do not replace applicable review procedures.
12. Deciding What Conclusions the Evidence Can Support
Responsibility for the research question does not end when data collection begins. Researchers must eventually determine whether the evidence answers the question and how far the conclusions can reasonably extend.
AI can assist in organizing findings, identifying alternative interpretations, and checking whether conclusions exceed the reported evidence.
However, it may also produce overly confident statements, particularly when asked to make results sound more significant or impactful.
A statistically significant association does not automatically establish causation. A small qualitative sample does not automatically support population prevalence estimates. Positive participant perceptions do not necessarily establish intervention effectiveness.
The researcher must determine what the findings warrant, acknowledge limitations, and resist conclusions that exceed the design.
What Can AI Decide Without Taking Over the Research?
Not every AI-assisted choice carries the same intellectual weight. Some tasks concern presentation and organization rather than the substantive direction of the inquiry.
AI may help reorganize a sentence, compare terminology, produce a preliminary alignment table, or identify ambiguities for the researcher to inspect.
Even these tasks require review when they affect scientific meaning, but they do not necessarily involve delegating the decision about what the study investigates.
| Research Activity |
Appropriate AI Contribution |
Decision That Remains With the Researcher |
| Problem selection |
Suggest possible problems and relevant questions |
Determine which problem is important and justified |
| Question formulation |
Generate and compare candidate wording |
Choose the intended inquiry |
| Scope refinement |
Identify possible boundaries and trade-offs |
Decide which boundaries are appropriate |
| Theoretical framing |
Explain and compare candidate frameworks |
Justify the adopted perspective |
| Variable identification |
Identify candidate constructs and relationships |
Establish their substantive and causal roles |
| Hypothesis formulation |
Propose possible predictions |
Determine which predictions are warranted |
| Method selection |
Compare methodological options and requirements |
Select and justify a defensible design |
| Interpretation |
Identify possible explanations and limitations |
Determine what conclusions the evidence supports |
The distinction is not between AI doing nothing and the researcher doing everything unaided. Scholarly work routinely involves collaboration, consultation, and the use of specialized tools. The important boundary is that substantive decisions must remain accountable to identifiable researchers who can explain and defend them.
Researcher Responsibility Is Not the Same as Working Alone
Research is often collaborative. Supervisors, coauthors, statisticians, research assistants, community partners, and methodological specialists may contribute substantially to decisions about questions and designs.
Retaining responsibility therefore does not mean refusing advice or personally performing every analytical task.
The difference is that human collaborators can contribute expertise, participate in scholarly deliberation, and, where appropriate, accept recognized responsibilities for their contributions.
Generative AI does not assume authorship accountability or responsibility for research integrity in the way human researchers do.
The International Committee of Medical Journal Editors, for example, states that AI-assisted technologies should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work. Researchers must also follow applicable disclosure requirements.
Although publisher and institutional policies vary, the underlying principle is relevant beyond biomedical publishing: AI assistance does not transfer responsibility for scientific decisions away from the researcher.
Researcher Oversight Must Be Substantive, Not Ceremonial
It is possible to retain nominal control while allowing AI to determine nearly every important element of a study.
A researcher might accept an AI-generated problem, question, theoretical framework, variables, hypotheses, and methodology, then describe the process as human-supervised because the final proposal was reviewed.
But meaningful oversight requires the ability to evaluate alternatives, recognize unsupported assumptions, reject inappropriate suggestions, and justify the adopted decisions independently.
Simply approving an output is not equivalent to understanding its scientific implications.
The researcher should be able to explain why the question was chosen, what evidence supports its rationale, which assumptions the design requires, and what limitations remain.
Watch Out
If you cannot explain why a research question, variable, hypothesis, or method belongs in your study without referring to AI's recommendation, the decision may not yet be sufficiently justified. Treat that as a reason to investigate the underlying rationale before proceeding.
Documenting AI Assistance Can Support Research Transparency
When AI materially contributes to research planning, maintaining a record of its role may help researchers reconstruct how important decisions were made.
Such records might include the purpose of AI assistance, significant alternatives considered, substantive changes adopted, sources used to verify claims, and the researcher's reasons for accepting or rejecting suggestions.
Documentation requirements differ among institutions, funders, publishers, and research activities. Researchers should consult the policies applicable to their work rather than assume that one disclosure practice is universally sufficient.
Records are especially useful when a proposed question changes after protocol approval, preregistration, or ethical review.
The purpose is not to preserve every trivial prompt indefinitely. It is to maintain an appropriate account of decisions that materially affect the study's scientific meaning and integrity.
