03 · What You Need to Know
Where AI Assistance Ends and Human Scientific Judgment Must Govern
Why Is Research Design More Than a Technical Selection Problem?
Choosing a research design involves more than matching a question to a methodological label.
Researchers must determine what they seek to understand, what evidence would count as an adequate answer, which assumptions are acceptable, and how the study's limitations affect interpretation.
For example, AI may recommend a randomized experiment to investigate whether AI tutoring improves academic performance.
However, the researcher must determine whether randomization is permissible, what comparison condition is educationally appropriate, which outcome represents meaningful learning, and whether the intervention can be implemented responsibly.
These decisions involve scientific reasoning, contextual knowledge, and ethical responsibilities that cannot be resolved solely through procedural recommendations.
Which Decisions Require Accountable Human Judgment?
| Research decision |
How AI may assist |
What humans must determine |
| Research problem |
Identify gaps, alternative questions, and possible contributions. |
Which problem is scientifically meaningful and worth investigating. |
| Theoretical framework |
Suggest theories, concepts, and possible relationships. |
Which explanations are justified and relevant to the phenomenon. |
| Study design |
Compare methodological approaches and assumptions. |
Which design can produce defensible evidence under actual conditions. |
| Causal inference |
Identify potential confounding, alternative explanations, and identification requirements. |
Which causal assumptions can reasonably be defended. |
| Measurement |
Suggest instruments, indicators, and operational definitions. |
What the constructs mean and whether measurements support intended interpretations. |
| Participant protection |
Flag ethical risks and possible safeguards. |
Which procedures are acceptable under applicable ethical and institutional requirements. |
| Implementation |
Identify resource requirements and operational dependencies. |
Whether the procedures can be conducted responsibly and as specified. |
| Interpretation |
Compare explanations and identify limitations. |
What conclusions the evidence supports and how uncertainty should be communicated. |
The distinction is not between AI performing simple tasks and humans performing difficult tasks. It concerns who has the authority and responsibility to make consequential scientific decisions.
Who Decides What Research Question Is Worth Investigating?
AI can suggest research questions based on existing literature, theoretical gaps, or practical problems.
However, identifying a question worth investigating requires judgment about scholarly significance, social relevance, feasibility, originality, and the consequences of the research.
For example, AI may recommend studying whether generative AI increases student productivity because the topic is frequently discussed.
A researcher may instead consider whether increased productivity reflects meaningful learning, superficial task completion, or changes in assessment practices.
The choice of what deserves investigation is not determined by topic popularity or the ease with which AI can generate a research proposal.
It requires the researcher to establish a defensible scientific purpose.
Who Determines Which Theory Should Guide the Study?
AI may recommend established theoretical perspectives based on terminology and common scholarly associations.
However, selecting a theoretical framework requires examining the original concepts, assumptions, explanatory mechanisms, and relevance to the research problem.
A theory should not be adopted merely because AI presents it confidently or because it is frequently used in similar studies.
The researcher must determine whether the theory provides an appropriate explanation and whether its limitations are understood.
This responsibility is particularly important when AI combines theories or proposes relationships that are not supported by primary theoretical sources.
Who Determines Whether a Causal Assumption Is Defensible?
AI can explain causal identification assumptions and identify possible violations.
However, deciding whether those assumptions are plausible requires substantive knowledge of the phenomenon, research setting, and data-generating process.
For example, a difference-in-differences analysis may require an assumption about how outcomes would have changed in the absence of an intervention.
AI can describe the assumption and suggest diagnostic analyses, but the researcher must evaluate whether it is credible in the actual context.
Hernán and Robins (2020) emphasize that causal inference depends on explicitly defined questions and assumptions.
The researcher cannot delegate the responsibility for defending those assumptions to a system that does not independently observe the relevant causal process.
Who Decides Whether a Research Design Is Ethically Acceptable?
Ethical acceptability requires accountable human and institutional judgment.
AI can identify potential concerns involving informed consent, confidentiality, participant burden, intervention allocation, and data protection.
However, it cannot authorize a study, grant ethics approval, or determine the acceptable balance of risks and benefits on behalf of responsible institutions.
The Belmont Report identifies respect for persons, beneficence, and justice as foundational ethical principles for research involving human subjects.
Applying such principles requires attention to the actual participants, procedures, and context.
For example, a statistically attractive control condition may be unacceptable if it deprives students of required educational support.
Researchers must therefore evaluate ethical considerations independently of methodological convenience.
Who Determines Whether a Construct Has Been Measured Appropriately?
AI may recommend existing instruments or generate candidate questionnaire items, but it cannot establish measurement validity merely by producing plausible indicators.
The researcher must define the construct, examine relevant theoretical and empirical evidence, and determine whether the proposed measurements support the intended interpretation.
The Standards for Educational and Psychological Testing emphasize validity as the degree to which evidence and theory support interpretations of test scores for proposed uses.
This perspective is particularly important in educational research, where constructs such as engagement, self-efficacy, motivation, and learning may be operationalized in substantially different ways.
AI assistance may support measurement planning, but the researcher remains responsible for the validity argument.
Who Decides Whether a Study Is Feasible?
AI can identify recruitment, timeline, funding, technical, and analytical requirements.
However, the researcher and responsible institutions must verify whether the necessary conditions are available.
For example, AI may recommend a multi-university trial, but it cannot independently establish that participating institutions will approve recruitment or share academic records.
Similarly, it may recommend a longitudinal design without knowing whether participants can be followed over the required period.
The researcher must determine whether the design can be implemented without compromising its essential methodological features.
This requires examining both operational feasibility and field-specific constraints.
Who Is Responsible When AI Recommends an Incorrect Method?
The researcher remains responsible for the methodological decisions incorporated into the study.
For example, if AI recommends a nonrandomized comparison and incorrectly describes it as a randomized experiment, the error must be corrected before the methodology is presented or implemented.
Similarly, if AI recommends an inappropriate statistical model, the researcher must evaluate its assumptions and suitability rather than treat the recommendation as authoritative.
AI assistance does not transfer authorship responsibility, scientific accountability, or institutional obligations to the system.
The International Committee of Medical Journal Editors' recommendations emphasize that authors remain responsible for the accuracy and integrity of their work when AI-assisted technologies are used.
Although publisher and institutional policies differ, the underlying principle of accountable human responsibility remains central to responsible research practice.
Does Human Judgment Mean Researchers Should Ignore AI Recommendations?
No. AI may provide useful methodological explanations, identify overlooked alternatives, and challenge assumptions that researchers have not explicitly examined.
For example, a researcher may initially propose a simple pretest-posttest design. AI might identify history and maturation as potential alternative explanations and suggest considering a comparison group.
That feedback can improve the researcher's understanding of the design.
However, the researcher must determine whether the proposed comparison is feasible, ethically acceptable, and capable of strengthening the intended inference.
Human judgment should be informed by evidence and criticism rather than insulated from them.
Can Human Judgment Itself Be Biased or Incorrect?
Yes. Human researchers may also rely on familiar methods, overlook assumptions, or prefer designs that confirm their expectations.
Recognizing the limitations of AI does not imply that human judgment is infallible.
The appropriate safeguard is a research process that exposes important decisions to scholarly evidence, transparent reasoning, methodological review, and appropriate institutional oversight.
AI can contribute to that process by presenting alternative explanations and identifying possible inconsistencies.
However, AI-generated criticism should be evaluated rather than treated as an independent guarantee of correctness.
What Is the Difference Between Delegating a Task and Delegating Scientific Judgment?
Some research planning tasks can be assisted or partially automated without transferring responsibility for the underlying decision.
For example, AI may help format a sampling plan, organize a methodological comparison, or draft code implementing a prespecified allocation procedure.
These tasks can be checked against explicit requirements.
By contrast, deciding whether the sampling plan supports the intended population inference or whether the allocation procedure is ethically and methodologically appropriate requires substantive judgment.
Delegating a bounded task
Using AI to perform or organize a procedure whose purpose, requirements, and acceptance criteria have been established and can be independently checked.
Delegating scientific authority
Allowing AI recommendations to determine consequential methodological choices without adequate human understanding, verification, or accountability.
The distinction helps researchers use AI efficiently without allowing convenience to replace scientific reasoning.
Should Researchers Always Follow Expert Review Over AI?
Expert review deserves serious consideration, but methodological disagreements should be resolved through evidence and reasoning rather than automatic deference.
A reviewer may identify a problem that AI overlooked, while AI may raise a legitimate alternative that a reviewer had not considered.
The researcher should examine the competing arguments, relevant methodological literature, and actual research conditions.
For complex or consequential decisions, additional specialist review may be appropriate.
The purpose of human oversight is not to substitute authority for evidence. It is to ensure that decisions are examined and justified by accountable individuals.
What Decisions Should Never Be Made Solely Because AI Recommends Them?
Researchers should be particularly cautious when AI recommendations affect the scientific purpose, participant welfare, validity of evidence, or interpretation of results.
| Decision |
Why independent judgment is essential |
| Changing the research question |
May alter the study's intended contribution and evidence requirements. |
| Selecting a causal identification strategy |
Requires defensible assumptions about the data-generating process. |
| Withholding or assigning interventions |
May affect participant welfare and ethical obligations. |
| Using identifiable participant data |
Requires appropriate authorization, safeguards, and governance. |
| Interpreting measurement scores |
Requires validity evidence supporting the intended construct interpretation. |
| Claiming intervention effectiveness |
Requires evidence and assumptions sufficient for the causal conclusion. |
| Changing essential protocol procedures |
May affect validity, comparability, ethics, and previously granted approvals. |
AI may assist in examining these decisions, but the authority to approve and defend them belongs to responsible researchers and institutions.
How Can Researchers Preserve Human Scientific Judgment While Using AI?
1. Define the scientific objective. Establish the research problem and intended contribution independently of AI-generated recommendations.
2. Use AI to explore alternatives. Request methodological options, assumptions, limitations, and competing explanations.
3. Identify consequential decisions. Determine which choices affect participant welfare, research validity, feasibility, or interpretation.
4. Verify the evidence. Consult relevant methodological literature, institutional requirements, and actual research conditions.
5. Obtain appropriate review. Seek specialist, supervisory, ethical, or institutional evaluation where necessary.
6. Document the rationale. Record why substantive recommendations were accepted, modified, or rejected.
7. Retain responsibility. Ensure that the research team can explain and defend the final design without relying on AI as an authority.
This process allows AI to contribute to methodological reasoning while preserving the accountability required for scientific research.
Watch Out
Do not treat AI's confidence, methodological vocabulary, or agreement with another AI system as evidence that a scientific decision has been validated. Consequential choices require independent justification and accountable oversight.
