03 · What You Need to Know
Where Human Scientific Judgment Must Remain Meaningful
Researchers Must Determine the Scientific Purpose of the Study
AI can generate possible research questions, identify apparent literature gaps, and suggest theoretical perspectives.
However, the decision about what a study should investigate requires consideration of scientific significance, available evidence, feasibility, ethical implications, and the researcher's disciplinary context.
A model may suggest a fashionable topic without establishing whether the proposed question addresses a genuine knowledge gap.
Researchers should examine the relevant literature and determine whether the question is sufficiently important, answerable, and appropriate for the intended research setting.
AI can broaden the set of possibilities, but the justification for the selected research problem must remain grounded in scholarly evidence and accountable human reasoning.
Researchers Must Justify the Study Design
Study design determines what kinds of conclusions the research can support.
AI may recommend experimental, observational, qualitative, mixed-methods, or other designs, but the suitability of those recommendations depends on the research question and practical constraints.
For example, a cross-sectional survey may identify associations but ordinarily cannot establish causal effects without additional assumptions and evidence.
Researchers must determine whether the proposed design adequately addresses the intended question and whether its limitations are acceptable.
Where specialized expertise is needed, consultation with qualified methodologists is appropriate. AI-generated recommendations should be evaluated rather than adopted solely because they appear comprehensive.
Ethical Decisions Cannot Be Delegated to an AI Recommendation
Research involving human participants requires attention to consent, confidentiality, risk, fairness, and the protection of participants' rights and welfare.
AI may help researchers identify possible ethical issues or organize documentation, but it cannot replace the responsibilities of investigators, ethics committees, or other authorized decision-makers.
For example, a model may recommend collecting additional personal information to improve predictive accuracy. Researchers must determine whether collecting that information is necessary, authorized, proportionate, and ethically appropriate.
Likewise, an AI system's suggestion that a participant is eligible for a consequential intervention should not bypass the approved eligibility and safety procedures.
Ethical authorization and accountability remain with the responsible humans and institutions.
Researchers Must Decide What Data Are Appropriate to Collect
Data collection decisions shape the evidence available for analysis.
AI can suggest survey items, interview questions, observation categories, or variables, but researchers must evaluate whether those instruments and measures adequately represent the intended constructs.
For example, an AI-generated questionnaire may contain items that appear relevant to AI literacy while failing to represent important dimensions of the construct.
Researchers should assess content relevance, validity evidence, participant burden, accessibility, and applicable ethical requirements.
AI-generated items may be useful candidates for instrument development, but their inclusion requires appropriate evaluation.
Researchers Must Establish Data Inclusion and Exclusion Criteria
Decisions about which observations or studies to include can materially affect research findings.
AI may assist with screening records, identifying possible duplicates, or flagging unusual observations.
However, the criteria governing inclusion and exclusion should be scientifically justified and applied consistently.
For systematic reviews, eligibility criteria should follow the review question and protocol. AI-assisted screening may be useful, but researchers must establish procedures for validating decisions and resolving uncertain cases.
For statistical analysis, removing observations because AI describes them as outliers is insufficient without a defensible methodological rationale.
Human accountability concerns the criteria, their application, and the consequences of exclusions, even when parts of the process are automated.
Researchers Must Select and Defend Analytical Methods
AI may recommend statistical tests, machine-learning algorithms, qualitative approaches, or data transformation procedures.
The final analytical choice must be justified by the research question, study design, data structure, and relevant assumptions.
For example, an AI system might recommend a complex machine-learning model when a transparent regression model adequately addresses the study's objective.
Alternatively, it might recommend an overly simple procedure that ignores clustering, repeated observations, or important confounding.
Researchers should compare suitable alternatives and explain why the selected approach is appropriate.
Where AI generates code, the implementation must also be verified independently of the method-selection decision.
Researchers Must Decide How to Handle Missing and Problematic Data
Missing observations, inconsistent records, and unusual values require methodological judgment.
AI can identify possible anomalies or suggest imputation procedures, but researchers must assess the reasons for missingness and the implications of different treatments.
For example, replacing missing survey responses with zero may distort a scale when zero is not a valid response category.
Likewise, excluding incomplete records may introduce bias when missingness is related to participant characteristics or outcomes.
Researchers should establish and document defensible procedures rather than allowing generated code to make unexamined data-handling decisions.
Researchers Must Interpret What the Evidence Actually Supports
Interpretation is a central scientific responsibility.
AI may propose explanations for observed patterns, compare results with literature, or identify alternative interpretations.
However, researchers must determine which conclusions follow from the evidence and which remain speculative.
A model may describe a statistically significant association as proof of causation or interpret a qualitative theme more broadly than the participant material supports.
Researchers should evaluate effect sizes, uncertainty, design limitations, contradictory evidence, and alternative explanations.
The final interpretation should reflect the strength of the evidence rather than the confidence of the AI-generated narrative.
Researchers Must Decide Which Findings Are Scientifically Reportable
AI can organize results and suggest which findings to emphasize, but researchers should not allow generated recommendations to determine reporting selectively.
Decisions about reporting should follow the study's objectives, protocol, disciplinary standards, and relevant reporting guidelines.
For example, an AI system might recommend highlighting statistically significant results while omitting null findings. Accepting this suggestion could create a misleading account of the study.
Researchers must ensure that reporting is complete and appropriately balanced.
Where exploratory analyses are conducted, they should be distinguished from prespecified analyses rather than presented as though they were planned from the outset.
Researchers Must Determine Whether an AI Output Is Acceptable
AI-generated content should not enter the research record solely because it is fluent, plausible, or consistent with the researcher's expectations.
Researchers must establish appropriate criteria for acceptance.
| Research decision |
How AI may assist |
What remains under human accountability |
| Research question |
Suggest possible questions or gaps |
Scientific significance, feasibility, and justification |
| Study design |
Compare possible designs |
Suitability for the research question and intended inference |
| Participant protection |
Identify possible risks |
Ethical decisions, safeguards, and authorized procedures |
| Data collection |
Draft candidate instruments |
Validity, appropriateness, consent, and data minimization |
| Data inclusion |
Flag possible exclusions or duplicates |
Criteria, verification, and final defensibility |
| Analysis |
Recommend methods or generate code |
Method selection, validation, and interpretation |
| Results reporting |
Organize tables or draft descriptions |
Accuracy, completeness, and balanced reporting |
| Conclusions |
Suggest explanations or implications |
Evidence-based interpretation and appropriate uncertainty |
The distinction is not between human action and automated action. It is between accountable scientific governance and unexamined delegation.
Can AI Make Operational Decisions Automatically?
Yes, in appropriately designed and authorized workflows.
Research routinely uses software to sort data, apply predetermined rules, calculate statistics, and execute analytical procedures.
AI systems may also perform specified operations automatically when their suitability has been established and the resulting risks are adequately managed.
For example, an approved screening system may automatically flag records for further review using validated criteria.
However, automation should operate within a framework established by responsible researchers. Consequential decisions require appropriate validation, oversight, and procedures for handling errors.
Automating an operation does not eliminate accountability for the criteria and conditions under which it is performed.
What Makes Human Oversight Meaningful?
Meaningful oversight requires more than placing a human approval step at the end of an automated process.
Reviewers need sufficient expertise, access to relevant evidence, time to evaluate the output, and authority to reject or modify the recommendation.
They should also understand the circumstances in which the AI system is likely to fail.
For example, a researcher who approves AI-generated classifications without reviewing source material may not provide meaningful oversight, even if every output receives a human confirmation.
Oversight becomes substantive when human decisions are based on independent criteria and evidence.
How Can Researchers Avoid Automation Bias?
Automation bias occurs when individuals place excessive weight on automated recommendations, potentially overlooking contradictory evidence.
AI systems may encourage such reliance when their outputs are fluent, detailed, or consistently plausible.
Researchers can reduce this risk by establishing decision criteria before consulting AI, reviewing relevant evidence independently, and explicitly considering alternative explanations.
For consequential decisions, independent review before exposure to an AI recommendation may sometimes help reduce anchoring effects.
However, no single procedure eliminates automation bias. Oversight should be evaluated according to the actual task and working conditions.
What if AI and a Human Expert Disagree?
Disagreement should prompt examination of evidence rather than automatic preference for either source.
An AI system may identify a genuine weakness in a human recommendation, while a human expert may recognize contextual limitations that the AI overlooked.
Researchers should compare the assumptions, evidence, and methodological reasoning supporting both positions.
Where the disagreement concerns a consequential analytical decision, additional independent evidence or qualified consultation may be necessary.
The final decision should be justified by the research requirements, not by the status of the person or system making the recommendation.
Who Is Responsible for the Final Manuscript?
Human authors remain responsible for the integrity of the work they submit or publish.
AI systems cannot fulfill the responsibilities associated with scholarly authorship, including accountability for accuracy, appropriate correction, and responses to questions about the research.
The International Committee of Medical Journal Editors states that AI-assisted technologies should not be listed as authors and that human authors remain responsible for their use.
Researchers should therefore ensure that AI-generated manuscript content has been appropriately reviewed and that relevant assistance is disclosed according to applicable journal policies.
Responsibility extends beyond wording to the methods, evidence, interpretations, and conclusions presented.
When Should Researchers Refuse to Delegate a Decision to AI?
Researchers should not delegate a consequential decision when the system has not been adequately validated, when the decision requires authorization that AI cannot provide, or when the consequences of error cannot be acceptably controlled.
They should also avoid unexamined delegation when the decision depends on methodological assumptions or contextual evidence that the research team cannot adequately evaluate.
Where AI assistance remains useful, its role can be restricted to generating options, flagging uncertainties, or supporting documentation.
The final decision should follow a defensible human-governed process.
