03 · What You Need to Know
Where Human Responsibility Cannot Be Reduced to Clicking “Accept”
“Never Delegate Completely” Does Not Mean “Never Use AI”
The distinction matters immediately.
A researcher might ask AI to critique a research question without asking AI to decide what the study should investigate. AI might suggest statistical approaches without being allowed to determine the final method. It might identify possible interpretations without deciding what the results mean.
AI assistance
The system contributes candidates, analysis, processing, critique, or other support that the researcher meaningfully evaluates.
Complete delegation
The system effectively determines the output or decision while the researcher lacks meaningful independent evaluation, justification, or control.
This distinction allows researchers to benefit from the broad range of tasks generative AI can help with without confusing assistance with the transfer of scholarly responsibility.
Do Not Delegate the Final Definition of the Research Problem
AI can be an excellent brainstorming partner. It can generate candidate topics, questions, hypotheses, constructs, and alternative framings.
But determining what is worth studying requires engagement with scholarship, context, feasibility, theory, evidence gaps, ethical implications, and often the needs of particular communities or disciplines.
A research question is not good because a language model can make it sound sophisticated.
The researcher should ultimately be able to explain why the problem matters, how it emerges from existing knowledge, what the study can contribute, and why the question is answerable using the proposed evidence.
Do Not Delegate Final Methodological Judgment
Generative AI can suggest methods with remarkable fluency. Ask how to analyze a dataset and it may confidently recommend regression, structural equation modeling, thematic analysis, random forests, propensity-score matching, grounded theory, or almost anything else that fits the vocabulary of the prompt.
Method selection, however, is not a terminology-matching exercise.
The method must fit the research question, design, sampling, measurement, data-generating process, assumptions, inferential objective, epistemological commitments where relevant, and practical constraints.
AI can help identify options or explain methods. The final choice must be defensible independently of the statement “the AI recommended it.”
Watch Out
If you cannot explain why a method is appropriate without referring to the AI's recommendation, you have not meaningfully retained methodological control.
Do Not Delegate Ethical Responsibility for Human Participants
AI can help researchers identify possible ethical concerns, draft participant-facing language, compare procedures with supplied requirements, or generate questions for an ethics application.
It cannot assume the researcher's ethical obligations toward participants.
Researchers and responsible institutions need to make decisions involving informed consent, privacy, confidentiality, risk, vulnerability, data handling, participant welfare, and compliance with ethics review requirements.
UNESCO's guidance on generative AI emphasizes protection of human agency, privacy, ethical validation, and human accountability. Its human-centred approach specifically argues that use of generative AI in research should remain under human control rather than displacing human agency.
An AI-generated reassurance that a procedure is “ethical” has no authority equivalent to the applicable ethics process or the judgment of responsible researchers and review bodies.
Do Not Delegate the Decision to Expose Sensitive or Protected Information
An AI system should not decide for you whether participant records, unpublished findings, confidential peer-review material, proprietary information, identifiable data, or other protected content may be entered into it.
That decision depends on applicable law, ethics approval, consent, contracts, institutional policy, data-processing arrangements, security requirements, and the particular system being used.
The European Commission's updated 2026 guidelines emphasize responsible information handling and expand attention to AI-related risks involving third parties and information-management environments.
Researchers should determine whether the system is an appropriate environment before providing the material, not ask the system afterward whether uploading it was acceptable.
Do Not Delegate Authoritative Verification of Sources to the Same System That Generated Them
Suppose an AI assistant gives you a reference. You then ask, “Are you sure this citation is real?” It responds, “Yes, I have verified it.”
You still do not have independent verification.
Generative systems can produce false or unsupported information confidently. NIST refers to this phenomenon as confabulation, including false content, incorrect logic, and fabricated citations presented in ways that may encourage inappropriate trust.
Verification needs an external reference point: the actual publication, an authoritative bibliographic database, official documentation, the original dataset, a validated calculation, or another suitable source of evidence.
An AI system can help locate or inspect evidence. It should not become the sole authority certifying the truth of its own generated claims.
Do Not Delegate Final Decisions About What Counts as Evidence
Research involves decisions about relevance and evidential weight.
Which studies genuinely address the research question? Which participant statement supports a theme? Which observation should be excluded as an artifact? Which source is authoritative enough to establish a claim? Which contradictory finding needs to be taken seriously?
AI can support classification and screening, sometimes at considerable scale. But when these decisions materially determine the evidence base, researchers need a validated procedure and meaningful oversight.
This is especially important in systematic evidence synthesis, qualitative interpretation, and other research where inclusion and exclusion decisions shape what the study ultimately concludes.
Do Not Delegate Interpretation of Findings Completely
AI is exceptionally good at producing explanations.
That is precisely why researchers need to be careful.
Give a language model a surprising statistical result and it can propose several mechanisms. Give it qualitative excerpts and it can produce themes. Give it a graph and it may tell a coherent story about the pattern.
Those outputs can be useful hypotheses for interpretation. They are not automatically findings.
Interpretation requires researchers to connect evidence with the research design, theoretical framework, measurement, context, uncertainty, alternative explanations, limitations, and existing scholarship.
AI may expand the set of interpretations you consider. It should not decide which interpretation the evidence warrants while you merely copy the explanation.
Do Not Delegate the Final Conclusions
The conclusion is where researchers state what they believe their evidence permits them to claim.
That requires judgment about uncertainty, limitations, generalizability, practical significance, causal language, competing explanations, and the difference between what the study observed and what researchers infer from those observations.
A generative system can draft a conclusion from supplied material. But if the researchers do not personally determine whether every substantive conclusion follows from the evidence, the most visible claims of the study have effectively been delegated.
This is also one reason generative AI can make weak research look more rigorous than it is. Polished language can conceal an inferential leap rather than repair it.
Do Not Delegate Final Responsibility for Citations
AI can help format references, identify missing bibliographic fields, locate possible sources, or compare citation information.
The authors remain responsible for ensuring that cited sources exist, bibliographic details are correct, quotations are accurate, and references actually support the claims attached to them.
This is not clerical fussiness. Citations are part of the evidential architecture of scholarly writing. A perfectly formatted nonexistent paper is still nonexistent.
Do Not Delegate Peer-Review or Editorial Responsibilities Where Confidentiality or Policy Forbids It
Researchers may be tempted to upload a manuscript they are reviewing and ask AI to identify weaknesses or draft reviewer comments.
Whether that is permissible depends on the confidentiality obligations and AI policies governing the review. A manuscript supplied in confidence is not the reviewer's data to distribute freely.
The same principle applies to grant proposals, examination materials, unpublished manuscripts from collaborators, and other restricted documents.
Researchers should check the relevant policy before supplying confidential third-party material to an external AI service.
Do Not Delegate Decisions You Lack the Expertise to Evaluate
This is perhaps the most important practical boundary.
AI can make advanced work accessible. A researcher with little programming experience can generate complex code. Someone unfamiliar with a statistical method can obtain a sophisticated interpretation. A novice in a neighboring field can produce technically fluent prose within seconds.
Accessibility is valuable. It also creates an evaluation problem.
If you lack enough expertise to recognize whether a consequential output is wrong, your approval is weak evidence of successful human oversight.
UNESCO's guidance stresses capacity building so that researchers can understand and critically evaluate generative AI outputs rather than surrendering human agency to the system.
Use AI to help learn what you need to understand. Do not use its fluency to bypass the need for that understanding.
Do Not Delegate Responsibility for the Final Research Record
Research eventually becomes part of a record: a thesis, article, dataset, report, conference paper, preprint, software package, policy recommendation, or other scholarly output.
Someone must stand behind what that record claims.
The European Commission's living guidelines center accountability, transparency, responsibility, and research integrity in the use of generative AI. The updated 2026 version retains those principles as AI capabilities continue to change.
AI cannot take responsibility for correcting a paper, responding to an integrity inquiry, explaining a methodological choice to a reviewer, protecting a participant, or defending a conclusion against the underlying evidence.
That responsibility remains human.
A Useful Boundary Is the “Defense Test”
Before fully delegating a research task, imagine that a skeptical expert asks you to defend the result without access to the AI conversation.
The defense test
Can you explain why this research question matters?
If not, AI has probably replaced rather than assisted your scholarly framing.
Can you explain why this method is appropriate?
If not, generated methodological language is substituting for methodological understanding.
Can you show where this factual claim comes from?
If not, the generated answer has not become evidence.
Can you explain why this interpretation follows from your data?
If not, the AI may be supplying the study's reasoning rather than assisting it.
Can you accept responsibility if this output is wrong?
If not, do not make the research depend on the output.
This test is deliberately demanding. Research has always required scholars to defend their choices. Generative AI changes how easily plausible choices can be produced, not the underlying need to justify them.
