Workspace
Blog

How to Use AI in Research Without Losing Responsibility

5 Nov 2026·

A practical guide to responsible AI use across literature review, data analysis, writing, and publication—with six checks for accountable research.

A student asks an AI assistant to suggest search terms for a literature review. Another asks it to summarize twenty articles. A researcher uses it to write code, classify interviews, choose a statistical model, or draft a discussion. Each activity can be described as “using AI in research,” yet each affects the work differently. A search suggestion may only save time. An incorrect summary can distort the argument. A flawed analysis can change the result. The ethical question therefore concerns what the system did, what evidence supports its output, and who remains responsible for the claim.

Rules have not fully settled these questions. Guidance from the International Committee of Medical Journal Editors (ICMJE), the American Psychological Association (APA), Nature Portfolio, Elsevier, and the European Research Area (ERA) differs in scope and in what authors must declare. It nevertheless points repeatedly to transparency, human responsibility, checking AI output, and protection of confidential material (American Psychological Association, 2025; Elsevier, 2026; European Commission, Directorate-General for Research and Innovation, 2026; International Committee of Medical Journal Editors, n.d.; Nature Portfolio, n.d.). The rules of a course, institution, funder, or journal may add further limits. Students need to check their assignment rules; researchers need to check the requirements that govern their study and publication.

This essay draws on selected scholarly articles and current guidance to explain a workable approach. AI can assist with substantial research tasks, but its contribution should be traceable, its inputs authorized, and its consequential outputs checked by methods appropriate to the claim. People must retain the ability and willingness to explain, defend, and correct the work. These conditions apply to literature reviews and class projects as well as advanced research. They do not turn a well-worded AI answer into evidence by themselves.

What Counts as AI Assistance in Research?

AI assistance extends beyond drafting sentences. A system might propose keywords, screen search results, transcribe interviews, extract information from documents, generate code, suggest hypotheses, run simulations, interpret patterns, create figures, or prepare a public summary. A review of AI in scientific discovery describes applications from hypothesis generation through experimental design and data interpretation (Wang et al., 2023). More autonomous systems can combine several such steps. These possibilities explain why rules limited to “AI-written text” leave much of research practice unaddressed.

A useful starting distinction is between assistance and evidence. Asking AI for possible search terms can improve a search strategy; the answer is not evidence that the search found all relevant studies. Asking for a summary may help decide what to read; the summary is not a substitute for checking the article. Generated code can be useful; its output is not trustworthy until the researcher understands what the code did and tests it. AI-generated hypotheses are candidates, not findings. Research claims must be supported by sources, observations, analysis, or tests that can be examined independently of the assistant’s fluency.

This distinction also helps avoid an unhelpful debate about whether any AI use is inherently good or bad. Nature Portfolio’s current policy, for example, distinguishes low-risk assistance from uses that influence interpretation and from uses that replace accountable scholarly judgment (Nature Portfolio, n.d.). That is a publisher policy, not a universal rule, but the underlying idea is useful: the more an AI action can change a research claim or harm someone, the stronger its oversight must be.

The Shared Principles—and Their Limits

Across the selected guidance, four principles recur. First, humans remain accountable. ICMJE says AI tools cannot be authors because they cannot assume responsibility for accuracy and integrity; authors must review AI-assisted material and its citations (International Committee of Medical Journal Editors, n.d.). The ERA’s October 2026 guidelines likewise say researchers remain responsible for output produced with AI and must provide appropriate oversight and verification (European Commission, Directorate-General for Research and Innovation, 2026). A name on a manuscript therefore represents more than having pressed “submit.”

Second, relevant AI use must be transparent. ICMJE asks authors to describe AI use in both a cover letter and the submitted work. APA gives task-specific examples, including disclosure of AI-assisted code, analysis, outlines, and substantial editing. The ERA guidance emphasizes disclosure of substantial uses and what the tools contributed (American Psychological Association, 2025; European Commission, Directorate-General for Research and Innovation, 2026; International Committee of Medical Journal Editors, n.d.). The exact threshold and location differ. A student should not assume that a journal’s permission overrides an instructor’s prohibition, or that one publisher’s exception applies elsewhere. When unsure, record the use, consult the applicable rule, and ask the instructor or editor before submission if the rule remains unclear.

Third, outputs must be checked. This is especially important when the output looks convincing. In a 2023 study, Walters and Wilder tested GPT-3.5 and GPT-4 on generated literature reviews and found fabricated references as well as errors in references to real works (Walters & Wilder, 2023). Their results do not measure the reliability of every AI system today. They do establish why a plausible reference should be searched in a bibliographic database or publisher site and why the source itself should be read before it supports a claim. The PNAS statement on scientific integrity similarly calls for verification of AI-generated data, images, and inferences, and for documenting synthetic material so it is not confused with observation (Blau et al., 2024).

Fourth, privacy, consent, and rights constrain what researchers may enter into a tool. A participant’s interview, a patient’s record, an unpublished manuscript received for peer review, and a licensed document each have conditions attached to them. Removing a person’s name may not remove all identifying information. APA warns authors about exposing participant information to a provider, while the ERA guidance cautions against putting unpublished review materials into unsecured AI services (American Psychological Association, 2025; European Commission, Directorate-General for Research and Innovation, 2026). Researchers should check the approved data-management plan, consent, institutional agreement, and tool’s terms before uploading material. If permission or security is unclear, use a permitted alternative or leave the material out.

These principles are necessary, but none alone settles a case. An honest disclosure does not repair a privacy breach. A private log does not validate an analysis. Human authorship does not make a result credible if the authors cannot explain how it was produced.

A Practical Standard: Responsible Delegation

The central practice is to match oversight to the task and the claim. Before using an AI output, researchers should be able to answer six questions. The questions synthesize the guidance and scholarly arguments reviewed here; they are a proposed working standard, not a quotation from a single policy.

  1. Purpose and authority: What task is AI doing, and is that use allowed by the assignment, ethics approval, data agreement, funder, or publication venue? Set limits before starting. For example, “suggest additional keywords” is a different instruction from “decide which studies are credible.”
  2. Data and rights: What will enter the system? Check whether the material contains private, confidential, unpublished, or restricted information. Use an approved environment where required. Do not assume that a public chatbot is an appropriate place for research data.
  3. Traceability and disclosure: Record the tool and version if known, the task, the type of input, the output used, and important human changes. Disclose consequential use in the location required by the course or venue. A reader should be able to tell whether AI helped polish prose, selected evidence, produced analysis, or proposed an interpretation.
  4. Competent oversight: Can the responsible researcher or team explain the method’s purpose, assumptions, and likely failure modes? If the answer is no, obtain qualified help or limit the claim. A student who cannot explain an AI-generated statistical test should not present its result as their own understood analysis.
  5. Claim-specific verification: What independent check addresses the likely error? Verify a citation against the article, a coded category against source material, generated code against test cases and data, and a scientific inference against alternative explanations. Ask whether the check could actually reveal a mistake rather than merely asking the same AI system to approve itself.
  6. Human accountability: Who takes responsibility for the decision, the resulting claim, and correction if it proves wrong? Identify the person or team before publication or submission, not only after a problem is found (Blau et al., 2024; European Commission, Directorate-General for Research and Innovation, 2026).

The fourth question needs care. Research has always relied on distributed expertise. A biologist may trust a statistician’s specialized analysis, and a student may learn a method from a supervisor without being able to invent it independently. Hardwig’s account of epistemic dependence explains why no individual researcher can personally reproduce or fully verify every part of complex science (Hardwig, 1985, pp. 344–348). The requirement is therefore a credible route to evaluation within the responsible team, not that every author can perform every task unaided. Yet AI output does not become trustworthy merely because a human expert’s work sometimes is. Its performance, provenance, and possible failure modes still require checks.

Messeri and Crockett add a further warning: AI can make researchers feel that they understand more than they do, especially when they use it outside their expertise or mistake prediction for explanation (Messeri & Crockett, 2024, pp. 52–54). A useful test of understanding is whether a researcher can explain why a method fits the question, identify what would count against the result, and state what the result does not show. If those answers are unavailable, the output may remain a lead or a tentative candidate rather than a finding.

How the Standard Changes Daily Research Work

The checks are easiest to apply at the moment a task is delegated. The following practices are deliberately concrete; they should be adapted to the methods and rules of each discipline.

Research taskProductive AI roleResearcher’s check before using the output
Framing a questionGenerate possible angles or objectionsChoose the question for its scholarly and social significance; check feasibility and ethics
Searching literatureSuggest terms, databases, or related worksRun a documented search; verify records and look for missing perspectives
Reading and reviewingMake provisional summaries or comparison tablesRead sources needed for central claims; inspect quotations, methods, and disagreements
Collecting or coding dataTranscribe or suggest labels where permittedConfirm consent and access; inspect original material and assess errors across cases
Coding and analysisDraft scripts, suggest diagnostics, or explain outputRun the code, preserve inputs and versions, test examples, and review assumptions
Writing and figuresImprove clarity or propose a layoutCheck every claim, citation, and image provenance; disclose use as required
Sharing resultsDraft a plain-language explanationEnsure the public summary has the same limits and uncertainty as the research

For a literature review, begin with a question and a search plan that can be inspected. AI may propose search strings or identify a paper, but the researcher should verify the paper’s bibliographic record, obtain the text, and judge whether it supports the specific sentence being written. A real article can be cited inaccurately; a correct title does not make an AI summary accurate. The ERA guidance explicitly notes both invented citations and incorrect summaries of real works (European Commission, Directorate-General for Research and Innovation, 2026). Keep a record of search terms, dates, databases, inclusion decisions, and any AI-assisted screening if the review’s method requires reproducibility. For an ordinary student essay, a simpler reading log may suffice, but central claims still need source checks.

For data work, the first check is permission. If data come from people, consent and confidentiality apply to AI-assisted transcription, coding, and analysis just as they do to other processing. If an AI tool generates synthetic data, mark it as synthetic and do not present it as observation (Blau et al., 2024). When AI extracts information from documents or labels responses, examine a sample against the originals and look for systematic omissions or errors. A high overall agreement score can hide mistakes concentrated in a small group or unusual cases, so the sample and error inspection should fit the research question.

For code and statistical analysis, preserve the original data, the code, and a record of transformations. Run generated code on known or small test cases. Check that variables mean what the code assumes, exclusions are justified, and the analysis answers the stated question. Ask someone with relevant expertise if the model or interpretation exceeds the team’s competence. “The code ran” establishes only that it executed; it does not establish that the design, measure, or inference is valid. The same applies to a polished figure or a fluent explanation of a model.

For writing, use AI to improve readability without allowing it to invent evidence. Inspect all citations, quotations, numbers, and conclusions in the final text. If an AI tool wrote or substantially reshaped a passage, describe that role according to the relevant rule. Elsevier’s current journal policy, for example, distinguishes basic grammar checks from substantive editing and asks for an AI declaration for manuscript preparation; AI used in research methods belongs in the Methods section (Elsevier, 2026). That example illustrates why a generic “AI was used” sentence may be too vague and why the declaration should name the task and extent of use.

A short, honest declaration might read: “We used [tool and version, if known] to propose search terms and to edit prose for clarity. We verified all references against their original sources and made the final selection and interpretation ourselves.” If AI drafted code or classified data, the declaration should say so and describe the validation in the methods. A student might write an analogous note for an assignment, subject to course rules. The declaration should describe what actually happened; it should not promise checks that were not performed.

When AI Agents Perform Several Steps

An AI agent can connect search, document reading, coding, analysis, and drafting. The ethical difficulty is that mistakes in an early step can silently shape later steps. A false citation can enter a review; a mistaken data transformation can become a clean-looking table; an unsupported interpretation can be polished into a confident conclusion. The ERA’s fourth-version guidance explicitly discusses agentic AI and warns that researchers may lose control by delegating too many tasks while remaining responsible for validity (European Commission, Directorate-General for Research and Innovation, 2026).

For such a workflow, keep stage boundaries. Specify the question and permitted sources before the agent starts. Require it to show the references and data it used. Review the search set before synthesis, inspect the source passages behind important claims, approve the analysis plan before results are known, and test code and outputs before drafting conclusions. Keep logs or reproducible artifacts when feasible. A human checkpoint must be able to reject a result, change a method, or stop the workflow. Merely accepting a final report is too weak when the intervening work cannot be reconstructed.

There is an open question about which intellectual decisions should remain exclusively human. Branas and Levine argue for strong human control over research questions, final interpretations, and judgments of significance; a multi-perspective PNAS essay shows that scholars disagree about how far AI’s role in science should extend (Binz et al., 2025; Branas & Levine, 2026). A practical position for students and researchers is to let AI offer candidate questions or explanations while people decide which purposes are worth pursuing, what evidence is adequate, and what conclusion can responsibly be shared. This keeps the choice and defense of a scholarly claim visible even as AI becomes more capable.

Responsibility Is Also Institutional

An individual cannot solve every problem through careful prompting. Institutions need clear rules, suitable tools for sensitive work, training in verification, and channels for honest disclosure. The ERA guidelines ask research organisations to support these conditions and to encourage disclosure without automatically penalizing it (European Commission, Directorate-General for Research and Innovation, 2026). Instructors can make assignment rules explicit: what AI assistance is allowed, what learning the student must demonstrate, and how use should be reported. Editors and journals can ask for useful details rather than an unexplained yes-or-no declaration.

Research communities should also ask which questions AI makes easier to pursue and which it neglects. Messeri and Crockett warn that shared reliance on similar tools can narrow methods, questions, and viewpoints even while the volume of output rises (Messeri & Crockett, 2024). Ethical use therefore involves more than catching fabricated references. It includes retaining intellectual diversity, attending to people affected by the research, and resisting the temptation to treat easy-to-produce answers as the most important ones.

Putting the Checks to Work in Beeblio

Beeblio is an AI research workspace that can assist across a project’s steps while keeping the working files available for inspection. A researcher can use it to explore a question and search for literature, organize references, read project documents, plan a study, prepare data, run analyses, make figures, draft a report, and export the work. Its project workspace keeps sources, data, analysis, and reports together, so an AI-assisted answer can be checked against the materials and artifacts behind it.

That continuity is useful only when the researcher applies the same checks at each stage. Verify a suggested paper against its original source before citing it. Confirm that uploaded data and documents are permitted for the tools used. Inspect cleaning decisions and generated code, test analyses against the study design, and trace figures back to their inputs. Review drafts for unsupported claims and disclose substantial AI contributions under the relevant rules. Beeblio can help carry out and document these steps; it does not approve a study, establish consent, validate a result by itself, or assume responsibility for a published claim.

Conclusion

AI can help at many points in research, from finding sources to writing code and communicating results. The unsettled rules need not leave researchers without a standard for action. Before using an AI-assisted output as part of a claim, ask what the system did, whether its inputs were permitted, how the contribution will be disclosed, what independent check could reveal an error, and who will answer for the result. If those questions cannot be answered, revise the workflow or narrow the claim. The aim is research that remains understandable, checkable, and correctable even when some of its work is delegated.

Scope and AI-Use Disclosure

This essay selectively synthesizes the cited sources; it is not a systematic review or an empirical study. Policy descriptions were checked against public guidance available on 5 November 2026; readers should consult their own current course, institutional, and venue rules. AI assistance included source discovery, bibliography organization, outlining, and prose drafting and revision. The Beeblio section was added for this publication using the product documentation and reviewed against the essay’s responsibility standard. The publishing author is responsible for approving the final argument, source support, and declaration before submission or publication.

References

American Psychological Association. (2025). APA journals policy on generative AI: Additional guidance. https://www.apa.org/pubs/journals/resources/publishing-tips/policy-generative-ai

Binz, M., Alaniz, S., Roskies, A., Aczel, B., Bergstrom, C. T., Allen, C., Schad, D., Wulff, D., West, J. D., Zhang, Q., Shiffrin, R. M., Gershman, S. J., Popov, V., Bender, E. M., Marelli, M., Botvinick, M. M., Akata, Z., & Schulz, E. (2025). How should the advancement of large language models affect the practice of science? Proceedings of the National Academy of Sciences, 122(5), e2401227121. https://doi.org/10.1073/pnas.2401227121

Blau, W., Cerf, V. G., Enriquez, J., Francisco, J. S., Gasser, U., Gray, M. L., Greaves, M., Grosz, B. J., Jamieson, K. H., Haug, G. H., Hennessy, J. L., Horvitz, E., Kaiser, D. I., London, A. J., Lovell-Badge, R., McNutt, M. K., Minow, M., Mitchell, T. M., Ness, S., … Witherell, M. (2024). Protecting scientific integrity in an age of generative AI. Proceedings of the National Academy of Sciences, 121(22), e2407886121. https://doi.org/10.1073/pnas.2407886121

Branas, C. C., & Levine, B. L. (2026). AI and authorship: Norms and uses to preserve human-led science. PNAS Nexus, 5(9), pgag277. https://doi.org/10.1093/pnasnexus/pgag277

Elsevier. (2026). Generative AI policies for journals. https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals

European Commission, Directorate-General for Research and Innovation. (2026). Living guidelines on the responsible use of generative AI in research. https://research-and-innovation.ec.europa.eu/document/download/2b6cf7e5-36ac-41cb-aab5-0d32050143dc_en

Hardwig, J. (1985). Epistemic dependence. The Journal of Philosophy, 82(7), 335–349. https://doi.org/10.2307/2026523

International Committee of Medical Journal Editors. (n.d.). Use of AI by authors. Retrieved November 5, 2026, from https://icmje.org/recommendations/browse/artificial-intelligence/ai-use-by-authors.html

Messeri, L., & Crockett, M. J. (2024). Artificial intelligence and illusions of understanding in scientific research. Nature, 627, 49–58. https://doi.org/10.1038/s41586-024-07146-0

Nature Portfolio. (n.d.). Artificial intelligence (AI): A risk-assessment framework for responsible AI use in research publishing. Retrieved November 5, 2026, from https://www.nature.com/nature-portfolio/editorial-policies/ai

Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5

Wang, H., Fu, T., Du, Y., Gao, W., Huang, K., Liu, Z., Chandak, P., Liu, S., Van Katwyk, P., Deac, A., Anandkumar, A., Bergen, K., Gomes, C. P., Ho, S., Kohli, P., Lasenby, J., Leskovec, J., Liu, T.-Y., Manrai, A., … Zitnik, M. (2023). Scientific discovery in the age of artificial intelligence. Nature, 620, 47–60. https://doi.org/10.1038/s41586-023-06221-2

Beeblio Team

researchaiethics

The Premise

Why we're building a full-stack research workspace — and finishing the job the IDE started.