A researcher asks an AI tool to improve the clarity of a paragraph they already wrote. Another asks it to suggest keywords for a database search. A third uploads interview transcripts and asks for themes. A fourth asks it to generate a literature review with references. All four are “using AI,” but the ethical and academic risks are very different.

That is why blanket statements such as “AI is cheating” or “AI is just another tool” are not enough for serious research. Generative AI can help with brainstorming, organisation, language, coding, data workflows, literature mapping, and accessibility. It can also fabricate references, misrepresent evidence, expose confidential data, reproduce bias, obscure authorship, and tempt researchers to replace analysis with plausible-sounding text.

Current publishing guidance reflects this more nuanced position. The International Committee of Medical Journal Editors (ICMJE), updated in January 2026, says AI-assisted technologies should not be listed as authors, humans remain responsible for accuracy and integrity, and AI use should be disclosed where applicable. Elsevier’s journal policy, updated in 2026, likewise allows supportive AI use but requires human oversight, source verification, transparency, privacy protection, and disclosure for substantive use. UNESCO’s guidance for generative AI in education and research takes a human-centred approach, emphasising agency, privacy, accountability, and ethical validation.

The most important lesson is that there is no single universal rulebook. Universities, supervisors, ethics committees, funders, journals, and disciplines can impose different requirements. A use that one journal accepts with disclosure may be restricted by another institution. The first line, therefore, is always local policy.

This guide offers a practical framework for deciding where AI can support research, where caution is required, and where the use becomes academically or ethically difficult.

The Five Questions to Ask Before Using AI in Research

1. Is this use permitted by my university, supervisor, ethics approval, funder, or target journal?
2. Am I giving the tool confidential, unpublished, copyrighted, personal, sensitive, or restricted information?
3. Is the AI supporting my thinking, or replacing a judgement that I am expected to make myself?
4. Can I independently verify the facts, references, calculations, coding, interpretation, or output?
5. Would I be comfortable describing exactly how I used the tool to a supervisor, examiner, editor, participant, or reader?

A Practical Green / Caution / Red-Line Framework

Important: these zones are a practical risk framework, not a substitute for the specific rules that apply to your course, institution, ethics approval, funder, or journal.

1. Brainstorming Research Ideas: Helpful, but Do Not Let AI Invent the Research Problem

AI can be useful at the early idea stage. You might ask for alternative ways to phrase a research question, possible stakeholder groups, related concepts, or search terms. Used this way, the tool can act like a brainstorming partner.

The line appears when the researcher accepts the output without checking whether the proposed gap is real. A language model can confidently suggest that “few studies have examined X in Y context” even when dozens of studies already exist. Research gaps must come from actual literature, not from the model’s impression of a field.

A safer workflow is to use AI to expand possibilities, then verify those possibilities in scholarly databases and primary sources. The final research problem, rationale, and contribution should emerge from your reading and judgement.

2. Literature Searching: Use AI to Generate Search Ideas, Not as a Substitute for Databases

AI tools can help generate synonyms, Boolean terms, related constructs, spelling variants, or keywords for databases such as Scopus, Web of Science, PubMed, Google Scholar, or discipline-specific indexes. That can save time when a topic uses several different labels.

However, an AI answer is not a transparent systematic search. It may omit important papers, favour highly visible sources, misunderstand inclusion criteria, or provide references that do not exist. If your methodology requires a reproducible literature search, you still need to document databases, dates, search strings, filters, screening procedures, and eligibility criteria.

For a systematic or scoping review, AI may support parts of workflow only where the protocol, software, ethics, and reporting requirements allow it. The researcher remains responsible for ensuring that screening and extraction decisions are reproducible and defensible.

3. AI-Generated References: Never Trust Them Without Verification

Reference fabrication is one of the clearest practical risks. Generative models can produce a perfectly formatted citation containing a plausible author, title, journal, year, volume, and DOI – while the article itself does not exist.

Elsevier’s current policy explicitly warns that AI-generated references can be incorrect or fabricated and places responsibility on authors to verify sources. ICMJE likewise states that humans must ensure appropriate attribution and full citations.

The safe rule is simple: never include a reference merely because AI produced it. Open the source. Confirm the title, authors, journal, year, pages or article number, DOI or URL, and most importantly, confirm that the source actually supports the claim you attach to it.

4. Summarising Articles: Useful for Orientation, Risky for Evidence Extraction

If you already possess an article and the platform permits the upload, AI can help explain difficult passages or create a preliminary summary. This may be useful for orientation, especially when reading outside your strongest field.

But summaries can flatten nuance. A model may confuse association with causation, overlook subgroup findings, ignore limitations, or overstate conclusions. It may also merge what the authors found with its own background knowledge.

For research writing, important claims should therefore be checked against the original paper. If you are building a literature matrix, extract the study design, sample, measures, findings, limitations, and context from the paper itself. Treat the AI summary as navigation, not as the final record of evidence.

5. Writing Assistance: Editing Is Different From Outsourcing Authorship

Many policies distinguish low-level language assistance from substantive content generation. Elsevier says basic spelling, grammar, and punctuation checks do not require its generative-AI declaration, but substantive changes to sentence structure or organisation should be disclosed. ICMJE similarly recognises that grammar correction is generally viewed differently from content generation.

The ethical difference is authorship. If you write the argument and use AI to tighten wording, the intellectual contribution remains visibly yours. If you ask AI to produce your literature synthesis, discussion, limitations, and conclusion from a few bullet points, the tool is now performing much more of the scholarly work.

A useful test is whether you can defend every sentence without referring back to the prompt. If an examiner asks why a paragraph makes a particular claim, you should be able to explain the evidence and reasoning yourself.

6. Paraphrasing: AI Does Not Make Borrowed Ideas Yours

Students sometimes use AI as a paraphrasing machine, assuming that changing the wording removes the need to cite. It does not. Citation is required because the idea, evidence, theory, or finding comes from someone else, not because the wording looks similar.

AI-assisted paraphrasing can also introduce errors by changing the strength or meaning of a claim. A cautious original statement such as “was associated with” may become “caused,” or “may indicate” may become “demonstrates.”

If you use a tool to improve phrasing, compare the result with the original source, preserve the original meaning, and retain the citation. Do not use AI to disguise copying.

7. Translation: Useful, but Scientific Meaning Still Needs Human Checking

Translation is a legitimate area where AI can improve access, but technical and academic language creates risks. Elsevier advises that AI-assisted translations should be checked by a person who understands both languages and the scientific content, and that substantive use should be disclosed under its policy.

A translation can alter terminology, remove uncertainty, or choose a word that changes a construct’s meaning. This matters particularly for participant materials, validated questionnaires, quotations, qualitative data, and cross-cultural research.

If a translated instrument is central to data collection, the researcher may need a formal translation and validation process rather than simply accepting an AI output.

8. Coding Assistance: AI Can Help Write Code, but It Cannot Validate Your Analysis for You

AI can generate R, Python, SPSS syntax, SQL queries, or other research code. This can be useful for debugging, learning syntax, or drafting a routine workflow. Elsevier’s current guidance says AI-assisted code used in research should be described in the Methods section and remains subject to the same validity and reproducibility standards as any other code.

Generated code can run without being correct. It may use the wrong statistical test, mishandle missing values, select an inappropriate package, leak information across train/test sets, or silently transform variables in a way you did not intend.

You should understand what the code does, test it on known cases, inspect outputs, preserve versioned scripts, document packages and settings, and if possible have important analysis reviewed by someone competent in the method.

9. Quantitative Analysis: The Model Cannot Choose Your Methodology Responsibly

An AI tool can explain when a t-test or regression might be used, suggest diagnostics, or help interpret output. The risk appears when the researcher asks, “Here is my dataset – find something significant.” That reverses good research design.

Analysis should follow the research question, study design, measurement level, assumptions, sampling plan, and pre-specified methods where applicable. AI can support the execution, but the researcher must justify the analytical choice.

Never report a statistical interpretation because the tool said it was significant. Verify the calculation, assumptions, confidence intervals, effect sizes, coding decisions, model specification, and the exact output from the software used.

10. Qualitative Analysis: Assistance Is Possible, but Meaning and Context Are the Research

Qualitative analysis is particularly sensitive because interpretation is not a mechanical afterthought. Coding, reflexivity, context, participant language, contradiction, and the researcher’s theoretical position all shape the analysis.

AI can help organise text, test candidate code labels, or compare coding ideas if policy and confidentiality rules allow it. But asking a model to “analyse these interviews and give me five themes” can hide how the themes were produced and may erase minority or contradictory voices.

If AI is involved in qualitative analysis, document exactly how. Was it used to suggest candidate codes? To compare with human coding? To cluster existing codes? To draft descriptions after the human analysis? The Methods section should make the role visible enough that another researcher can understand the process.

11. Confidentiality: Do Not Upload What You Do Not Have the Right to Share

This is one of the most important lines. Unpublished manuscripts, interview transcripts, patient information, student records, confidential company data, grant proposals, identifiable participant data, and proprietary datasets may be subject to ethics approvals, contracts, data-protection law, or confidentiality agreements.

Elsevier warns that AI tools may process, store, reuse, or train on uploaded content and tells researchers to check the tool’s terms before sharing unpublished or personally identifiable material. ICMJE similarly warns that manuscripts submitted for review are privileged communications and should not be uploaded into AI systems where confidentiality cannot be assured.

The principle is broader than publishing. Before uploading research material, ask whether the participant consent form, ethics approval, data-management plan, contract, or institutional policy allows transfer to that provider. If you are unsure, do not upload the data until you have approval.

12. Peer Review and Grant Review: Confidential Material Is a Special Case

Researchers sometimes assume they can upload a paper they are reviewing and ask AI to identify weaknesses. That can violate confidentiality. ICMJE says reviewers must follow journal policy and maintain manuscript confidentiality. NIH goes further for its own peer-review process: NIH prohibits peer reviewers from using generative AI to analyse or formulate critiques of grant applications because uploading the content would violate confidentiality requirements.

This is a useful general lesson. Peer review is not your content to share. Even if AI would make the review faster, the research belongs to someone else and has been entrusted to you under specific conditions.

When reviewing manuscripts, grants, theses, confidential reports, or unpublished proposals, treat the relevant review policy as controlling.

13. Data Fabrication and Synthetic Output: Never Blur the Boundary

Generating fake participants, interviews, survey responses, laboratory observations, quotes, or results and presenting them as collected evidence is research misconduct, regardless of whether the fabrication was produced manually or by AI.

Synthetic data can have legitimate research uses – for simulation, method development, privacy-preserving experimentation, training, or testing – but it must be clearly identified as synthetic and methodologically justified. It cannot silently replace empirical data that your study claims to have collected.

Similarly, AI-generated quotations should never be presented as participant quotations. A plausible sentence is not evidence.

14. Images and Figures: Ask Whether the Image Is Evidence, Illustration, or Analysis

AI-generated images may be acceptable for some explanatory or illustrative purposes, depending on the publisher and rights involved. They are much more problematic when the image is presented as research evidence.

A generated microscopy image, clinical image, field photograph, or experimental observation cannot stand in for data that were never collected. When AI is used to create, alter, segment, enhance, or analyse research images, the method and extent of manipulation should be transparent and consistent with discipline and journal rules.

Researchers should also check licensing and copyright terms. An image generator’s availability does not automatically mean every output can be published in every context.

15. Authorship: AI Cannot Take Responsibility for the Paper

ICMJE and Elsevier are explicit that AI tools should not be listed as authors. Authorship requires accountability: approving the final work, taking responsibility for accuracy and integrity, addressing questions, and meeting ethical obligations.

An AI system cannot accept those responsibilities. If it generated wording, code, images, summaries, or analytic assistance, the human authors remain responsible for the result.

This also means “the AI made the mistake” is not a defence. Once material appears in your thesis, paper, report, or proposal, you own the responsibility for checking it.

16. Disclosure: What Should You Say?

Disclosure requirements differ. Some institutions require a declaration for any generative-AI use. Some journals exempt basic spelling or grammar checks but require disclosure of substantive assistance. Some disciplines require AI used in data analysis or figure generation to be described in the Methods section.

A good disclosure identifies the tool, purpose, and human oversight without exaggerating what the tool did. If the AI affected the research method, explain it where methods are reported. If it assisted manuscript preparation, follow the journal’s declaration instructions.

This is an example only. Use the wording and location required by your university, journal, funder, or research protocol.

17. When AI Use Should Appear in the Methods Section

If AI is part of how the research was conducted, it is not merely a writing aid. Examples include AI-assisted coding, image classification, automated transcription, data extraction, model-assisted screening, generation of synthetic comparison data, or use of an LLM as part of an intervention.

In those cases, reproducibility matters. Record the tool or model, version or access date where relevant, prompts or procedures where methodologically important, parameters, input data, human-review steps, and validation approach. If outputs can change over time, explain how you preserved enough information for another researcher to understand what happened.

The exact level of detail depends on discipline and study design, but “we used AI” is not a sufficient method.

18. Bias and Homogenisation: AI Can Make Research Sound More Certain and More Similar

Generative models learn patterns from large collections of existing material. Their outputs may reproduce dominant viewpoints, underrepresent less visible communities, or produce polished answers that flatten disagreement.

UNESCO’s guidance stresses human agency, inclusion, equity, cultural and linguistic diversity, and plural viewpoints. These issues matter in research because the most probable answer is not necessarily the most accurate, locally relevant, or theoretically interesting answer.

When using AI to brainstorm interpretations or summarise a field, deliberately look for missing perspectives. Ask which populations, regions, languages, methods, or theoretical traditions are absent. Then investigate those gaps through actual scholarship.

19. The Verification Rule: Every AI Output Needs a Different Kind of Check

20. A Practical “Would I Still Own This?” Test

Before keeping AI-generated material, close the tool and ask whether you can explain the reasoning yourself. Could you defend the methodological choice? Could you find the original source? Could you reproduce the analysis? Could you explain why one theme was retained and another rejected?

If the answer is no, the tool has moved from assistance into substitution. The solution is not necessarily to stop using AI; it is to return to the underlying task until the intellectual ownership is genuinely yours.

Common AI-in-Research Mistakes

• Assuming a polished answer is a correct answer. Fluency is not evidence of factual accuracy.
• Copying references without checking them. Fabricated citations can undermine an entire literature review.
• Uploading participant data to a public tool. This may violate consent, ethics approval, privacy, or contractual obligations.
• Using AI to create a research gap. The gap must be demonstrated through literature, not generated from probability.
• Letting AI choose a statistical method after seeing the data. Analysis should follow design and research questions rather than fishing for significance.
• Treating generated themes as qualitative analysis. Interpretation requires context, reflexivity, and an auditable analytical process.
• Hiding AI use because the output was edited. Editing does not automatically remove a disclosure obligation.
• Listing AI as an author. Major publishing guidance rejects this because AI cannot take accountability.
• Assuming one journal’s policy applies everywhere. Rules vary across institutions, publishers, courses, disciplines, and funders.
• Failing to keep a record of important AI use. Documentation becomes difficult when disclosure or methods reporting is required later.

A Researcher’s AI Decision Checklist

• ☐ I checked my institution, course, supervisor, ethics, funder, and target-journal rules.
• ☐ I know whether the AI provider stores or reuses my inputs.
• ☐ I have not uploaded confidential or restricted information without permission.
• ☐ I can explain exactly what the AI contributed.
• ☐ I independently verified all factual claims and references.
• ☐ I understand and can reproduce any code or analysis I kept.
• ☐ I did not use AI to fabricate data, quotations, participants, or observations.
• ☐ I preserved citations for ideas that came from real sources.
• ☐ I documented substantive AI use during the research process.
• ☐ I disclosed AI use where the applicable policy requires it.
• ☐ The final interpretation and argument are mine.
• ☐ I am willing to take full responsibility for every sentence, table, figure, result, and citation.

A Simple Traffic-Light Example

Green-ish: language polish. You write a discussion paragraph, ask AI to identify unclear sentences, accept only wording changes you understand, verify that no meaning changed, and disclosure rules say basic grammar editing does not need declaration.

Yellow: literature synthesis support. You provide your own verified notes from 20 papers and ask AI to suggest possible thematic groupings. You then return to the papers, make the grouping yourself, document the assistance, and disclose if required.

Yellow: coding support. You ask AI to draft R code for a regression specified in your protocol. You inspect the code, test it, verify assumptions and outputs, preserve the script, and report the AI role if applicable.

Red: invented references. You ask AI for ten recent studies and paste the citations into your thesis without opening them.

Red: confidential peer review. You upload an unpublished manuscript or grant proposal you were asked to review into a public AI chatbot.

Red: fabricated qualitative evidence. You ask AI to create plausible interview quotations to fill gaps in your findings.

Useful Backlinks & Current Guidance

• ICMJE – Use of Artificial Intelligence in Publishing – Updated 2026 guidance on AI use by authors, reviewers, and editors, including authorship, disclosure, accuracy, and confidentiality.
• Elsevier – Generative AI Policies for Journals – Current policy covering manuscript preparation, disclosure, methods use, AI-assisted code, privacy, references, authorship, and confidential manuscripts.
• UNESCO – Guidance for Generative AI in Education and Research – Human-centred guidance emphasising agency, ethics, privacy, inclusion, validation, and responsible use in education and research.
• NIH – Generative AI and Peer Review Confidentiality – NIH policy prohibiting generative AI use for analysing and formulating critiques of NIH grant applications because of confidentiality requirements.
• KMAS – Academic support for research planning, literature review structure, referencing, formatting, methodological clarity, and responsible academic workflows.

Final Thought: AI Can Support Research, but It Cannot Carry Your Responsibility

The safest way to think about AI in research is not as a replacement researcher, but as a tool operating inside a human-controlled research process. It can accelerate routine work, expose alternative ideas, improve language, help debug code, and support organisation. Those benefits are real.

But the closer the task moves toward evidence selection, interpretation, authorship, confidential data, methodological judgement, or claims that readers must trust, the stronger the need for human control, verification, documentation, and disclosure.

The line is crossed when convenience begins to hide accountability: when fabricated references are accepted, when confidential material is uploaded without permission, when AI-generated analysis is presented as the researcher’s own reasoning, or when the tool is used to create evidence that never existed.

Good research has always required judgement. AI changes the tools available to researchers, but it does not change who is responsible for the truthfulness, transparency, reproducibility, and integrity of the final work.