Using AI in Research: Where Assistance Ends and Responsibility Begins

Artificial intelligence can help researchers work faster. But can it help them produce research they can actually defend?

A postgraduate student is preparing a dissertation.

The literature review is taking longer than expected. There are dozens of articles to examine, several complicated concepts to understand and an approaching submission deadline.

The student opens an AI assistant and asks:

“Find ten important studies on my research topic and summarise their findings.”

Within seconds, the system produces a convincing response.

There are author names, publication years, journal titles, methodological descriptions and concise summaries of findings.

The student is relieved.

But an important question remains unanswered:

How much of this information has actually been verified?

The references may be genuine, inaccurate or entirely fabricated. The summaries may overlook qualifications in the original studies. The apparent agreement between researchers may be an artefact of the AI system’s response rather than a conclusion supported by the literature.

The technology has produced something that resembles scholarly work.

Whether it constitutes trustworthy research depends on what the researcher does next.

This distinction is central to the responsible use of artificial intelligence in academic research.

AI can assist with research activities. It cannot assume responsibility for the credibility of the resulting research.

1. The wrong question: Is using AI in research acceptable?

Discussions about artificial intelligence in academic research frequently begin with a simple question:

Is using AI allowed?

The question matters because universities, supervisors, publishers and funding bodies establish rules governing the use of AI.

But it is not sufficient.

AI is not one activity, and research is not one activity.

Using an AI tool to suggest search keywords is different from asking it to invent a literature review.

Using it to improve the grammar of an author-written paragraph is different from submitting an argument whose reasoning the author does not understand.

Using it to generate sample programming code is different from trusting an untested analysis of confidential research data.

The responsible question therefore has several parts:

• What research task is AI performing?
• Is that use permitted by the applicable rules?
• What information is being shared with the system?
• How will its output be independently evaluated?
• What intellectual decisions remain with the researcher?
• What use must be documented or disclosed?

These questions help distinguish assistance from inappropriate delegation.

AI use should be evaluated in relation to the task, risks, evidence and institutional context—not merely the presence or absence of a particular tool.

2. AI can generate answers without establishing their truth

Generative AI systems can produce coherent explanations, plausible summaries, programming code and well-structured academic language.

Their fluency is useful, but it creates a distinctive risk.

A response may sound authoritative even when its underlying claims are incorrect, incomplete or misleading.

Consider a researcher exploring the relationship between social media use and academic performance.

An AI assistant might provide several seemingly relevant publications. Some may be real. Others may contain incorrect journal titles, mismatched author names or invented findings.

The researcher cannot establish the reliability of those claims by asking the same system whether they are accurate.

A credible verification process requires consulting appropriate original sources.

For scholarly references, the researcher should examine the actual publication record, verify its bibliographic details and read the relevant content.

For quantitative claims, the researcher should inspect the original dataset, definitions, calculations and analytical assumptions.

For quotations, the wording and context must be checked against the source.

The central principle is straightforward:

An AI-generated claim is a proposition to examine, not evidence merely because it has been generated.

Worked example: a literature search

A student asks an AI assistant to identify studies supporting a particular hypothesis.

The system produces a list of ten papers.

An irresponsible approach would be to copy the references into the dissertation and reproduce the summaries without checking them.

A more defensible approach is to use the suggestions as preliminary leads, search scholarly databases, retrieve the actual publications and evaluate their relevance independently.

Even when every suggested publication is genuine, the list is not automatically comprehensive or unbiased.

Important contradictory studies may be missing.

The researcher must therefore develop and document a search strategy appropriate to the review’s purpose.

SRS Research Note

Finding literature and evaluating literature are different intellectual tasks.

AI may help with discovery, but a trustworthy literature review still requires critical engagement with the underlying scholarship.

3. Assistance becomes problematic when judgment is outsourced

Research depends on decisions that cannot be justified solely by reference to a tool’s recommendation.

A researcher must decide which questions are meaningful, which concepts require clarification, which methods are appropriate and which conclusions the evidence can support.

AI may help explore alternatives.

But those alternatives must be assessed through methodological reasoning.

Imagine a researcher asks an AI assistant:

“Which statistical test should I use for my data?”

The system recommends a regression model.

That recommendation may be useful as a starting point.

However, the researcher must still examine the research question, variable definitions, study design, sample characteristics, statistical assumptions and limitations of the proposed analysis.

Using the recommended method without understanding these issues transfers an essential research decision to a system that cannot take responsibility for its consequences.

The same principle applies to qualitative research.

An AI assistant may propose possible themes in interview material, but researchers must assess whether those themes are supported by the actual data and consistent with their analytical approach.

In some qualitative methodologies, researcher interpretation and reflexivity are central to the process. Those intellectual responsibilities cannot be replaced simply by receiving automated labels.

A useful distinction is this:

AI assistance offers material that the researcher can evaluate. Outsourcing judgment means allowing the system’s response to substitute for that evaluation.

4. Worked example: AI-generated statistical code

Consider a doctoral researcher examining the relationship between students’ weekly study time and examination results.

The researcher knows which variables are relevant but wants help writing code to perform an analysis.

An AI assistant generates a script.

The code appears technically correct and runs without producing an error.

Does that mean the analysis is valid?

Not necessarily.

A script may execute successfully while using inappropriate assumptions, mishandling missing values, selecting the wrong variables or interpreting an association as evidence of causation.

The researcher should therefore review the script line by line, test it with known or synthetic examples where appropriate, inspect the output and check whether the statistical method is suitable for the study.

The researcher should also document any material AI contribution according to the project’s reporting requirements.

A defensible workflow might look like this:

1. The researcher specifies the intended analytical task.
2. AI suggests possible code or an implementation.
3. The researcher inspects and tests the code.
4. The analytical assumptions are evaluated.
5. Results are independently checked.
6. The final procedure is documented and appropriately disclosed.

The key issue is not whether AI helped write the code.

It is whether the researcher can explain, reproduce and defend the analysis.

5. Confidential research data require special protection

Many research activities involve information that cannot be freely shared.

Examples include interview transcripts, identifiable survey responses, medical records, unpublished research manuscripts, restricted institutional documents and commercially sensitive datasets.

Uploading such material to an AI service may involve processing, storage or access arrangements that conflict with confidentiality obligations, participant consent or institutional requirements.

A researcher should not assume that a tool is suitable for confidential material simply because it is widely available.

Before using AI with research data, the researcher should establish:

• Whether the proposed processing is permitted by the research ethics approval and consent arrangements.
• Whether institutional policies authorise the particular tool.
• What information the service collects, retains or shares.
• Whether suitable contractual and technical protections exist.
• Whether identifiable information can be avoided altogether.

Removing participant names may not be sufficient to prevent re-identification, especially when transcripts contain distinctive personal details.

Worked example: qualitative interviews

A researcher has interviewed postgraduate students about experiences of academic supervision.

The transcripts contain sensitive descriptions of supervisors, departments and personal circumstances.

The researcher wants an AI assistant to generate an initial thematic summary.

An unsafe approach would be to upload the full transcripts to an unapproved public service without examining confidentiality and consent requirements.

A safer preliminary exercise could involve experimenting with fabricated interview excerpts to understand the tool’s capabilities.

For actual participant data, the researcher must use processes authorised by the relevant ethical, legal and institutional arrangements.

This may mean using approved infrastructure—or deciding not to use AI for that task.

Convenience does not override confidentiality.

6. Bias and omissions are research problems, not merely technical imperfections

Even when an AI-generated response contains no obvious factual errors, it may still provide a distorted account of a subject.

For example, an AI-assisted search may disproportionately surface certain languages, regions, disciplines or publication traditions.

A generated explanation may present one theoretical perspective as though it were universally accepted.

A summary may omit findings that contradict the apparent dominant position.

Researchers should therefore consider not only whether individual statements are accurate, but also whether the overall representation is sufficiently balanced and relevant.

This is particularly important in interdisciplinary and context-sensitive research.

A framework developed primarily from one cultural or institutional setting should not automatically be treated as applicable everywhere.

The researcher must ask:

Whose evidence is represented? Whose perspectives are missing? What assumptions shape the explanation?

These are established concerns in critical research practice. AI makes them newly visible and sometimes harder to detect.

7. AI cannot be the accountable author of a research publication

Academic authorship is not simply the act of producing text.

It carries responsibility for the accuracy, integrity, originality and scholarly claims of the submitted work.

Organisations such as the International Committee of Medical Journal Editors (ICMJE) and the Committee on Publication Ethics (COPE) state that AI tools should not be listed as authors because they cannot fulfil those responsibilities.

Researchers remain accountable for material included in their work, including content produced with AI assistance.

That means authors must be able to verify sources, identify errors, address questions about methods and ensure that the final work complies with relevant publication requirements.

This responsibility cannot be transferred to a chatbot, software application or automated system.

The practical implication is important:

A researcher should not submit an argument, analysis or interpretation that they are unable to understand and defend.

8. Transparency: what should researchers disclose?

Transparency is an important component of responsible AI use.

But disclosure requirements are not identical across all universities, disciplines, journals and research activities.

Some publishers distinguish between language editing and substantive content generation.

Others specify where AI assistance must be reported, particularly when it contributes to analytical procedures, data processing or manuscript preparation.

For example, ICMJE guidance asks authors to disclose relevant AI-assisted technologies and explain how they were used. COPE also emphasises transparency about AI involvement in manuscript preparation and research activities.

Researchers should consult the policies applicable to the specific destination rather than assuming that one general disclosure statement satisfies every requirement.

A useful record of AI assistance may include:

• The tool and provider.
• The date or version, where relevant.
• The specific research task.
• The nature of the output used.
• The verification and revisions performed.
• Whether data or other research material were submitted.
• The disclosure required by the institution or publisher.

Illustrative disclosure statement

“During preparation of this manuscript, an AI-assisted tool was used to suggest grammatical and stylistic improvements to author-written text. The author reviewed and revised all suggested changes and takes full responsibility for the final content.”

For analytical assistance, a more specific statement may be necessary.

“AI-assisted software was used to generate preliminary code for the data-cleaning procedure. The researcher reviewed, tested and modified the code, independently verified the outputs and documented the final procedure in the methods section.”

These examples are illustrative. Researchers should adapt their disclosures to what actually occurred and to the requirements of the relevant university or publisher.

Transparency is not a substitute for validity.

Disclosing that AI generated an incorrect result does not make the result acceptable.

9. A practical framework for responsible AI use

Before using AI in a research task, researchers can apply a five-part decision framework.

1. Permission

Is AI use permitted for this task under the applicable institutional, ethical, funding and publishing requirements?

If not, do not use it.

2. Protection

Can the task be completed without exposing confidential, personal or restricted information?

If sensitive material is involved, verify whether the proposed process is authorised and appropriately protected.

3. Verification

Can the AI-generated output be independently checked against evidence, original sources, reproducible calculations or established methods?

If meaningful verification is impossible, the output may not be suitable for research use.

4. Judgment

Does the researcher retain control over the research question, methodological choices, interpretation and conclusions?

AI should not become an unquestioned authority.

5. Transparency

Can the researcher accurately record and disclose the assistance provided?

Material contributions should be documented in accordance with the relevant requirements.

SRS Responsible AI Framework

Permission → Protection → Verification → Judgment → Transparency

These questions should be considered before, during and after AI-assisted work.

A task may initially appear appropriate but become unsuitable if the tool requests restricted data, produces unverifiable references or cannot support a defensible analytical process.

10. The deeper question: what does it mean to conduct research?

AI introduces a useful challenge to the way we understand research itself.

If a system can produce a polished paragraph, a plausible literature review or a functioning statistical script within seconds, which parts of the research process remain distinctly intellectual?

The answer cannot simply be the physical act of writing or calculating.

Research involves asking questions whose answers matter.

It involves deciding what counts as appropriate evidence, recognising the limitations of that evidence, examining alternative explanations and defending conclusions against criticism.

These responsibilities remain essential whether a researcher uses a notebook, statistical software, a search engine or an AI assistant.

The tools may change.

The requirement for intellectual accountability does not disappear.

AI can be valuable when it helps researchers explore possibilities, clarify ideas, test approaches and perform appropriate tasks more effectively.

It becomes problematic when fluency replaces evidence, convenience overrides ethical obligations or the researcher stops evaluating the results.

The central distinction is therefore not between research conducted with AI and research conducted without it.

It is between research in which AI is used under responsible human judgment and research in which that judgment has been surrendered.

The most useful question a researcher can ask is not:

“Can AI do this for me?”

It is:

“If AI helps me do this, can I still explain, verify and take responsibility for the result?”

That is where responsible assistance begins.

And where the researcher’s responsibility must remain.

References and further reading

International Committee of Medical Journal Editors. (2026). Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals: Use of Artificial Intelligence in Publishing. https://www.icmje.org/recommendations/browse/artificial-intelligence/

Committee on Publication Ethics. (2023; position statement updated 2024). Authorship and AI tools. https://doi.org/10.24318/cCVRZBms

Miao, F., & Holmes, W. (2023). Guidance for Generative AI in Education and Research. UNESCO. https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research

Researchers should also consult the latest AI-use policies of their institution, funding body and intended publication venue.