1. Introduction
Many guides to generative AI in research are organised by tool. This arrangement has limitations: tools change frequently, their capabilities overlap considerably, and a list of tools says little about the more important decision of whether a given task can be delegated safely.
This guide is therefore organised by research task, and around a single principle that applies with different force to each.
2. The competence trap
One property of these tools is particularly relevant to novice researchers, and it explains why their guidance needs to differ from guidance for experienced researchers.
This runs counter to a common intuition. It may seem that beginners have the most to gain, because there is more that they do not yet know. In practice, beginners may carry the greatest risk, because the verification that makes these tools safe depends on the very expertise they are still developing.
This does not mean that novice researchers should avoid such tools. It suggests that the safest uses are those in which output can be checked against something other than the user's own judgement: a source, a dataset, a test that either passes or fails, or a supervisor. The recommendations that follow are based on this principle.
3. Uses by research task
3.1 Finding literature: orientation only
Appropriate uses: becoming oriented in an unfamiliar area, for example by asking about the main theoretical positions, the meaning of key terms, current debates, or the methods typical of a field. This can provide the vocabulary needed to search effectively.
Inappropriate uses: compiling a reference list. When asked for sources, language models may produce plausible authors, titles, journals, and years, and occasionally a well-formed DOI that resolves to nothing or to an unrelated work. Such references are difficult to identify by inspection, because they closely resemble genuine ones.
A safer alternative: use the model to generate search terms, and then search established databases such as Scopus, Web of Science, ERIC, PubMed, or Google Scholar. Some tools now search real indexes and return working links. These present a different level of risk, but the same principle applies: sources should be opened and read rather than cited from a generated summary.
As a rule, nothing should be cited that the researcher has not opened and read. This single practice may prevent the most damaging error available to a novice researcher.
3.2 Reading and understanding: a comparatively strong use
This is an area in which the potential value appears high and the risk comparatively low for new researchers, and it may be underused. Examples include:
- requesting a plainer explanation of a dense paragraph;
- asking what a statistical procedure does and what assumptions it makes, and then checking the answer against a textbook;
- asking for the distinction between two frequently confused concepts; and
- asking what question a paper set out to answer, and then reading the paper to assess whether the summary is accurate.
The risk is lower in these cases because the source is available for comparison. The explanation can be checked against the text being explained. This illustrates the general form of a safer use: the model interprets a document the user can consult, rather than asserting a fact the user cannot check.
Errors still occur. Models may smooth over genuine disagreement in a literature, presenting it as consensus, or report tentative findings as established. An explanation that seems tidier than the field itself warrants particular caution.
3.3 Developing a research question: with a supervisor involved
Appropriate uses: testing a proposed question, for example by asking what assumptions it makes, what would need to be true for it to be answerable, what objections a critical reviewer might raise, and which designs could address it. Used in this adversarial way, models can provide an inexpensive preview of objections that may later arise in review.
Weaker uses: identifying what is worth studying. A model has limited insight into what a field currently needs, what a supervisor's group is well placed to pursue, or what data can realistically be obtained. Questions it proposes tend to be plausible but generic, which may make them difficult to defend in a proposal.
3.4 Methods and analysis: always verify
Code is among the more productive uses, subject to a firm condition. Generated R, Python, or SPSS syntax is often correct and may be faster to produce than writing it manually, but it always requires verification. Useful checks include running the code on data with known results, confirming the sample size after each filtering step, and ensuring that the test matches the study design.
The more serious risk is not code that fails to run, which is readily noticed, but code that runs without error and computes the wrong quantity, for example by using the wrong variable, applying unintended listwise deletion, or relying on a default setting that does not match the analysis assumptions.
Choosing a statistical approach: a model's suggestion may serve as a starting point, but it should be confirmed against a methods text or with a statistician. Models may suggest tests that sound defensible but are inappropriate for the data structure; treating clustered data as independent is a recurring example.
Qualitative coding calls for particular care. A model may be useful for a first pass over a large corpus or as a second perspective on codes the researcher has developed. It is not, however, a substitute for close engagement with the data, which is central to qualitative analysis. Any such use should be documented in detail, including which suggestions were retained or rejected; see Declaring AI use in a methods section.
3.5 Writing: useful, with a caveat about skill development
Generally legitimate uses: editing for grammar and clarity, particularly for researchers writing in English as an additional language, for whom the cost of professional editing represents a genuine inequity that these tools may partly reduce. Other reasonable uses include tightening long paragraphs, suggesting a structure for a section, drafting a plain-language summary, and drafting responses to reviewers.
Uses best avoided: delegating interpretation. The discussion section is where the author explains what the findings mean, and it is arguably the part of a paper most clearly the author's own. A model can produce a competent-sounding discussion of results it does not understand, and reviewers familiar with the field may recognise this.
There is also a longer-term consideration. Academic writing is not merely the transcription of finished thoughts; it is often the means by which thinking develops. Drafting is frequently where an author discovers that an argument does not hold, that two findings conflict, or that a claim cannot be justified. Delegating the drafting may therefore also delegate this discovery. Early in a career, the reasoning developed through drafting is a central part of what researchers are expected to acquire.
A practical compromise adopted by many experienced users is to write the first draft independently and then use a tool to improve it. This preserves the thinking while reducing the time spent on polishing.
3.6 Peer review: not appropriate
A manuscript under review should not be entered into a general-purpose AI tool. Doing so breaches the confidentiality accepted when agreeing to review, regardless of how the tool handles the text, and many publishers now prohibit it explicitly. Researchers who feel unqualified to review a manuscript can decline and explain why, which is a normal and respected response.
4. A verification routine
The recommendations above can be summarised as a single routine. Before any model output enters the work:
- Citations: open each one, confirm that it exists, and confirm that it supports the claim made. The second check is important, because attaching a genuine paper to a claim it does not support appears to be a more common error than citing a non-existent one.
- Factual claims: trace each to a source that has been read.
- Code: test it against a case with a known answer.
- Statistical choices: confirm them against a methods reference or with an expert.
- Interpretations: ensure that they are the researcher's own and could be defended without notes.
If a step cannot be carried out, this is not a reason to omit it. Rather, it suggests that the tool should not yet be used for that task.
5. Disclosing AI use
Journals increasingly require disclosure of AI use, and institutions may apply stricter rules to theses and coursework. In general, language editing is acknowledged in the acknowledgements, use that affects the evidence is reported in the methods, and no AI system is listed as an author. Template wording for common cases is provided in Declaring AI use in a methods section.
Keeping a running record of the tools used, and for what purpose, during the research is advisable. Reconstructing this information months later, for example at the proof stage, is difficult and prone to error.
6. Capacities worth developing independently
Some capacities cannot be delegated without losing much of their value, and they are worth naming explicitly because delegation has become so easy:
- Critical reading. Judging whether a study's design supports its conclusions is central to research training, and there appears to be no shortcut that develops this ability.
- Knowledge of the literature. This means familiarity with the arguments themselves, and with who disagrees with whom and why, rather than with summaries of them.
- Close engagement with data. Particularly in qualitative research, insight tends to emerge from sustained contact with the material.
- Constructing an argument. As noted above, writing is often how researchers discover whether they have an argument.
- Judging quality. This is a capacity that models do not possess in any meaningful sense, and one on which a research career ultimately depends.
7. A reasonable starting position
For researchers at the beginning of their careers, the following may serve as a defensible default:
- Lower-risk uses: explaining texts being read, drafting code that will be tested, and editing prose the researcher has written.
- Uses requiring verification: generating search terms, suggesting methods, and critiquing the researcher's own framing.
- Uses to avoid: generating citations, writing interpretations, and any involvement in peer review.
- In all cases: open every source, test every script, and keep a record of tools used.
8. Conclusion
None of these recommendations is unique to generative AI. They reflect the long-standing expectation that researchers understand, and can defend, everything published under their name. The technology has made this expectation easier to neglect, but no less important.
Related reading: Writing for international journals discusses the drafting conventions with which these tools are often used, and What SJR actually measures explains the journal metrics researchers encounter when choosing where to submit.