AI in Research: Literature Discovery, Verification and Transparent Disclosure

Separate assistance from evidence

AI may help generate search terms, organise notes or suggest alternative explanations. These functions can support research, but they do not establish the truth of a statement. A fluent summary is not evidence that the cited study exists, used the reported method or reached the described conclusion. Decide which tasks the tool may assist and which checks remain the researcher's responsibility.

The European Commission's living guidance addresses responsible generative AI use in research.^[1] A practical implication is to preserve an auditable connection between a claim and its original evidence. Do not allow a generated answer to become the only record of how a conclusion was reached.

Use AI to improve discovery carefully

Begin with a clearly framed question and a search plan appropriate to the discipline. An AI tool can suggest synonyms, related concepts or alternative terminology, which the researcher can test in relevant databases. Record the databases, search expressions, dates and selection decisions required by the chosen review method.

For a systematic review, convenience does not justify replacing a reproducible search with an undocumented chatbot conversation. If an AI-assisted screening method is used, describe the method, human checks and known limitations. Test it against a sample that includes difficult cases, not only obvious matches. Keep the protocol consistent or document changes transparently.

Verify references and interpretations

Open the original publication and confirm the title, authors, year and persistent identifier where available. Read the relevant section rather than relying solely on an abstract. Check whether the source is a preprint, final publication, correction or retracted item. Compare the generated interpretation with the actual population, methods, findings and limitations.

Maintain a claim-evidence note for consequential statements. It can contain the claim, source location, limitations and the researcher's interpretation. This is especially useful when a tool merges several studies into one apparently simple conclusion. Verify quotations word for word and avoid copying copyrighted material beyond permitted limits.

Protect data and unpublished work

Before using a service, determine whether the material includes participant information, confidential peer-review content, intellectual property or restricted partner data. Review the applicable institutional approval and contractual terms. Removing names may not eliminate re-identification risk when a dataset contains distinctive combinations of attributes.

UNESCO's education and research guidance discusses privacy and ethical considerations surrounding generative AI.^[2] Locally, define approved environments and prohibited inputs. Do not upload material merely because a tool offers a convenient interface. Where approval is uncertain, use synthetic examples or non-sensitive text until the responsible authority has assessed the proposed use.

Illustrative example: preparing a literature map

Imagine a doctoral researcher studying feedback practices. An AI assistant proposes search terms and lists several apparently relevant papers. The researcher verifies each entry in scholarly databases, finds one inaccurate citation and excludes it. They then read the remaining papers and create their own comparison of study settings and methods.

The AI tool helps organise a draft table, but the researcher checks every cell against the source. Their records distinguish suggested search terminology from source-based extraction. The final review explains the actual search and screening process rather than claiming that the AI-generated list was a comprehensive evidence search.

Document meaningful assistance

Keep proportionate records of the tool, version or access date, purpose, inputs permitted by policy and verification performed. Requirements differ across institutions, funders and publishers, so check the rules relevant to the intended output. A useful disclosure explains what the tool assisted with and which substantive decisions remained with the authors.

Avoid both extremes: concealing substantial assistance and adding a vague disclosure that says nothing about its role. Generated analysis code must be reviewed and tested like other research code. If an AI system is part of the research method itself, report enough detail for readers to assess the method and its limitations.

Retain intellectual ownership of the work

Researchers should be able to explain their questions, methods, analyses and conclusions independently of the tool. Build verification time into the project plan; apparent drafting speed can be offset by checking and correction. Discuss proposed use with supervisors and collaborators before it affects shared work.

AI assistance is most defensible when it improves a bounded task while leaving a clear evidence trail. The research record should show what was done, what was checked, what remains uncertain and who is accountable for the resulting claims.