A researcher asks an AI assistant for twelve recent papers on a narrow topic. The list arrives in seconds. Every title sounds plausible. Every author has the right kind of name. Each entry includes a journal, year and DOI.
R. Chen, L. Morales and T. Singh · Journal of Applied Learning Science · 2024
10.1234/jals.2024.1187The citation looks complete. The DOI has the right shape. The journal title sounds respectable. But a search of the journal, Crossref and the authors’ records finds nothing.
The AI did not retrieve a hidden paper. It assembled a citation that resembled the citations in its training data.
AI can suggest where to look. Only a source you can open, identify and inspect can become evidence.
This distinction sits at the centre of responsible AI-assisted research. A polished answer can help the work move faster, but polish is not provenance. The researcher still needs a path from every claim back to the underlying paper, dataset, observation or calculation.
First separate the work
Academic research is not one task
“Use AI for research” can describe very different activities. Each carries a different level of risk.
Framing
Clarifying the question, identifying variables, defining terms and exposing assumptions.
Discovery
Finding keywords, authors, journals, databases, related concepts and possible search paths.
Evidence review
Reading real papers, comparing methods, judging limitations and deciding what the evidence supports.
Analysis
Working with data, code, statistics, qualitative material, models and sensitivity checks.
Writing
Explaining the research question, method, evidence, uncertainty and contribution.
Accountability
Disclosing methods, protecting participants, preserving records and taking responsibility for the final work.
AI may assist with every stage, but it cannot take responsibility for any of them. It cannot confirm that consent was valid, that an exclusion criterion was applied consistently or that a conclusion fairly represents the data. Those remain human research duties.
Where AI can genuinely help
The strongest uses are usually bounded, inspectable and reversible. You can see what the system did, compare it with the original material and reject the result without damaging the study.
Generate search language
Ask for synonyms, older terminology, spelling variants, acronyms and related concepts before building a database query.
Map a field
Request a provisional list of debates, methods, schools of thought or commonly measured outcomes, then verify that map against the literature.
Explain difficult passages
Use AI to restate a statistical method, philosophical distinction or technical paragraph in simpler language while keeping the original open beside it.
Challenge a plan
Ask for missing assumptions, alternative explanations, confounders, edge cases and reasons a proposed design might fail.
Inspect code
Use AI to explain functions, identify obvious bugs, propose tests or translate a method into pseudocode before a researcher reviews and runs it.
Improve communication
Draft a plain-language summary, reorganize a dense paragraph or create questions for a presentation without changing the underlying result.
These uses save time because they support thinking around the research. They do not replace the evidence trail.
Literature discovery is not the same as literature evidence
An AI system may know that a topic is associated with certain authors, journals or phrases. It may also retrieve information from connected search tools. But the user still needs to know which mode is operating.
The model predicts a plausible answer
It may mention real work, combine details from different papers or fabricate a complete reference.
The system returns records from a source
The records are more traceable, but ranking, coverage, filters and database errors still need review.
Ask the tool directly:
- Are these references retrieved from a live database or generated from the model?
- Which database was searched?
- What date was the search run?
- What exact query and filters were used?
- Can you provide a stable DOI, PMID, accession number or publisher record?
If the system cannot answer those questions, treat the output as a list of leads—not as a bibliography.
How to verify every citation before using it
A citation is verified only when the researcher can establish that the work exists and that it supports the claim being made.
- 01
Find the authoritative record
Use the publisher, journal, Crossref, PubMed, a library catalogue, an institutional repository or another field-appropriate index.
- 02
Match the identity
Confirm the title, authors, journal or venue, year, volume, pages and persistent identifier.
- 03
Open the source
Do not cite a paper because its title sounds relevant. Read at least the abstract, method, results and limitations needed for your claim.
- 04
Check the claim against the paper
AI summaries often strengthen cautious findings. “Associated with” can become “causes,” and “in this sample” can disappear.
- 05
Check status and corrections
Look for retractions, expressions of concern, corrected versions, updated datasets or later replications.
The citation rule
Never cite the AI’s description of a paper when you can cite and inspect the paper itself.
Using AI to read papers without surrendering interpretation
Long papers are difficult to navigate, and AI can help readers enter them. The danger appears when a generated summary becomes a substitute for reading the parts that matter.
Ask for structured extraction rather than a general summary:
Then compare each extracted item with the paper. Pay special attention to tables, footnotes, appendices and supplementary files. They often contain qualifications that disappear in a short summary.
A summary can be accurate and still be insufficient
A model may correctly report the headline finding while omitting the weak sample, unusual measurement choice, failed robustness test or conflict between subgroup results. Research judgment depends on those details.
AI can assist with data and code, but reproducibility must survive
Generated code is useful when it remains inspectable. It becomes dangerous when the researcher cannot explain what it does or reproduce the output without the chat session.
- Review inputs, assumptions and expected outputs.
- Check package names, versions and licences.
- Remove secrets, identifiers and restricted data.
- Test on a small, non-sensitive sample.
- Compare results with a known case or independent method.
- Inspect warnings, missing values and excluded records.
- Save the code, environment and exact data version.
- Document every manual change after generation.
For statistical analysis, ask the AI to state the assumptions of the proposed test and the consequences when those assumptions fail. Do not accept a method merely because the code executes.
Running code proves that the computer accepted the instructions. It does not prove that the instructions answered the research question.
Do not upload confidential or unpublished material by habit
Research material may contain participant information, commercial secrets, unpublished findings, grant proposals, embargoed manuscripts or data governed by contracts and ethics approvals.
Before using an external AI service, establish:
Peer-review manuscripts require particular care. Nature Portfolio and ICMJE guidance warns that uploading confidential submitted manuscripts to external generative AI systems can violate the confidentiality expected in review.
When the rules are unclear, do not upload the material. Use an approved institutional tool, a properly governed local system or a de-identified synthetic example instead.
A safer AI-assisted research workflow
- 1
Define the research question yourself
Write the population, phenomenon, comparison, outcome, boundaries and purpose before asking the model to improve the wording.
- 2
Use AI to expand the search vocabulary
Generate synonyms and related concepts, then build and save the actual database query.
- 3
Search authoritative databases
Record the database, query, date, filters and number of results. Do not rely on an untraceable generated list.
- 4
Verify and read the sources
Confirm identifiers, inspect the paper and record why it was included or excluded.
- 5
Use AI for structured assistance
Request extraction, comparison, counterarguments or code explanations tied to material you can inspect.
- 6
Validate every substantive output
Check claims, calculations, code, quotations and interpretations against the source data and independent methods.
- 7
Preserve the record
Save prompts, outputs, tool versions, code, data versions and human edits needed to explain the final result.
- 8
Disclose according to policy
Describe material AI use in the manuscript, cover letter, methods, acknowledgements or other required location.
Build an audit trail that survives the chat window
A research process should remain understandable after the AI conversation is closed. Someone else should be able to see where an idea came from, what changed and how the final conclusion was reached.
| Research item | What to preserve | Why it matters |
|---|---|---|
| Literature search | Databases, queries, dates, filters and screening decisions | Shows how the evidence set was formed |
| AI interaction | Tool, version, prompt, output and date | Makes the assistance visible and reviewable |
| Data | Source, permissions, cleaning steps and version | Connects the analysis to the correct material |
| Code | Scripts, packages, environment and tests | Allows the analysis to be rerun |
| Interpretation | Human decisions, rejected alternatives and uncertainty | Prevents generated prose from hiding judgment calls |
| Disclosure | Policy checked and statement submitted | Supports transparency and publication compliance |
Before submitting or publishing AI-assisted research
The bottom line
Let AI widen the search. Do not let it replace the chain of evidence.
AI can make academic work faster. It can suggest terminology, expose assumptions, explain methods, compare passages, inspect code and help communicate a result.
It can also produce the appearance of scholarship without the substance: references that do not exist, summaries that erase uncertainty, code that runs the wrong analysis and prose that hides where human judgment ended and automated generation began.
The researcher remains accountable for the method, the evidence and the final claim.
Sources and further reading
- Google Trends — Artificial Intelligence Search Trends
- ICMJE — Use of Artificial Intelligence in Publishing
- ICMJE — Use of AI by Authors
- Nature Communications — Artificial Intelligence Editorial Policy
- Nature Methods — Using AI Responsibly in Scientific Publishing
- UNESCO — Guidance for Generative AI in Education and Research