AI as a Research Assistant
Generative AI can accelerate almost every step of the research process. It can help formulate questions, explore evidence, analyse data, draft a manuscript and simulate peer review. The scientific responsibility does not move with the task.
How to use AI without outsourcing the science
See the process
Ask AI to explain what it is doing and why. A polished output is not evidence that the underlying reasoning is sound.
Challenge before accepting
Ask what could make the conclusion wrong, which variables were ignored, what assumptions were made and which words go beyond the data.
Verify outside the model
Check references, journal requirements, factual claims and important evidence against the original source or a trusted scientific database.
AI can support every step, but every step still needs a researcher check
What AI can do and what you still need to check
The examples below follow the same logic used in the TIP26 exercises. They are not a fixed recipe. The point is to keep the reasoning visible.
Research question
Start with the clinical problem, not with the model.
Generate candidate questions, refine scope, define population, exposure, comparison and outcomes, and suggest alternative formulations.
Is the question clinically important, answerable with the available data and sufficiently precise? Use Consensus.app or another evidence source to see what is already known before committing to the question.
Existing evidence
Use AI to navigate evidence, not to replace source verification.
Summarise themes, compare studies, identify possible evidence gaps and help formulate more precise literature searches.
Does the cited paper exist? Does it actually support the claim? Is the evidence direct, current and relevant to the population and outcome you are studying? Verify important claims in the original literature.
Understand the data
Start simple and look before you model.
Describe variables, summarise groups, detect missingness, create visualisations and surface unexpected patterns.
What did AI choose to show? What did it not show? Equal group size does not mean comparable groups. Check baseline differences, follow up, missingness and how each variable was defined.
Analysis strategy
A model is a scientific choice, not just a technical operation.
Suggest statistical approaches, generate code, compare models, perform adjusted analyses and create figures and tables.
Why are these variables included? What was omitted? Could there be confounding by indication? Does a composite variable duplicate its components? Does the model answer the actual research question? Do not include an analysis in a manuscript if nobody on the research team can explain and defend it. More analysis is not automatically better analysis.
Results
Separate what was observed from what it might mean.
Turn an analysis into tables, figures and concise scientific Results text.
Are all numbers actually generated by the analysis? Has AI introduced a new analysis or interpretation? Is the sentence descriptive, or has it quietly become causal?
Clinical interpretation
This is where language can become stronger than the evidence.
Translate statistical findings into clinical language, identify possible implications and compare alternative interpretations.
Association is not causation. A non significant difference is not evidence of equivalence. Clinically meaningful requires a justified clinical threshold. Ask which exact words you would challenge as a reviewer.
Missing data and selection
Who is absent can change what the observed result represents.
Compare completers and non completers, identify patterns in missingness and suggest sensitivity analyses.
Who is excluded from the analysis? Is missingness plausibly random? Can you know the direction of bias from the observed data? Do not run an advanced correction method unless you understand the assumptions it introduces.
Outcome choice
There may be no single best outcome.
Compare multiple domains, identify trade offs and explore how conclusions change when different outcomes are prioritised.
Who decides what counts as the best outcome? Urinary, bowel, sexual and general health can point in different directions. Patient preferences may be part of the scientific question rather than a secondary detail.
Manuscript drafting
Draft section by section so you can see what changes.
Draft title, Methods, Results, Discussion and Introduction from the research process, then assemble the verified sections into one submission manuscript.
Has AI invented recruitment procedures, ethics approval, software, prespecified analyses, eligibility criteria or methods that were never provided? Results should contain only analyses that were actually performed. A polished Methods section is not the same as a reproducible Methods section. When the manuscript is ready for independent review, freeze that version before opening a fresh reviewer conversation.
References
A plausible citation is not a verified citation.
Suggest relevant literature and draft an Introduction with references.
Verify every important reference against the original source. A real reference can still be the wrong reference for the claim. Check the population, treatment, outcome, timing and exact statement being supported. AI can also fabricate publications or combine details from different sources.
Peer review and journal fit
AI can simulate review, but the reviewer also needs reviewing.
Simulate methodological review, editorial screening and scientific peer review. Use separate fresh conversations when you want the reviews to be independent of the authoring process and of each other.
Do you agree with the reviewer? What did it miss? Did it invent a journal requirement or overstate a methodological problem? AI outputs are shaped by conversation history, so separate creation from independent review. Check journal requirements on the official journal site. Detailed review prompts are provided below.
Final accountability
Automation does not transfer scientific responsibility.
Accelerate almost every operational step and create increasingly polished research outputs.
Can you explain the research question, the data, the model, the assumptions, the uncertainty, the evidence and the final claim without relying on the AI answer? If not, the work is not ready to carry your name.
Use longer prompts when you have time to learn from the review
The timed TIP26 challenge uses shorter review prompts. For real work, a longer review can be more useful because it asks the AI to explain the problem, show where it appears and translate technical concerns into consequences for the science. Run these in separate fresh conversations using the same frozen manuscript.
Methodological reproducibility review
Can another researcher understand and reproduce what you actually did?
Do not accept technical criticism just because it sounds sophisticated. Ask the reviewer to explain unfamiliar terms, and return to the analysis record or code before changing the manuscript.
Editorial screening
Would an editor spend journal and reviewer resources on this paper?
An editor and a methodological reviewer answer different questions. A technically sound analysis can still be a weak paper if the contribution is unclear. Verify current journal scope and submission requirements independently.
Scientific peer review
How much confidence should a reader place in the interpretation and clinical conclusion?
Peer review is another AI output, not a verdict. Decide which comments you agree with, which require checking, what the reviewer missed and whether its proposed solution would actually answer the scientific problem.
Stop and check when AI uses words like these
The words are not automatically wrong. They signal that the claim may require stronger evidence than a descriptive association.
Ask: What exactly in the data or evidence allows us to make that claim?
Do not separate AI use from research governance
Before uploading research data, confirm what your institution, study protocol, ethics approval, data processing agreements and chosen AI service allow. Do not upload identifiable or confidential participant data to a general AI service unless this is explicitly permitted and appropriately protected. Synthetic teaching data are not a model for handling real patient data.