The honest picture
AI tools are now good enough to be useful in research and unreliable enough to be dangerous. Both are true at once, and most of the public conversation picks one. Our view is simple: use them for tasks where errors are cheap to catch, measure how often they are wrong on your own material, and keep a named person responsible for everything that goes out.
Where AI helps
- Drafting and summarizing support, with a person checking every claim against the source.
- Screening assistance in reviews, to prioritize records for human decision.
- Structured data extraction, when accuracy is checked on a sample.
- Code assistance for analysis and data cleaning, reviewed and tested like any code.
- Brainstorming search terms, synonyms and angles you may have missed.
Where it fails
- Invented citations, quotes and figures that look entirely plausible.
- Hidden bias inherited from training data, hard to see in any single output.
- Results that cannot be reproduced when the model or its settings change.
- Privacy leaks when participant or unpublished data is pasted into a tool.
- Confident wrong answers, delivered in the same tone as right ones.
Where AI fits across the research lifecycle
The same tool can be harmless at one stage and risky at the next. Choose a stage to see where AI tends to help and what to watch for.
Question
Good uses
- Brainstorming related questions and angles
- Listing terms and synonyms for an initial search
- Testing whether a question is stated clearly
Watch outs
- Summaries of a field that cite papers you have not checked
- Questions that drift toward what the tool finds easy to answer
- False confidence that a gap in the literature exists
Design
Good uses
- Checklists of design threats to consider
- First drafts of protocol sections for expert revision
- Rewording survey items for readability
Watch outs
- Sample size figures produced without a real calculation
- Generic designs that ignore your setting and constraints
- Survey wording that shifts meaning while sounding smoother
Data
Good uses
- Structured extraction from documents, validated on a sample
- Drafting cleaning code for review
- Spotting inconsistent coding across sources
Watch outs
- Uploading identifiable or confidential data to external tools
- Extraction errors that look like real values
- Undocumented changes to the data that nobody can trace
Analysis
Good uses
- Code assistance, with tests and human review
- Explaining unfamiliar functions or error messages
- Drafting plain-language descriptions of results you have checked
Watch outs
- Methods chosen because the tool suggested them
- Running many models until one looks good
- Numbers in text that do not match the actual output
Writing
Good uses
- Tightening sentences you wrote
- Checking consistency of terms across a long document
- Drafting lay summaries for careful revision
Watch outs
- Citations you have not read, or that do not exist
- Claims that quietly overstate the evidence
- Undisclosed use where a journal or funder requires disclosure
Review
Good uses
- Checklists for your own draft before submission
- Organizing reviewer comments into themes
- Comparing versions to confirm requested changes were made
Watch outs
- Pasting manuscripts you are reviewing into external tools
- Treating automated critique as a substitute for expert reading
- Policies on AI in peer review that differ by journal and funder
Principles for responsible use
- Hallucination
- Output that is fluent and confident but false, such as a citation to a paper that does not exist.
- Precision and recall
- Of the items a tool flagged, the share that were correct (precision); of the items it should have flagged, the share it found (recall).
- Inter-rater agreement
- How often two independent reviewers reach the same decision, beyond what chance alone would produce. A common reference point for judging a tool.
- Human in the loop
- A workflow where a person reviews and can overrule each consequential output before it is used.
A named person is accountable
Every claim, number and citation has a human owner who has checked it. "The tool said so" is never the source.
Validate against a human-checked sample
Before relying on a tool, compare its output with careful human decisions on a sample of your own material, and measure the errors.
Document everything
Record the tool, the model version, the settings and the prompts, so the work can be described honestly and repeated.
Protect data
Participant data and unpublished work stay out of tools whose terms allow retention or training, unless approvals and agreements explicitly permit it.
Follow and record disclosure rules
Funders and journals set their own rules on AI use and they change. Check the current policy for each submission and keep a record of what you disclosed.
Never cite what you have not read
If a reference came from a tool, find it, read it, and confirm it says what you claim, or remove it.
What we offer
Validating a tool on a convenient example and then trusting it everywhere. Accuracy varies by topic, document type and wording. Validate on a sample from the material you will actually use.
- AI-assisted evidence synthesis
- Review workflows that use AI for screening or extraction where it helps, with validation measured against human decisions and reported in the methods.
- AI-use policies and disclosures
- A short, practical policy for a lab or team: what is allowed, with which data, how it is documented, and template disclosure statements to adapt for each submission.
- Pilot evaluations
- A test of one AI tool or model on one specific research task, with accuracy measured against a human-checked reference sample and a clear recommendation at the end.
- Privacy and de-identification review
- A review of what data would go to an AI tool, what the tool's terms allow, and what must be removed or changed before anything is shared.
- Workshops for research groups
- Practical sessions built around your group's own tasks and data types, covering good uses, failure modes and how to document AI use.
How a pilot works
A pilot answers one question: is this tool accurate enough, for this task, on this material, to be worth using? It ends with a decision and a record of how it was reached.
Define the task
One task, one tool, a clear definition of a correct output, and the error rate you can tolerate.
Build a reference sample
A sample of your own material, judged independently by people, with disagreements resolved and agreement measured.
Test
Run the tool on the same sample with fixed settings and recorded prompts.
Measure errors
Precision, recall and the kinds of mistakes made, compared with how well people agreed with each other.
Decide and document
Adopt, adapt or drop the tool, with a short report you can cite in methods or policy.
Illustrative numbers for an AI screening tool checked against human decisions on 1,000 records. Accuracy is 91.8% and sensitivity is 91%, yet 18 relevant records were lost. Whether that is acceptable is a decision for the review team, and the check is what makes it visible.
Questions
It can help with parts of the drafting, but the ideas, the judgment and the responsibility must be yours, and reviewers tend to notice generic text. Some funders have specific rules about AI use in applications and in review. Check the current policy of your funder before you start, and keep a record of how you used any tool.
Policies differ between journals and change often. Some permit certain kinds of AI assistance when it is disclosed; others restrict it more tightly. Check the current policy of your target journal, and we can help you write an accurate disclosure statement.
Often it is not, at least not without checks. It depends on the data, your approvals and agreements, and the tool's terms on retention and training. Participant and unpublished data deserve particular care. A privacy review before anything is uploaded is the safe default.
You check it, and you measure how often it is wrong. For anything consequential, compare the tool against careful human judgments on a sample of your own material before relying on it, and keep a person reviewing outputs afterward.
There is no single answer. The right tool depends on the task, your data and your institution's rules, and the honest answer is often found by running a small pilot.
Only where we have agreed it with you, for tasks where it is appropriate, and never with confidential material in tools that might retain it. Any use is documented so you can disclose it accurately.