8 min read
Intermediate
Draft: this article is under editorial review and may change.
A reference that looks perfect, with plausible authors, a real journal and a believable year, can still be invented. Language models write text that resembles what usually comes next, and to them a citation is just another pattern. This risk of HallucinationWhen an AI states something false with confidence, for example inventing a journal article that does not exist. shapes everything in this guide.
AI can still make you a faster, sharper researcher. It can stress-test your question, screen abstracts against fixed criteria and play a tough reviewer of your draft. But research rests on claims others can check, so this workflow puts a check between every stage and keeps every judgement with you.
The idea in 30 seconds
Use only sources you supply
AI works on records you exported, never on its memory of the literature.
- Use only sources you supply AI works on records you exported, never on its memory of the literature.
- Check before moving on A gate between every stage stops errors travelling.
- AI as sceptical reviewer Ask it to attack your design, not to praise it.
- You interpret Themes, findings and claims are yours.
- Disclose AI use Keep a simple log and follow journal and university policy.
- Every reference is guilty until checked.
Key takeaways
If you remember nothing else, remember these.
- Use AI on sources you supply. Never cite a reference from a model until you have opened it and checked its DOIA permanent ID that finds the real journal article. or database record.
- Put a verify gate between stages: nothing moves forward until you have checked it.
- Make AI your sceptical reviewer, not your co-author. Critique, screening and code are good uses; interpretation and claims are yours.
- Disclose AI use under your journal’s and university’s current policy, and keep confidential manuscripts out of AI tools.
A pipeline with a gate between every stage
Think of a research project as a production line. Work flows from the question to the final disclosure statement, and errors flow with it. A study wrongly excluded at screening doesn’t announce itself later; it quietly becomes part of your findings.
The gates are your quality checks, like an inspector between stations. Nothing moves on until it passes.
Tap a stage, or press Play to send the work through every gate.
Every prompt here uses the same five layers: AIM, Action verbs, Named methods, Calibrated dialogue and Verify. If an AI AI agentAn AI that does a multi-step job for you (planning, opening files, using tools and checking its work) instead of only replying once. will work through your files on its own, read the agent briefing guide too.
Sharpen the question first
A weak research question wastes every hour that follows. Two well-known frameworks help. FINER, associated with Hulley and colleagues, tests whether a question is worth asking. Population, Intervention, Comparison, Outcome: four boxes that make a question about the effect of something precise. structures questions about the effect of an intervention, and it works well beyond medicine.
FINER in five tiles
Feasible
Can you do it with your time, money, skills and access?
- Feasible Can you do it with your time, money, skills and access?
- Interesting Will your field care about the answer?
- Novel Does it add something not already known?
- Ethical Can you protect participants and get approval?
- Relevant Will the answer matter for practice or policy?
Here is an accounting-education example in PICO form. Population: diploma accounting students at two polytechnics. Intervention: weekly AI-generated practice on bank reconciliations. Comparison: the usual tutorial. Outcome: scores on a common end-of-semester test.
AI is useful here as a critic. Paste your draft question, objectives and planned method, and ask it to test each FINER criterion the way a grant panel would.
You are a sceptical senior reviewer on a Malaysian research grant panel.
My draft research question, three objectives and planned method are pasted below. Method: an online survey of 300 SME owners in Selangor, 18 months, one research assistant.
Evaluate the question against each FINER criterion, giving a rating (strong / adequate / weak) and one sentence of evidence for each. Then rewrite the question in PICO form if it fits, or explain why it does not.
Use the FINER criteria (Hulley and colleagues) as your checklist, and say which criterion you are least sure about.
What would you need to see in my proposal to be convinced this is feasible in 18 months?One limit to remember: the AI can’t know whether your gap is real. A claim that “no study has examined X in Malaysia” must come from your own searches of Scopus, Web of Science, MyCite or your field’s databases, not from a model’s impression.
Critique your design with a sceptical reviewer
Most of us are too close to our own designs to see their weak points. Give the AI a reviewer’s role and a reviewer’s job, and it becomes a useful sparring partner.
A named method that fits well is the Imagine your project has already failed, then list the most likely reasons. It surfaces risks that optimism hides., popularised by psychologist Gary Klein. You ask the AI to imagine the study has already been rejected, then explain why.
You are a methodologist who reviews quantitative and mixed-methods designs for an education research journal.
My study design is pasted below: research question, sampling plan (polytechnic lecturers in three states), instruments, and analysis plan.
Critique the design and rank the five most serious threats to validity, from most to least serious. For each, name the threat, explain how it could change my conclusions in one or two sentences, and propose one practical fix.
Run a pre-mortem: assume the study was rejected after review, and explain the most likely reasons.
Which of my assumptions are you least able to judge without more information?Treat the list of threats to Whether your study really measures what it claims, and whether its conclusions hold up. as a checklist to think through, not a verdict. Some threats won’t apply to your context; others will be ones you genuinely missed. Discuss the serious ones with your supervisor, and get ethics approval before you collect any data.
Search and screen, using only sources you supply
This is where the risk of invented references is highest. Asking a chatbot to “find papers on” a topic invites made-up titles, wrong authors and wrong years. The rigorous pattern splits two jobs, like a librarian who plans the search and a reader who checks each book:
- AI designs the search. It suggests key concepts, synonyms and A search line that joins words with AND, OR and NOT, for example: (SME OR 'small business') AND Malaysia. for each database. It lists no papers.
- You run the searches. Use the real databases, record the date, database and number of hits, and export the results.
Then AI can help you screen, but only the abstracts you paste in, against criteria you fixed in advance.
Find me 20 recent papers on AI adoption in Malaysian SMEs and summarise them.
What you typically get back
A tidy list of 20 references with summaries. Several cannot be found in any database, two real papers are credited to the wrong authors, and the summaries mix findings from different studies.
You are a research assistant who screens records for a systematic review reported to PRISMA 2020.
Inclusion criteria are below, followed by 25 abstracts I exported from Scopus, labelled R01–R25. Use only these abstracts. Do not add, rename or cite any other source.
Screen each abstract and return a table: ID | Include / Exclude / Unsure | criterion | one-line reason quoting a phrase from the abstract. Stop after R10 so I can check your decisions.
Record every decision so it can be counted in a PRISMA 2020 flow diagram.
Which criterion was hardest to apply, and how should I tighten it?What you typically get back
A table of decisions you can audit line by line, each linked to a quote from the abstract. You correct two decisions in the first batch, tighten one criterion, and resolve every 'Unsure' yourself.
For a Systematic reviewA literature review that follows fixed, reported search rules., a second human reviewer is still essential. Check the AI’s decisions on the first batch and agree your criteria before it continues, because a wrongly excluded study disappears silently.
PRISMA 2020An international checklist for reporting how review studies were found and chosen. is a reporting guideline, not a screening method. It tells you what to report, such as a flow diagram showing how many records you kept and dropped at each stage. Report honestly how AI was used.
Extract data into a fixed table
Pulling numbers out of papers is repetitive and error-prone, so it is tempting to hand it over. The danger is that models fill gaps with plausible values. The fix is to demand evidence: for every value, a page number and a short quote, and “Not reported” when the paper is silent.
For a review of AI adoption in Malaysian SMEs, your form might ask for the sample size, sector, state, how adoption was measured and the main result. Each value comes with its page number and quote, so you can check any cell in seconds.
That makes guesses visible and your spot-checks fast. Design the extraction form yourself, test it on two or three papers, and record who extracted what.
Investigate a surprising result
Sometimes your data says something you didn’t expect. Suppose your SME survey finds that owners who attended a government digital-skills course report lower AI adoption than owners who didn’t.
Explaining a contradictory result is a diagnostic task, so this is where the Chris Sanders method fits. Ask the AI for rival explanations, including plain chance, and for the check that would rule each one out.
You are a quantitative social scientist who reviews survey studies of Malaysian SMEs.
Survey of 300 SME owners in Selangor. Owners who attended a government digital-skills course report lower AI adoption (mean 2.1 on a 5-point scale) than owners who did not (mean 2.6). My variable list and sampling notes are pasted below. You have not seen the raw data.
Investigate why this result might occur. List at least four rival explanations, including chance. For each, give one check I can run on my own data to support or rule it out. Do not pick a winner.
Use the Chris Sanders method to complete this task. First, restate in three bullets how you will apply it here.
What would you need to know about how owners were chosen for the course?Then you run the checks. Perhaps owners were sent on the course because their businesses were already struggling. The AI can list ideas; only your data can decide between them.
Qualitative coding with your own codebook
In qualitative work the analysis is interpretive, and the interpreter is you. AI can still help in a disciplined way, once you have built a CodebookYour list of labels and meanings for tagging interview text. from your own close reading.
The AI applies your codes to anonymised excerpts, quotes the words behind each code, rates its confidence and lists excerpts that fit no code. That gives you a second reader and a quick way to spot overlapping definitions.
You are a qualitative research assistant experienced in codebook thematic analysis.
My codebook (8 codes with definitions and examples) and 12 anonymised interview excerpts from polytechnic accounting lecturers, labelled P01–P12, are pasted below.
Apply the codebook to each excerpt. Return a table: excerpt ID | code(s) | the exact words that justify each code | confidence (high / medium / low). List excerpts that fit no code separately. Do not create new codes or themes.
Apply my codebook consistently, as in codebook (template) analysis. I am the analyst; you are a second reader who flags misfits.
Which code definitions overlap enough to confuse a second coder?Name the approach correctly in your methods section: this is codebook (template) analysis. Braun and Clarke’s reflexive thematic analysis is different, because it treats coding as the researcher’s own interpretive work. Examiners notice the difference.
Be careful with consent. Interview transcripts often identify people even after names are removed. Upload only what your ethics approval and consent forms allow, to tools your institution has approved.
Analysis code you run and check
For quantitative work, AI can write analysis code for you. The rule is simple: you run the code and read the results yourself, not the model.
For developers: technical detailsOptional: skip this if you only use the chat app
Ask for R or Python code that explains each step and the assumptions behind each test. Then:
- run it on a small sample and compare the result with a manual calculation or a known example;
- check that missing values, reverse-coded items and outliers are handled the way your analysis plan says;
- read the output yourself, rather than asking the AI to interpret numbers it has not seen.
If you share data with an AI tool, share the structure (variable names and types, a few fake rows) rather than real participant data, unless your approvals allow otherwise.
Explore the whole workflow
Tap through the seven stages below. Each one shows what AI can do, what you must do, a prompt to adapt and the main risk at that stage.
Tap a step (or use the arrow keys). Each one shows what AI can do, what stays with you, a prompt you can copy and a risk to watch.
Step 1 of 7
Sharpen the question and the design
What AI can do
- Test your question against FINER criteria.
- Rewrite it in PICO form where that fits.
- Play a sceptical reviewer of your methodology.
What you must do
- Own the question and its contribution.
- Check the gap against literature you have actually read.
- Plan ethics approval before collecting data.
You are a sceptical senior reviewer for a Malaysian research grant panel.
My draft research question, objectives and planned method (a survey of 300 SME owners in Selangor) are pasted below.
Critique the question against each FINER criterion (feasible, interesting, novel, ethical, relevant) and list the five most serious weaknesses in the design, ranked, each with one fix.
Run a pre-mortem: assume the study failed and explain why.
What would you need to see in my proposal to be convinced this is feasible in 18 months?
Risk to watch
The AI cannot know whether your gap is real. "No studies exist on X" claims must come from your own database searches, not from the model.
Writing, editing and disclosure
AI can help you restructure a messy draft, tighten wordy paragraphs, check that every claim has a citation from your own reference list, and point out where you overclaim. Used this way, it is a demanding editor. The argument and the voice stay yours.
Many publishers and research-integrity bodies, including COPE, say AI tools can’t be listed as authors, because they can’t take responsibility for the work. Many journals also ask how you used generative AI. Rules differ between journals, and universities have their own rules for theses.
The practical habit is a simple AI-use log: tool, date and task, kept as you work. At submission, read the current policy of the journal and your university, and write your statement from the log. Your supervisor and graduate school can tell you what applies to a thesis.
What AI can help with, and what stays yours
Across the whole pipeline, the split is the same. AI speeds up the parts that can be checked; you keep the parts that make the work scholarship.
| AI can help | Only you decide |
|---|---|
| Search terms and Boolean strings | Which databases to search, and the final search |
| First-pass screening of abstracts you supply | Every inclusion decision and every “Unsure” |
| Rival explanations for a surprising result | Which explanation your data supports |
| A second reading with your codebook | Your themes and what they mean |
| Editing your wording | Every claim, citation and your name as author |
Peer review: keep manuscripts confidential
When you review for a journal or a grant body, the manuscript is confidential. Uploading it to an AI tool may share unpublished work with a third party, and many journals and funders forbid it. Even where it is allowed, the judgement in a review is supposed to be yours.
If you want help, use AI only on your own words: for example, to tighten comments you have already written, without pasting any manuscript text. Check the reviewer guidelines first, and when in doubt, ask the editor.
Check your understanding
Nice work!
Lesson complete
You made it to the end. Take the quiz above if you have not yet, for bonus XP.
- 0 XP
0 day streak