The Use of Artificial
Intelligence in Research
CAIT, FTSM UKM
3.15 – 4.30 pm
FEP UKM
Between 3.15 and 4.30 this afternoon…
From an audit of 2.5 million papers: roughly 146,900 hallucinated citations in 2025 publications alone.
Nearly all of them passed peer review. Nobody set out to deceive. They simply did not check.
It is not steady. It is accelerating.
Rate of papers containing at least one fabricated reference. Sixfold in two years.
Study in The Lancet (2026), Columbia University researchers — over 2 million papers and 97 million citations. The 2024 bar is interpolated for display; 2026 covers the first seven weeks.
AI is where this morning's four slots collide
Beyond compliance
No form to sign here. Only culture is left.
UKM integrity
Fabrication and falsification can now happen without intent.
Social science ethics
Did your participants consent to third-party processing?
Data governance
Every prompt is a data transfer.
This slot is not a fifth topic. It is a stress test on the first four.
Six parts, one hour
Reality — what this tool actually is
Two live demonstrations
Use — across the seven stages of research
Including a prompting demo and how to choose tools
Five ethical boundary lines
Accountability · transparency · confidentiality · validity · justice
A practical framework
What you can do tomorrow morning
Four cases — you vote
Red, amber or green
Closing
What becomes expensive when everything else becomes cheap
Q&A in the final 15 minutes. I would rather have sharp questions than polite ones.
Seven stages. AI touches all of them.
Idea
Questions and gaps.
Review
Summary and synthesis.
Design
Instruments, translation.
Collection
Participant transcripts.
Analysis
Coding, interpretation.
Writing
Drafting and editing.
Dissemination
Abstracts, peer review.
The public conversation
Focuses on stage 6 — writing. The easiest to detect, and the least dangerous.
The real risk
Stages 3, 4 and 5 — participant data and methodological judgement. Damage here cannot be undone.
It is not a fact engine.
It is a language probability engine.
One task only
Predicting the word most likely to come next.
No ground truth
It does not “know” a source exists. It produces the convincing shape of one.
Confidence is a style
The assured tone comes from academic prose patterns — not from certainty.
So what is actually happening when it answers? Let's watch.
Inside a single sentence
Illustrative simulation: the mechanism is accurate, the numbers are invented for teaching.
One word per click
The same mechanism, applied to references
Every part is right
Perfect format. Plausible page range. The journal name sounds exactly right.
The whole does not exist
No valid DOI, no database record. It was not copied — it was generated.
Which is why it passes
Reviewers check whether a reference looks right. The shape is right.
Hallucination is not rare. It is a base rate.
Rates are higher for specialised and under-represented topics. Researchers working on local context carry the highest risk.
JMIR Mental Health (2025); cross-model benchmark (2026); Walters & Wilder (2023) reported 55% for GPT-3.5 and 18% for GPT-4.
The model speaks in the voice of its training data
Constructs migrate
“Wellbeing”, “entrepreneurship” — answered using definitions dominant in an English-language corpus. Your construct shifts meaning without you noticing.
Patterns reflect history
Is the pattern it identifies a real phenomenon, or an artefact of the algorithm?
Deceptive uniformity
AI summaries flatten disagreement. A live debate becomes a false consensus.
In social science, a tool that flattens nuance produces wrong findings in a way that looks entirely reasonable.
When the output becomes the input
Each output became the next input.
No new information ever enters
Each pass keeps what is probable and drops what is rare. Fine detail dies first; only bold, high-frequency shapes survive.
The accurate name for this
Model collapse, not drift. Drift is the world moving around a fixed model; collapse is a model trained on its own output, losing the tails.
Why this reaches your discipline
About 146,900 fabricated citations entered the record in 2025 — and that record is training data. Local scholarship sits in the tails, and tails go first.
Careful, original, human-collected data is what stops this. Not a sentimental claim — it is the finding.
Clip compressed and silenced for this deck; the printed PDF substitutes four sampled frames. Model collapse: Shumailov, I., et al. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759.
Once it leaves, you no longer know where it goes
There is no undo button. Every stop is outside your control — and the route is different every time.
Never enter these into public AI tools
Human data
Interview transcripts, recordings, case notes, contact details — anything that could identify a participant.
Student material
Student records and assessment material not already public.
Others' confidential work
Unpublished manuscripts, embargoed findings, grant proposals, peer reviews, examiner reports.
Third-party property
Instruments, datasets, code or material under a confidentiality obligation.
Default position when in doubt: do not upload.
Seven stages, two columns
| Stage | Useful for | Not for |
|---|---|---|
| 1 · Idea | Generating alternative questions; stress-testing your argument | Deciding what the literature gap is, unchecked |
| 2 · Review | Summarising papers you have already read | Finding references without verification |
| 3 · Design | Checking items for clarity and ambiguity | Translating instruments without validation |
| 4 · Collection | Drafting protocols and interview guides | Processing participant transcripts in public tools |
| 5 · Analysis | Generating code; a second lens on your themes | Replacing a second coder |
| 6 · Writing | Language fluency and argument structure | Generating findings or discussion |
| 7 · Dissemination | Drafting lay summaries; alternative titles | Reviewing other people's manuscripts |
The right column is not merely a moral prohibition — every item is a task the model cannot perform reliably.
Two ways of asking. Two kinds of answer.
The first way
Author, B. (2019). A plausible title…
Author, C. (2022). A plausible title…
The second way
Article 2 — construct: financial behaviour; instrument: not stated
Contradiction: two different definitions of the same construct
The difference is not prompt length. It is where the information came from — from you, or from the model's memory.
Illustrative example; actual output varies by tool and version.
Five parts — and the template they build
Context
Who you are, what study, for which reader.
A fixed source
The material you supply — and a ban on anything else.
One task verb
List, compare, classify. Not “please help me”.
Output format
Fields you specify — so the gaps become visible.
An uncertainty rule
The single sentence that saves the most trouble.
I am a [role] at [faculty], writing [type of output] on [research topic].
Use ONLY the material I have attached below. Do not add sources, references or figures from anywhere else.
List the construct and instrument used by each article, and state any contradictions between them.
Return a table with the fields: Article | Construct | Instrument | Sample size | Main finding.
If a field is not in the text I supplied, write “not stated”. Do not estimate and do not fill gaps.
Four of these five parts are your work — knowledge of your own study. Which is why a good prompt cannot simply be copied from someone else.
Six patterns worth keeping
Extract to a table
“From the attached, extract [fields] into a table. Write not stated where absent.”
Why: gaps become visible.
Steelman the argument
“Give the three strongest objections a critical reviewer would raise.”
Why: nothing leaves; rehearses Reviewer 2.
Find the disagreement
“Where do these sources conflict? Quote the conflicting lines.”
Why: undoes the flattening of debate.
Explain, then verify
“Explain [concept], then list what I should verify independently.”
Why: output becomes a checklist.
Translate & back-translate
“Translate these items, then translate your translation back.”
Why: surfaces drift early.
Structure, don't write
“Propose three alternative outlines. Outlines only — no prose.”
Why: the sentences stay yours.
What every pattern has in common: you supply the material, and the model does something structural to it.
Choose by engine type, not by brand
No search
Good for drafting, structure, explanation and code. Dangerous for anything needing a source.
e.g. ChatGPT · Claude · Gemini · Copilot
Fetches first, answers second
Good for finding literature and querying your own documents. A citation may exist yet not support the claim.
e.g. Elicit · Consensus · Scite · Semantic Scholar · NotebookLM
Built for one task
Screening, extraction, references, qualitative coding. Narrow scope, far easier to audit and report.
e.g. Rayyan · Covidence · ASReview · Zotero · NVivo / MAXQDA / ATLAS.ti
Searches, runs code, chains steps
Most capable, hardest to audit. Condition: log every step, or you cannot write your Methods section.
e.g. deep-research modes in major assistants; agents that browse and execute code
Runs on your own machine
Data never leaves. Lower performance, but often the only legitimate option for participant data.
e.g. Ollama · LM Studio · local-first QDA tools
The first question is not “which is smartest”, but: does it search? can it be audited? does my data leave?
Examples current as of August 2026 — illustrations of each type, not endorsements. If a tool shows no links to real documents, it is a generative engine.
Literature screening that survives review
Search in a real database first
Scopus, Web of Science, EconLit. AI does not replace the search — it reorders what the search returned.
Deduplicate and import to a screening tool
Rayyan, Covidence or ASReview — decisions stay in one auditable, exportable place.
Let AI rank relevance; you screen
Ranking changes the order you read in. It does not change what is included.
Hand-screen a random 10% and compute agreement
This single step converts “I used AI” into a reportable method.
Report tool, version, settings and agreement rate
In Methods, alongside your PRISMA flow.
AI can reorder the pile. It cannot decide what is in it. Keep the inclusion decision human and the disagreement rate visible.
Qualitative coding: a second lens, not a second coder
You code a subset by hand first
Your codebook exists before any tool sees the data — otherwise the model's categories quietly become yours.
Anonymise — or keep it local
Strip identifiers, or run a local model so transcripts never leave the machine.
Ask for candidate themes on that subset
Require a supporting quote for every theme, so each one is traceable.
Compare with your codebook — and keep the conflicts
Where you disagree, write an analytic memo. That memo is the finding.
Report which step AI touched and how you validated it
UKM's baseline standard: validation must be reported, not merely performed.
The value is in the disagreement, not the agreement. If the model only confirms what you already coded, it has told you nothing.
Analysis code: the fastest win, with one rule
Describe the data, not the data itself
Variable names, types and structure are enough. Raw rows are rarely needed.
Ask for commented code, run it on a subset
Stata, R or Python. The comments are how you learn what it assumed.
Check one result by hand
One mean, one cross-tab, one coefficient. If it matches, continue.
Ask for the same analysis a second way, then compare
Divergence between two routes is the cheapest bug detector you have.
Why this is the safest use of all
No participant data need be uploaded, the output is testable against a known answer, and errors announce themselves as errors rather than as fluent prose.
The one rule
Code that runs is not code that is correct. A silent logic error still produces a table — and the table still looks publishable.
You are responsible for every line, including the ones you did not type.
Where AI presses on research ethics
Accountability
Who answers when the output is wrong?
Transparency
What must be declared, and where?
Confidentiality
Is this data yours to give away?
Validity
Can it be repeated?
Justice
Whose context becomes the default answer?
Each one already has an official document behind it. You do not need to invent policy.
Accountability cannot be transferred to a tool
A consistent position
- COPE: AI tools cannot be listed as authors.
- MOHE (2024): generative AI technology may not be listed as a co-author of any academic output.
- UKM (2025): using a tool does not transfer responsibility to the tool's provider.
A simple test
If you cannot defend a sentence in a viva or an integrity inquiry, that sentence does not yet belong in your manuscript — no matter who typed it.
Not every use needs to be declared
- Grammar, spelling, thesaurus
- Smoothing sentences you wrote yourself
- Reference formatting
- Substantive draft text
- Literature summaries entering the manuscript
- Analysis code and data coding
- Suggested themes, interpretations, arguments
- Translation of research instruments
The line is not how much AI was used. The line is whether AI contributed intellectual content.
Declared in the Methods and Acknowledgements sections. For UKM theses: Appendix B2 of the UKM Generative AI Guideline.
Participant data is not yours to give away
Coding is not anonymisation
In a small population, context alone can identify a participant.
Transcription counts
A 2025 research-ethics editorial recommends informing participants and obtaining explicit consent.
Peer review
NIH prohibits reviewers from uploading application content to online AI tools.
“If my participants reread the form they signed, would they expect their words to be sent to a foreign company's servers?”
If it cannot be repeated, it is not a method
Qualitative coding
It can be a second lens. It cannot be a second coder.
Review screening
It can speed things up — if accuracy rates are reported and human cross-checked.
Analysis code
Subtle bugs still produce output. You test every line.
What must be recorded
Tool · version · date · prompt · data supplied · output accepted · how it was verified.
The answer is correct — for another country
The model's answer
illustrativeWhat never appears
Malaysian contextThe gap between these two columns is your expertise. The model gives no warning that it is answering about somewhere else — and your reader will not notice either.
Illustrative of a typical response pattern, not a transcript of any particular model. Weaker performance on under-represented languages and contexts is documented in the studies cited on slide 11.
Five lines, principles you already know
| Boundary line | Principle | Official reference |
|---|---|---|
| 1 · Accountability | Accountability | COPE · MOHE (2024) · UKM (2025) |
| 2 · Transparency | Honesty | Publisher policies · UKM Appendix B2 |
| 3 · Confidentiality | Respect for participants | Ethics approval · UKM data baseline |
| 4 · Validity | Reliability | ERA Living Guidelines |
| 5 · Justice | Fairness and equity | National AIGE (2024) · UNESCO |
There is not one new principle here. AI does not create new ethics — it makes old ethics much easier to breach without noticing.
You are not starting from nothing
UKM Generative AI Guideline
Data baseline standards, the ETHICAL framework, and the official disclosure form (Appendix B2).
Generative AI Guidelines for Higher Education
AI cannot be a co-author; all content must be properly acknowledged.
National AI Governance & Ethics (AIGE)
Seven principles of responsible AI, including fairness, privacy and accountability.
ERA Living Guidelines & publisher policies
European Commission (updated 2026), COPE, Elsevier, Springer, Sage, Wiley.
And closest of all: your own ethics approval. The scope on that form is the real contract with your participants.
Traffic lights, mapped to real FEP work
Red
- Identifiable transcripts
- Manuscripts or theses under review
- Uncollected data reported as real
- Citing a source you have not opened
Amber
- Literature summaries — every source checked
- Theme suggestions on anonymised data
- Review screening — cross-validated
- Instrument translation — with back-translation
- Condition: declared and recorded.
Green
- Grammar and spelling
- Reorganising your own notes
- Explaining concepts you then verify
- Playing devil's advocate against your argument
An honest note: the UKM guideline itself warns that traffic-light models create an “enforcement illusion” — this is a communication tool, not a control mechanism.
ETHICAL — seven steps UKM has already adopted
Examine
Check university, funder and target journal policies.
Think
Consider social impact, not just efficiency.
Harness
Understand the technology enough to know when it drifts.
Indicate
State your use openly and specifically.
Critically engage
Do not take output at face value.
Access secure versions
Institutional accounts, not free public ones.
Look at agreements
Read the terms: what happens to your data.
Why it helps
Seven steps, not seven abstract values.
Eacersall et al. (2024). Adopted in the UKM Generative AI Guideline (2025), section 2.2.
If you remember only one thing this afternoon
Six practices that reduce risk immediately
Give the minimum
Summarise at a high level. Do not paste raw transcripts.
Anonymise first
Strip identifiers before any tool touches the data.
Institutional accounts
Their data terms differ entirely from free accounts.
Turn off training
Check retention settings. It is not the default, and it changes.
Local models
For sensitive data, nothing ever leaves. Adequate for many text tasks.
Decide early
Agree tools and data types with your supervisor at the start, and record it.
The actual words you can copy
Three rules for good disclosure
Specific — name the actual task · Assessable — readers can judge its effect · State the verification — what you did to check it.
What it looks like in a real journal
Declaration of Generative AI and AI-assisted technologies in the writing process
“During the preparation of this work the author(s) used…” — followed by the tool name, the purpose, an acknowledgement that the authors reviewed and edited the result, and a statement of full responsibility for the content.
- Improving language and readability
- Checking spelling and grammar
- Translating parts of the manuscript
- Labelling detected communities in a dataset — analysis, not writing
All of these articles were published. Disclosure does not block publication — it protects you when someone asks later.
Data: Xu, Z. (2025). Patterns and purposes: A cross-journal analysis of AI tool usage in academic writing. arXiv:2502.00632 — declarations in Elsevier open-access journals, 2024.
If you supervise, examine or review
Supervisors
- Discuss it at the first meeting, not the final draft
- Agree permitted tools and prohibited data; record it
- Request Appendix B2 when use is material
Examiners
- Never upload a manuscript or thesis
- Test understanding through oral questioning
- An invalid DOI is an early signal
Ethics reviewers
- Ask on the form: AI used at which stage?
- Check the consent form covers third parties
- Check the data storage and transfer plan
“No adverse academic decision shall be made solely on algorithmic inference.”
UKM Generative AI Guideline (2025) — AI detector scores are probabilistic signals that can produce false positives.
Four things you can do tomorrow morning
Read the UKM Generative AI Guideline (2025)
Section 2.2 and Appendix B2 in particular. Thirty minutes.
Add one AI clause to your consent template
Done once, applies to every future project.
Have a five-minute conversation with each researcher you supervise
Permitted tools, prohibited data, how it will be recorded.
Start a use log for your current project
A single text file is enough.
All four can be done in under two hours, and none requires anyone's approval.
Interview transcripts
Coding is not anonymisation. Among 28 entrepreneurs in one district, business type and contextual detail identify individuals easily.
Approval a year ago does not cover this. The consent form almost certainly says nothing about overseas third-party processing.
What should have happened: thorough anonymisation and a local model; or a protocol amendment to the ethics committee.
Instrument translation
The translating: amber. No participant data was exposed, and assisted translation is a legitimate starting point if declared.
Distributing it unvalidated: red. This is a validity failure. Semantic equivalence is not psychometric equivalence — and 400 responses cannot be rescued afterwards.
What should have happened: back-translation, expert panel, cognitive pilot, and reported reliability for the Malay version.
Peer review
The breach happened at the moment of upload, not in the writing. An unpublished manuscript is confidential work entrusted to a reviewer.
Explicitly prohibited. NIH bars reviewers from uploading application content; UKM lists unpublished manuscripts among prohibited material.
Writing it himself restores nothing. If time does not permit a proper review, the right answer is to decline the invitation.
Thesis language editing
The use is legitimate. Language improvement on human-authored text is assistive use; MOHE places no restriction on proofreading tools.
First condition: supervision. UKM requires supervisor permission for material use in a thesis, and a supervisor may restrict it early on to protect the development of writing skill.
Second condition: it stays green only while it is genuinely language. Once the tool starts suggesting arguments, it has crossed into amber.
What remains valuable is what cannot be generated: choosing the right question, recognising suspicious data, understanding context that exists in no corpus, and being willing to say “I am not yet sure”.
This tool's weakness is your strength
Models are weakest on under-represented languages, unfamiliar contexts, and topics specific to one place.
That is a precise description of this faculty's work — the Malaysian economy, local household behaviour, Islamic finance, domestic policy.
The gap in the world's corpus is your field of expertise. Not something to be protected from AI.
Being built now
A Malay foundation language model and campus inference infrastructure at UKM — text analysis without data leaving the campus.
Possible today
Open-source models on your own machine are already adequate for many text tasks.
The most valuable contribution
Quality datasets and instruments validated in Malay.
Three pledges, if you are willing
I will not cite what I have not read
I will not give away data that is not mine
I will write down what I did
None of these requires new policy, a meeting, or anyone's approval. All three begin tomorrow morning, with you.
Questions
Slides, disclosure templates and links to every official document will be shared through the workshop secretariat.
ILMU · IMAN · AMAL
Checklist before using AI
Before you start
- UKM, funder and target journal policies checked
- Permitted tools agreed with your supervisor
- Institutional account, or a local model for sensitive data
- Retention and training settings checked
While using it
- Only the minimum information supplied
- Data anonymised before any tool touches it
- Tool, version, date and prompt recorded
- Every source opened and read before citing
Before you submit
- Every reference verified against the original source
- Analysis output verified and the step reported
- Disclosure included in Methods and Acknowledgements
- Appendix B2 completed if this is thesis work
The three-question test
- Is this data mine to give away?
- Can I defend this output without the AI?
- Am I willing to write it in the Methods section?
Three yeses — proceed and declare. One no — stop.
Official documents and references
Universiti Kebangsaan Malaysia (2025). Guideline on Responsible Use of Generative Artificial Intelligence in Teaching, Learning, and Research Supervision. Centre for Teaching and Curriculum Development, UKM.
Kementerian Pendidikan Tinggi (2024). Garis Panduan Penggunaan Teknologi Kecerdasan Buatan Generatif (KBG) dalam PdP Pendidikan Tinggi. JPT, MOHE.
MOSTI (2024). National Guidelines on AI Governance & Ethics (AIGE).
European Commission (2026). Living Guidelines on the Responsible Use of Generative AI in Research (updated version). ERA Forum, DG RTD.
COPE (2023, updated 2024). Authorship and AI tools: COPE Position Statement.
NIH (2023). The Use of Generative AI Technologies is Prohibited for the NIH Peer Review Process (NOT-OD-23-149).
Eacersall, D., et al. (2024). Navigating ethical challenges in generative AI-enhanced research: The ETHICAL framework. arXiv:2501.09021.
Xu, Z. (2025). Patterns and purposes: A cross-journal analysis of AI tool usage in academic writing. arXiv:2502.00632.
Linardon, J., et al. (2025). Influence of topic familiarity and prompt specificity on citation fabrication using LLMs. JMIR Mental Health, 12, e80371.
Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13.
Samuel, G., & Wassenaar, D. (2025). Joint editorial: Informed consent and AI transcription of qualitative data. JERHRE.
UKM Research Ethics Secretariat. Guidelines and forms. ukm.my/jepukm
Disclosure of AI use: Claude (Anthropic) was used in August 2026 to search and screen policy documents, structure the slide sequence, and generate the display code. All policy sources were checked against the original documents by the speaker, who takes full responsibility for the accuracy and final content.