01 / 56

NNND6673 Research Communication Skill · Master of Clinical Nutrition

Scientific
Communication
using AI

Let AI carry the message. Never let it change the meaning.

Ts. Dr. Mohd Nor Akmal Khalid

Center for Artificial Intelligence Technology (CAIT)
Faculty of Information Science and Technology, UKM

WED 21 OCT 2026 · 09:00–11:00
BILIK SEMINAR 2, BLOK D · FSK · UKM KUALA LUMPUR

Photo: “Kuala Lumpur Skyline at dusk 1”, Walkerssk, CC0, via Wikimedia Commons.

02 / 56

Before we start · hands up

“Spinach became famous for its iron because a scientist put a decimal point in the wrong place, making it look ten times richer than it is.”

— a story told in textbooks, lectures and the BMJ

A · True B · False C · Not sure Hands up · 20 s

03 / 56

Verdict

A myth
about an error.

No decimal slip has ever been found in the sources. The “decimal” story itself spread through citations nobody checked.

Early iron values were high for other reasons: contaminated samples and fresh vs dried confusion.
Popeye’s creator credited spinach for vitamin A in 1932, not iron.
The decimal tale was repeated in the BMJ (1981) and copied from there, again and again.

Fluent + plausible + repeated = “known fact”. That is also how an AI hallucination travels from a chatbot into a poster.

Sutton, M. (2010). Spinach, iron and Popeye: Ironic lessons from biochemical history and the importance of healthy scepticism. Internet Journal of Criminology. Rekdal, O. B. (2014). Academic urban legends. Social Studies of Science, 44(4), 638–654. Schwarcz, J., “Spinach and iron”, McGill Office for Science and Society.

04 / 56

By 11:00 you should be able to…

1

Choose

the right AI tool for each format and audience: paper, figure, poster, talk, patient leaflet.

WRITE · ADAPT

2

Verify

every number, claim, citation and figure an AI tool touches, before a reader sees it.

CHECK · PROTECT

3

Disclose

AI use the way UKM, ICMJE and your target journal expect.

DECLARE · DEFEND

05 / 56

Today’s route

09:10

Who owns the message?

Audiences · UKM AI levels · evidence from nutrition

09:35

Writing for journals

Reporting guidelines · AI as critic · citations

10:15

Show, tell, translate

Figures · posters · talks · patient leaflets

10:45

Protect & disclose

Patient data · detectors · declarations

4 live demos

1 × 5-min break (10:10)

1 running example

0 paid tools needed

Demos use tools available through your UKM account (Gemini, NotebookLM) or free tools.

06 / 56

01

Block 1

Who owns the message?

One finding, many audiences, and where AI belongs.

⏱ 09:10 – 09:35

Photo: “Fountain and logo of UKM”, Wolfiewhite, CC0, via Wikimedia Commons (cropped).

07 / 56

THE MESSENGER

Wording, format, layout

AI drafts, shortens, translates, restyles and checks readability in seconds.

Cheap to change. Safe to delegate, with checks.

THE MESSAGE

Finding, size, certainty

What you found, how big it is, how sure you are, and what a clinician or patient should do with it.

Yours. Every rewrite is a chance for it to drift.

Let AI carry the message. Never let it change the meaning.

08 / 56

N

Nurul, MSc Clinical Nutrition

Our running example · fictional

HER ONE FINDING

Design: 12-week RCT, one KL clinic
People: 120 adults with type 2 diabetes
Compared: Pinggan Sihat low-GI meal plan vs usual dietary advice
Result: HbA1c −0.4 percentage points (95% CI −0.7 to −0.1); 8.2% → 7.8%
Caveats: Single centre · participants not blinded · 12 weeks only

ONE FINDING · SIX FORMATS · SIX CHANCES TO DRIFT

1

Abstract

250 words, CONSORT

2

Paper

Discussion, hedging

3

Figure

Honest chart

4

Poster

One-line headline

5

Talk

10 minutes + Q&A

6

Patient

Leaflet in BM

Fictional worked example. Pinggan Sihat Malaysia: Ministry of Health Malaysia healthy-plate concept (quarter grains, quarter protein, half vegetables and fruit).

09 / 56

Same finding, four readers

Journal reviewer

Abstract

“…mean difference in HbA1c −0.4 percentage points (95% CI −0.7 to −0.1; p = 0.01) at 12 weeks.”

preciseCIno advice

Conference visitor

Poster headline

“A low-GI Pinggan Sihat plan lowered HbA1c modestly in 12 weeks.”

one line“modestly”

Ward dietitian

Case meeting

“Worth offering; expect a small drop, about 4 mmol/mol. Adherence was 7 in 10.”

mmol/molpractical

Patient

Leaflet (BM)

“Pesakit yang mengikut pelan makan ini selama 12 minggu mempunyai bacaan gula darah purata yang sedikit lebih rendah.”

plain BM“sedikit”

The words change in every column. The size (“modest”, “small”, “sedikit”) and the certainty must not.

HbA1c conversion: 1 percentage point ≈ 10.9 mmol/mol (IFCC). Fictional example.

10 / 56

HMS SALISBURY · MAY 1747 · 12 SAILORS WITH SCURVY, 6 PAIRS

🧍🧍

Cider

🧍🧍

Elixir of vitriol

🧍🧍

Vinegar

🧍🧍

Sea water

🧍🧍

Oranges & lemons

🧍🧍

Garlic, mustard purge

174717531795

trialbookNavy ration

Dulu vs Kini · clinical nutrition’s first trial

1747: good evidence, weak persuasion

James Lind’s citrus pair recovered within days. One was fit for duty in six.
He published six years later, inside A Treatise of the Scurvy, tentatively and without pressing the case.
The Navy issued daily lemon juice in 1795: about 48 years after the trial.

Kini: evidence that is not communicated well changes nothing. AI can help you be read, but only if the evidence underneath is right.

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779.

11 / 56

A page from a 1597 edition of Ibn Sina’s Canon of Medicine — Ibn Sīnā, al-Qānūn (1597 ed.)

Dulu vs Kini · plain-language health writing

9th–11th century: medicine for people without a doctor

al-Rāzī wrote Man lā yaḥḍuruhu al-ṭabīb, “For one who has no physician”: a self-help manual using cheap, easy-to-find remedies.
Ibn Sīnā condensed medicine into a teaching poem of about 1,326 verses, the Urjūza fī al-ṭibb, so students could memorise it.
Same knowledge, different reader, different form.

Kini: a plain-language summary and a patient leaflet are the same idea. AI makes the rewrite fast; checking it is still the physician’s duty.

US National Library of Medicine, Islamic Medical Manuscripts: “Monographs” (al-Rāzī) and “Poetry” (Ibn Sīnā’s Urjūza, usually 1,326 verses). Image: “Avicenna canon 1597”, public domain, via Wikimedia Commons.

12 / 56

UKM’s five levels of AI use

Click a level ↓ · examples from research communication

1

No AI

2

AI for planning

3

AI collaboration

4

Full AI

5

AI exploration

Every UKM assessment declares one level. Not stated? Ask your course coordinator. Undeclared AI use breaches academic integrity.

UKM Guideline on Responsible Use of Generative AI in Teaching, Learning and Research Supervision (2025), AI-Enabled Assessment Levels Framework; adapted from the AI Assessment Scale (Perkins et al., 2024).

13 / 56

Influence comes from the use, not the tool

Same tool. Ask: “What am I handing over?”

🔒 gemini · UKM account

EMULATED

Suggest five poster titles for my finding. I’ll choose and edit.

✦

1. Small plate, smaller HbA1c…
2. …

Level 2 You keep the decision.

🔒 gemini · UKM account

EMULATED

Fix grammar in my abstract. Do not change any number, unit or hedge.

✦

Edited text with 6 tracked changes. Numbers unchanged.

Level 3 You keep the argument.

🔒 gemini · UKM account

EMULATED

Write the discussion for my diabetes trial.

✦

“This study demonstrates that low-GI diets are an effective treatment…”

Level 4 You hand over the thinking.

Emulated chat windows for teaching. Levels: UKM Guideline (2025).

14 / 56

The one rule for today

DELEGATE · TASKS

grammar, clarity, word count
reformatting for a new audience
first-draft translation you verify
layout and chart code you audit
predicting audience questions

KEEP · JUDGMENT

the finding and its size
the uncertainty: CIs, limits, hedges
the clinical recommendation
who the reader is and what they need
every citation and every number

Delegate tasks. Never delegate judgment.

15 / 56

Evidence 1 · Nutrients, 2024

Chatbot diet advice: mostly right, never complete

ChatGPT-3.5 answers vs international guidelines, 7 conditions · % rated appropriate

NAFLD

73%

Dyslipidaemia

71%

Type 2 diabetes

71%

Obesity

70%

CKD

70%

Hypertension

67%

Sarcopenia

56%

Two answers contradicted guidelines (snacking in obesity; supplements in NAFLD).
Key recommendations were missing; CKD protein advice ignored the stage.
A separate study: 56 allergy “robo-diets” included almond milk in a nut-free plan.

Plausible and incomplete is the hardest error to see, for patients and for authors.

Ponzo, V., et al. (2024). Is ChatGPT an effective tool for providing dietary advice? Nutrients, 16(4), 469. doi:10.3390/nu16040469. Niszczota, P., & Rybicka, I. (2023). The credibility of dietary advice formulated by ChatGPT: Robo-diets for people with food allergies. Nutrition.

16 / 56

🔒 chatbot · 2024

EMULATED

What can chloride be replaced with?

✦

Chloride can often be substituted with other halides, such as bromide, depending on the application. Context matters.

No health warning. No question about why you are asking.

The doctors’ own re-test with ChatGPT-3.5, paraphrased

Evidence 2 · Annals of Internal Medicine: Clinical Cases, Aug 2025

The salt swap that ended in psychosis

A 60-year-old man who had studied nutrition wanted to cut chloride from his diet.
After consulting ChatGPT, he replaced table salt with sodium bromide for 3 months.
Paranoia, hallucinations, an involuntary psychiatric hold: bromism, a 19th-century diagnosis.

The answer was true in a cleaning-chemistry context. In a diet context it was poison. Context is part of the message.

Eichenberger, A., Thielke, S., & Van Buskirk, A. (2025). A case of bromism influenced by use of artificial intelligence. Annals of Internal Medicine: Clinical Cases. doi:10.7326/aimcc.2024.1260. Emulated chat reconstructs the authors’ description; the patient’s own log was unavailable.

17 / 56

K

Keluarga Bahagia 🌸
Mak Long, Pak Ngah, Nurul, +9

Mak Long
AI kata air rebusan serai boleh gantikan ubat kencing manis. Sudah dua minggu Mak Long minum 👍08:02

Pak Ngah
Betul! Dalam video itu pun doktor kata sama. Kongsikan 🙏08:05

Mak Long, jangan berhenti makan ubat. Nanti Nurul terangkan, ya.08:31 ✓✓

Meme break · think about it

Your patients already ask AI.

Your leaflet now competes with a chatbot that is fast, kind, confident and always awake.

Accuracy alone will not win. Clear, warm, locally relevant writing might.

What would make Mak Long trust your one-page leaflet over her chatbot?

Original meme in a messaging-app format, created for this lecture. Names are fictional.

18 / 56

02

Block 2

Writing for journals

AI as critic, not author. Structure first, then style.

⏱ 09:35 – 10:10

Photo: “Tun Seri Lanang Library, UKM (221127) 01”, *angys*, CC BY-SA 4.0, via Wikimedia Commons (cropped).

19 / 56

What journals and UKM allow in 2026

Allowed

Language, clarity and readability
Organising your own content
Literature search support
Plain grammar checks (no declaration needed at Elsevier)

Allowed if declared

Substantive rewording or restructuring
Translation
AI in data analysis or figures: Methods
AI in writing: declaration / acknowledgments

Never

AI as an author
AI-written core content passed off as yours
Unverified references
Uploading a manuscript you are reviewing

Always read the target journal’s policy page before you submit. Image rules are usually the strictest.

Elsevier, Generative AI policies for journals (updated June 2026). ICMJE Recommendations, current version, sections on AI-assisted technology (icmje.org). UKM Guideline (2025), §2.2.

20 / 56

Before any prompt

Your reporting guideline is your outline

CONSORT 2025

Randomised trials · 30 items

Nurul’s trial

STROBE-nut

Nutritional epidemiology · 24 extra items

Dietary assessment, intake data

PRISMA 2020

Systematic reviews

Flow diagram + checklist

CARE

Case reports

e.g. the bromism case

CHART

Studies that test health chatbots

New in 2025

All free at the EQUATOR Network.

Best first job for AI: check your draft against the checklist. It is grounded, boring and genuinely useful.

Hopewell, S., et al. (2025). CONSORT 2025 statement. BMJ, 389, e081123. Lachat, C., et al. (2016). STROBE-nut. PLoS Medicine, 13(6), e1002036. Page, M. J., et al. (2021). PRISMA 2020. BMJ, 372, n71. Huo, B., et al. (2025). CHART statement, published in six journals incl. BMJ Medicine and JAMA Network Open.

21 / 56

DEMO 1 · ABSTRACT

Checklist audit of Nurul’s abstract

🔒 notebooklm · Nurul — Abstract check

EMULATED

SOURCES

CCONSORT 2025 checklist

AMy abstract (v3)

JJournal author guide

Check my abstract against the CONSORT 2025 checklist. Table: item · present? · quote from my abstract. Don’t rewrite anything.

Item	Present?	Evidence
Identified as randomised in title	Yes	“…a randomised controlled trial” 2
Primary outcome, effect size, CI	Yes	“−0.4 percentage points (95% CI…)” 2
Trial registration	Missing	—
Harms / adverse events	Missing	—
Blinding	Yes ✓	“participants were blinded”

WATCH FOR

Two real gaps found in seconds: registration and harms

A tick with no citation. Nurul’s participants were not blinded

Click every citation number. A tick is a claim too.

Emulated interface for teaching, not a product screenshot. Fictional abstract.

22 / 56

PHD Comics break

In 2009, a joke. In 2026, a two-second prompt.

An AI abstract fills the template perfectly: fluent, generic, and true of a hundred other trials.

Your abstract should be true of only yours: your population, your number, your caveat.

UKM treats an AI-written abstract as AI-written core content.

Strip: “Abstract Mad Libs”, Jorge Cham, PHD Comics (14 Jan 2009). “Piled Higher and Deeper” by Jorge Cham · www.phdcomics.com. Reproduced for non-commercial teaching under PHD Comics’ permission-by-email policy. UKM Guideline (2025), §2.2.

23 / 56

Change one verb

“Write it” → “Critique it”

Hands over the thinking

“Here are my results. Write a strong discussion section, 4–5 paragraphs, citing relevant studies.”

Argument, citations and claims come from the tool.

Keeps you the author

“Here are my results table and my draft discussion. List unsupported claims, causal words, missing comparisons and unit errors. Don’t rewrite.”

You write; the AI checks. Level 3, declared.

24 / 56

DEMO 2 · PAPER

Critique, not write: Nurul’s discussion

draft · Discussion ¶1

EMULATED

This trial proves that the Pinggan Sihat plan reverses type 2 diabetes. HbA1c fell by 0.4%, which is clinically important. These results agree with previous studies on low-GI diets. Malaysian clinics should adopt this plan for all patients.

Results table: 8.2% → 7.8% · −0.4 percentage points (95% CI −0.7 to −0.1) · single centre · 12 weeks

🔒 gemini · UKM account

EMULATED

✦

1. “proves” — one RCT cannot prove; suggest “suggests” or “indicates”.

2. “reverses” — 7.8% is still above most targets; no remission was measured.

3. “0.4%” — ambiguous: percentage points or relative? Add mmol/mol.

4. “previous studies” — no citations; which studies, what effect sizes?

5. “all patients” — one clinic, 12 weeks; generalisability not shown.

I have not rewritten your text.

Emulated interface; fictional text.

25 / 56

The quiet risk in “just fix my English”

Hedge drift: when polishing changes the science

NURUL’S SENTENCE

Higher fibre intake was associated with lower HbA1c, and the plan may help some patients improve glycaemic control.

“POLISHED” BY AN AI TOOL

Higher fibre intake ~~was associated with~~ reduces HbA1c, and the plan ~~may help some~~ helps patients ~~improve~~ achieve glycaemic control.

Association → cause · “linked to” becomes “reduces”

Some → all · qualifiers quietly vanish

Improve → achieve · a modest effect becomes a target met

Prompt: “Edit grammar only. Keep every hedge, number and unit. Show changes.” Then compare line by line (Word: Compare documents).

Illustrative example constructed for teaching.

26 / 56

TASK Critique my discussion against my results.
CONTEXT MSc clinical nutrition; RCT, n=120, T2DM; journal: Clinical Nutrition ESPEN.
SOURCES [results table] [my paragraph] [CONSORT 2025]
CONSTRAINTS Don’t rewrite. Don’t add references. Keep numbers, units and hedges.
FORMAT Numbered list: sentence → problem → why.
CHECK Say “unsure” rather than guess.

From prompting to context

Personas (“act as a professor of dietetics…”) matter less than real sources.
Constraints protect the meaning: numbers, units, hedges.
Asking for “unsure” makes errors visible instead of silent.

27 / 56

The Lancet · audit of ~2.5M biomedical papers

Fabricated references are rising fast

1 in 2,828

2023

1 in 458

2025

1 in 277

EARLY 2026

About 4,000 fabricated references in 2,810 papers.
Most affected papers had just one or two: a hallucination that slipped past.
This is our literature: biomedicine, nutrition, health sciences.

Topaz, M., et al. (2026). Fabricated citations: An audit across 2.5 million biomedical papers. The Lancet, 407(10541), 1779–1781. Bar heights proportional to rate.

28 / 56

Five checks for every AI-surfaced citation

1

Exists

DOI resolves; found in PubMed, Scopus or the publisher site.

2

Matches

Authors, year, title, journal all correct. No “Frankenstein” mixes.

3

Read

You read the relevant part, not just the abstract.

4

Supports

It says what you cite it for: same population, same outcome.

5

Standing

Not retracted (publisher page, Retraction Watch).

UKM rule: never include a citation you have not personally checked against the original. In nutrition, check #4 hardest: a rat study cited for a human claim is a common slip.

UKM Guideline (2025), §2.3: Accuracy, bias, fabricated references and integrity risks.

29 / 56

Open chat

Answers from training data
May be out of date
Can invent plausible sources
Origin of each claim unclear

Source-grounded (NotebookLM)

Answers only from your uploads
Every answer links to a passage
You choose the corpus
Still misreads: click the citation

A GOOD NOTEBOOK FOR A MALAYSIAN NUTRITION PAPER

Your included papers · CONSORT/STROBE-nut checklist · Malaysian Dietary Guidelines 2020 · CPG Management of Type 2 Diabetes Mellitus (6th ed., 2020) · journal author guide

National Coordinating Committee on Food and Nutrition (2020). Malaysian Dietary Guidelines 2020. Ministry of Health Malaysia (2020). CPG: Management of Type 2 Diabetes Mellitus (6th ed.).

30 / 56

Rehat · Stretch

5:00

Back at 10:15 · Block 3: Show, tell, translate

Renungan · before we make anything public

“The seeker after the truth is not one who studies the writings of the ancients and … puts his trust in them, but rather the one who suspects his faith in them and questions what he gathers from them.”

Pencari kebenaran bukanlah orang yang mempelajari tulisan orang terdahulu lalu mempercayainya begitu sahaja, tetapi orang yang meragui kepercayaannya terhadap tulisan itu dan mempersoalkan apa yang diperolehnya.

Ibn al-Haytham (c. 965–c. 1040)
Al-Shukūk ʿalā Baṭlamyūs (Doubts concerning Ptolemy), trans. A. I. Sabra

31 / 56

03

Block 3

Show, tell, translate

Figures, posters, talks and patient leaflets.

⏱ 10:15 – 10:45

Photo: “Petronas Towers at night in 2026 (DSCF0796)”, Yu Chu Chin, CC BY-SA 4.0, via Wikimedia Commons (cropped).

32 / 56

John Snow's 1854 map of cholera deaths in Soho, London — Snow, 1854: one bar per death

Nightingale, 1858 (schematic redraw)

Dulu vs Kini · figures that changed policy

A picture can win the argument

John Snow mapped cholera deaths around the Broad Street pump: the cluster made the case.
Florence Nightingale’s “rose” showed that far more soldiers died of preventable disease (blue) than of wounds (red).
Neither had software. Both chose one comparison that mattered.

Kini: AI draws a chart in seconds. Choosing which comparison deserves the picture is still your job.

Snow, J. (1855). On the Mode of Communication of Cholera (2nd ed.). Image: “Snow-cholera-map-1”, public domain, via Wikimedia Commons (cropped). Nightingale, F. (1858). Notes on Matters Affecting the Health, Efficiency, and Hospital Administration of the British Army. London: Harrison. Rose redrawn schematically for this slide; wedge sizes are illustrative, not her data.

33 / 56

AI and figures: four rules (Elsevier, June 2026)

ALLOWED · DECLARE IN CAPTION

Explanatory images

Flow charts, study schematics, timelines.

e.g. CONSORT-style flow diagram layout

ONLY IF REPRODUCIBLE

Data charts

Plots must come straight from your data, by a method in Methods.

e.g. code you ran on your dataset

NEVER CREATE OR ALTER

Primary images

Microscopy, radiology, patient photos.

e.g. food photos for intake, DXA, ultrasound

NO GENERAL IMAGE AI

Graphical abstracts

Use dedicated scientific illustration tools instead.

e.g. BioRender, your own drawing

Nutrition-specific trap: food photographs used to estimate intake are research data. Never “enhance” them with AI.

Elsevier, Generative AI policies for journals (updated June 2026): images, graphical abstracts and cover art.

34 / 56

Figure 1 · labels as published

RECONSTRUCTED

Integrity case · February 2024

The figure nobody read

A review in Frontiers in Cell and Developmental Biology used Midjourney figures, declared in the paper.
Labels were gibberish: “testtomcels”, “Dissisilcied”, “dck”. An anatomically absurd rat went viral.
It passed peer review and was retracted within days of publication.

Declaring AI use is necessary, not sufficient. Read every label, every axis, every unit.

Reported in Popular Science, “A ridiculous AI-generated rat penis made it into a peer-reviewed journal” (Feb 2024); Frontiers retraction notice (2024). Figure reconstructed schematically with the reported labels; not the original image.

35 / 56

DEMO 3 · FIGURE

Gemini in Colab: generate the chart, then audit it

🔒 colab · nurul_hba1c.ipynb

EMULATED

✦ Plot mean HbA1c at baseline and 12 weeks by group, from df.

import matplotlib.pyplot as plt
m = df.groupby(['group','week'])['hba1c'].mean().unstack()
m.T.plot(kind='bar')
plt.ylim(7.6, 8.4)      # ← truncated axis
plt.ylabel('HbA1c')        # ← unit missing
plt.title('Diet cuts HbA1c') # ← causal claim

A 5% relative change
looks like 75%

AUDIT BEFORE IT LEAVES COLAB

Axis starts at 7.6: exaggerates the drop

No unit, no 95% CI, no n per group

Title makes a causal claim the caption can’t support

Keep the code: your figure is now reproducible, as the journal requires

Emulated interface; synthetic data. Bars drawn to show the visual effect of the truncated axis.

36 / 56

MethodsRCT · n=120 · T2DM · KL clinic · 12 weeksInterventionPinggan Sihat low-GI plan vs usual adviceLimitsSingle centre; unblinded

MAIN FINDING

A low-GI Pinggan Sihat plan lowered HbA1c modestly in 12 weeks.

−0.4 percentage points (95% CI −0.7 to −0.1)

FigureChange in HbA1c by group, with 95% CIAsk meAdherence was 71%. What would help it?

Mock-up in the #betterposter layout: the finding in plain words, centre stage

Poster

Visitors give you seconds, not minutes

AI for: 10 headline variants; cutting 600 words to 200; “where does the eye land first?” on a photo of your draft.
Not for: the data figure, or a headline stronger than your CI.
Test the headline on a non-nutrition friend in 10 seconds.

Layout inspired by Morrison, M. (2019), #betterposter. Mock-up with fictional content.

37 / 56

🔒 notebooklm · Nurul — Talk prep

EMULATED

From my abstract and paper, list the 6 hardest questions a clinical audience might ask. Cite the passage each comes from.

✦

1. Participants weren’t blinded: could expectation explain the drop? 3

2. Is 0.4 points clinically meaningful for your population? 5

3. How did you measure adherence? 2

4. Will it hold beyond 12 weeks? 6

Talk · 10 minutes + Q&A

AI as a rehearsal partner

Hard questions, grounded in your paper, before the real audience asks them.
Timing: about 130 words a minute; ask it to cut your script to 1,300 words.
Audio Overview: hear two AI hosts discuss your paper. What they stress is what a listener takes away.

You still give the talk. Answers you only read from AI will not survive a follow-up question.

Emulated interface. NotebookLM features: Audio Overviews, Video Overviews, and (since late 2025) Slide Decks and Infographics; Google, “8 ways to make the most out of Slide Decks in NotebookLM” (Dec 2025).

38 / 56

One click makes slides, a video, an infographic. Now what?

Good as a draft

A first storyboard for your talk
Spotting what your paper sounds like it says
Study aid for your own revision

Not a deliverable until checked

Every number, unit and arrow on every frame
Generated pictures of food, organs, people
Voices and visuals you must still declare

Try it: generate an infographic of Nurul’s paper, then circle every claim that is stronger than her CI allows. That list is your editing job.

39 / 56

Evidence 3 · J Med Internet Res, 2025

LLMs can make leaflets readable

US heart, cancer and stroke patient pages · reading grade level · NIH target ≤ 6

Original

10.1

ChatGPT

7.6

Gemini

6.6

Claude

5.6

Bar length ∝ grade level (scale 0–12)

Texts got about half as long.
2 of 60 Gemini and Claude rewrites were inaccurate.
Omitted content was not assessed, and only English was tested.

Readability up, completeness unknown. In BM, grade formulas don’t apply: test with real patients.

Will, J., Gupta, M., Zaretsky, J., Dowlath, A., Testa, P., & Feldman, J. (2025). Enhancing the readability of online patient education materials using large language models. J Med Internet Res, 27, e69955. doi:10.2196/69955

40 / 56

DEMO 4 · PATIENT

From finding to a BM leaflet, and back again

🔒 gemini · draft 1

EMULATED

AI DRAFT (FORMAL BM)

Pelan makan Pinggan Sihat terbukti dapat menyembuhkan kencing manis dalam 12 minggu. Paras HbA1c anda akan turun. Anda tidak perlu lagi bergantung kepada ubat.

4 problems: cure claim · jargon · certainty · dangerous advice

nurul · draft 2 (edited by hand)

EMULATED

NURUL’S VERSION

Dalam satu kajian, pesakit yang mengikut pelan makan Pinggan Sihat selama 12 minggu mempunyai bacaan gula darah purata yang sedikit lebih rendah. Ujian ini dipanggil HbA1c. Teruskan ubat anda dan bincang dengan pegawai dietetik sebelum menukar pemakanan.

THEN CHECK

Back-translate to English: did “sedikit” survive?

Teach-back with 3 patients: can they say it in their own words?

Clinician review before printing

Emulated interface; fictional text. Teach-back: ask patients to explain the information in their own words to confirm understanding.

41 / 56

Writing for the public

Same number, two headlines

Relative · sounds big

“Blood sugar cut by 5%”

0.4 ÷ 8.2 ≈ 5%. Correct arithmetic, misleading impression.

Absolute · honest

“HbA1c 8.2% → 7.8%”

A small drop; most participants still above target.

1 finding

1 absolute number

1 limitation

1 link to the paper

Prompt any caption with: “Use absolute numbers. Keep the limitation. No cure words.”

Fictional numbers from the worked example.

42 / 56

📄

The paper: “…was associated with a modest reduction…”

📰

The press release: “Diet plan lowers blood sugar”

📱

The AI caption: “SCIENCE PROVES rice causes diabetes 🍚❌”

Meme break · think about it

Every new audience is a chance for the claim to grow.

Nobody in the chain lied. Each step just made it a little more shareable.

Read the last version with the first one beside it. If you would not sign the caption, don’t post it.

Original meme in a three-panel escalation format, created for this lecture.

43 / 56

Nurul’s route: one finding, six formats

01

Abstract

CONSORT check

AI: found 2 gaps

Nurul: caught a false tick

02

Paper

Discussion

AI: listed 5 problems

Nurul: rewrote by hand

03

Figure

HbA1c chart

AI: wrote the code

Nurul: fixed axis, CI, title

04

Poster

Headline

AI: 10 variants

Nurul: kept “modestly”

05

Talk

Q&A prep

AI: hard questions

Nurul: answered herself

06

Patient

BM leaflet

AI: first draft

Nurul: removed “cure”, tested

Six rewrites. One meaning. Every check was hers.

44 / 56

04

Block 4

Protect & disclose

Patient data, detectors, and a declaration that says something.

⏱ 10:45 – 10:58

Photo: “Merdeka 118 illuminated at night”, Wlsrudals, CC0, via Wikimedia Commons (cropped).

45 / 56

What never goes into a public AI tool

👤

Patients

Diet records, food photos, lab results, case notes, interview recordings, even “de-identified” small samples.

📄

Unpublished work

Manuscripts, theses, grant proposals, and anything you are asked to review.

🏥

Institutional

Hospital protocols, internal reports, data under an ethics or data-sharing agreement.

Health data is sensitive personal data under Malaysia’s PDPA 2010. Unsure? UKM’s default: don’t upload. Use synthetic data to test prompts.

Personal Data Protection Act 2010 (Act 709), as amended 2024. UKM Guideline (2025), §2.3: Data security, confidentiality and privacy.

46 / 56

PHD Comics strip: The Academic Review Process

PHD Comics break

2014: the goldfish reviews it. 2026: the goldfish might be an LLM.

Reviewers must never upload a manuscript into an AI tool. They may use a private tool to polish their own words, and must declare it.

Assume your submission is read by humans who know these rules, and some tools that don’t.

Strip: “The Academic Review Process”, Jorge Cham, PHD Comics (12 Nov 2014). “Piled Higher and Deeper” by Jorge Cham · www.phdcomics.com. Reproduced for non-commercial teaching under PHD Comics’ permission-by-email policy. Reviewer rules: Elsevier, Generative AI policies for journals (June 2026).

47 / 56

About AI detectors

Scores are probabilistic. False positives happen, more often for non-native English writers.
UKM won’t decide misconduct on a score alone.
Concerns go to academic judgment and an oral check of your understanding.

Your best protection

Drafts, version history, notes, analysis code, an AI-use log, and a clear declaration.

“No adverse academic decision shall be made solely on algorithmic inference.”

UKM Guideline (2025), §2.3: Use of AI-detection tools (e.g., Turnitin, ZeroGPT, GPTZero). Liang, W., et al. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779.

48 / 56

Try it · click the options

Disclose specifically, not vaguely

“AI was used to assist in writing” tells an editor nothing.

Where: Elsevier → separate AI declaration; ICMJE → acknowledgments (writing) or Methods (analysis, figures). Template: UKM Appendix B2 · GAIDeT

AI-use declaration builder (GAIDeT-style)

INTERACTIVE

TASKS DELEGATED TO AI

✓Language editing ✓Checklist check ✓Translation (BM) ✓Readability ✓Figure code ✓Q&A rehearsal

TOOLS

✓Gemini ✓NotebookLM ✓Gemini in Colab

Builder inspired by GAIDeT: Suchikova, Y., Tsybuliak, N., Teixeira da Silva, J. A., & Nazarovets, S. (2025). Accountability in Research. For submissions, follow the journal’s own form.

49 / 56

Five minutes a week

Keep an AI-use log from today

sheets · AI-use log · Nurul MSc 2026

EMULATED

Date	Tool	Task	Input (sanitised)	Kept	Verified how
02 Nov	NotebookLM	CONSORT check, abstract	Abstract + checklist	2 gaps	Checked each item myself
05 Nov	Gemini	Critique of discussion	My paragraph + table	4 / 5 points	Rewritten by hand
09 Nov	Gemini in Colab	HbA1c figure code	Synthetic sample	Edited	Axis, CI, units checked
14 Nov	Gemini	BM leaflet draft	Published finding only	Structure	Back-translation + 3 patients

It becomes your declaration, your evidence if questioned, and your memory at viva time.

Emulated spreadsheet with fictional entries.

50 / 56

Click a tool to open it

A communication toolbox at zero cost

AllAINon-AI

GeminiAI

Critique, editing, translation drafts

UKM account

NotebookLMAI

Grounded checks, Q&A prep, audio/video drafts

UKM account

Google ColabAI

Reproducible charts with Gemini help

Free

Zoteronon-AI

References and citation styles

Open source

PubMednon-AI

Check that a citation exists

Free

EQUATOR Networknon-AI

CONSORT, STROBE-nut, PRISMA, CHART

Free

Datawrappernon-AI

Clean charts for posters and the web

Free tier

BioRendernon-AI

Scientific illustration, graphical abstracts

Free tier

CanvaAI

Posters and slides; check any AI images

Free tier

Hemingway Editornon-AI

English readability check

Free web

Retraction Watchnon-AI

Is the paper still standing?

Free

PRPM Dewan Bahasanon-AI

Check formal BM terms

Free

Free tiers and features change; check before relying on one. Canva’s AI image features fall under the figure rules in Block 3.

51 / 56

Six things to remember

01

Carry the message, keep the meaning.

Size and certainty survive every rewrite.

02

Structure before style.

Your reporting guideline is the outline.

03

Critique, don’t write.

You draft; AI finds the gaps.

04

Every number, citation and label is yours.

Check them; nobody else will.

05

Patients get plain, warm, tested words.

Absolute numbers, no cure words, teach-back.

06

Never upload what you must protect. Declare the rest.

PDPA, UKM levels, journal policy, AI-use log.

52 / 56

Practice task · 30 minutes

One finding, three versions

1

Pick

One result from your own project, or from a published nutrition RCT.

2

Write three

An abstract sentence, a poster headline, a formal BM patient sentence. Yourself first.

3

Critique & log

Ask AI to critique all three for drift in size and certainty. Log it. Write a two-line declaration.

Your course coordinator’s brief and AI level take precedence over this suggestion.

53 / 56

Closing · every generation fears its new tool

We have been here before

c. 370 BCE

WRITING · PLATO, PHAEDRUS

“[T]his invention will produce forgetfulness in the minds of those who learn to use it… the appearance of wisdom, not true wisdom.”

King Thamus, judging the invention of writing

1986

CALCULATORS · WASHINGTON, D.C.

Mathematics teachers picketed the NCTM meeting over calculators in classrooms. Textbook author John Saxon warned of “a tremendous potential for destruction” of students’ skills.

2026

GENERATIVE AI · THIS ROOM

“Will AI make us stop writing, and stop thinking?”

The case report, the rat and the citation audit say the risk is real, if we stop checking.

Plato, Phaedrus 275a, trans. H. N. Fowler (1914), Loeb Classical Library (public domain). “The great calculator debate”, The Christian Science Monitor, 9 May 1986.

54 / 56

Closing · were they wrong?

Right about the risk. Wrong about the ending.

Writing

Memory did change. Writing also gave us journals, case reports, leaflets: everything in today’s lecture.

Calculators

Mental arithmetic did weaken. We still teach it first, then hand over the calculator, and expect you to notice a wrong answer.

Learn to say it yourself. Then delegate the wording. Always know when the meaning changed.

P.S. The only reason we know Socrates distrusted writing is that Plato wrote it down.

55 / 56

Questions & discussion

Terima kasih
Thank you

akmal@ukm.edu.my · akmal.app

Let AI carry the message. Never let it change the meaning.

Photo: “Gong UKM”, Wee Hong, CC BY-SA 4.0, via Wikimedia Commons (cropped).

56 / 56

References & provenance

Eichenberger, A., Thielke, S., & Van Buskirk, A. (2025). A case of bromism influenced by use of artificial intelligence. Annals of Internal Medicine: Clinical Cases. doi:10.7326/aimcc.2024.1260

Elsevier. (2026). Generative AI policies for journals (updated June 2026).

Hopewell, S., et al. (2025). CONSORT 2025 statement: Updated guideline for reporting randomised trials. BMJ, 389, e081123.

Huo, B., et al. (2025). Reporting guideline for chatbot health advice studies: The CHART statement. BMJ Medicine; JAMA Network Open (joint publication).

ICMJE. Recommendations for the conduct, reporting, editing, and publication of scholarly work in medical journals (current version), AI-assisted technology. icmje.org

Lachat, C., et al. (2016). Strengthening the Reporting of Observational Studies in Epidemiology–Nutritional Epidemiology (STROBE-nut). PLoS Medicine, 13(6), e1002036.

Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779.

Lind, J. (1753). A Treatise of the Scurvy. Edinburgh.

Ministry of Health Malaysia. (2020). Clinical Practice Guidelines: Management of Type 2 Diabetes Mellitus (6th ed.).

National Coordinating Committee on Food and Nutrition. (2020). Malaysian Dietary Guidelines 2020. Ministry of Health Malaysia.

Nightingale, F. (1858). Notes on Matters Affecting the Health, Efficiency, and Hospital Administration of the British Army. London: Harrison.

Niszczota, P., & Rybicka, I. (2023). The credibility of dietary advice formulated by ChatGPT: Robo-diets for people with food allergies. Nutrition.

Page, M. J., et al. (2021). The PRISMA 2020 statement. BMJ, 372, n71.

Perkins, M., Roe, J., & Furze, L. (2024). The AI Assessment Scale revisited. arXiv:2412.09029.

Ponzo, V., et al. (2024). Is ChatGPT an effective tool for providing dietary advice? Nutrients, 16(4), 469. doi:10.3390/nu16040469

Rekdal, O. B. (2014). Academic urban legends. Social Studies of Science, 44(4), 638–654.

Snow, J. (1855). On the Mode of Communication of Cholera (2nd ed.). London: John Churchill.

Suchikova, Y., Tsybuliak, N., Teixeira da Silva, J. A., & Nazarovets, S. (2025). GAIDeT: Generative AI Delegation Taxonomy. Accountability in Research.

Sutton, M. (2010). Spinach, iron and Popeye: Ironic lessons from biochemical history and the importance of healthy scepticism. Internet Journal of Criminology.

Topaz, M., et al. (2026). Fabricated citations: An audit across 2.5 million biomedical papers. The Lancet, 407(10541), 1779–1781.

Universiti Kebangsaan Malaysia. (2025). Guideline on Responsible Use of Generative AI in Teaching, Learning and Research Supervision.

US National Library of Medicine. Islamic Medical Manuscripts at the NLM: Monographs (al-Rāzī); Poetry (Ibn Sīnā).

Will, J., et al. (2025). Enhancing the readability of online patient education materials using large language models. J Med Internet Res, 27, e69955.

PHD Comics strips © Jorge Cham, used with permission requested by email. Photos credited on each slide. Emulated interfaces are teaching reconstructions, not product screenshots.

ScientificCommunicationusing AI

A mythabout an error.

By 11:00 you should be able to…

Choose

Verify

Disclose

Today’s route

Who owns the message?

Writing for journals

Show, tell, translate

Protect & disclose

Who owns the message?

Wording, format, layout

Finding, size, certainty

Abstract

Paper

Figure

Poster

Talk

Patient

Same finding, four readers

Abstract

Poster headline

Case meeting

Leaflet (BM)

1747: good evidence, weak persuasion

9th–11th century: medicine for people without a doctor

UKM’s five levels of AI use

No AI

AI for planning

AI collaboration

Full AI

AI exploration

Same tool. Ask: “What am I handing over?”

The one rule for today

Chatbot diet advice: mostly right, never complete

The salt swap that ended in psychosis

Your patients already ask AI.

Writing for journals

What journals and UKM allow in 2026

Your reporting guideline is your outline

CONSORT 2025

STROBE-nut

PRISMA 2020

CARE

CHART

Checklist audit of Nurul’s abstract

In 2009, a joke. In 2026, a two-second prompt.

“Write it” → “Critique it”

Critique, not write: Nurul’s discussion

Hedge drift: when polishing changes the science

NURUL’S SENTENCE

“POLISHED” BY AN AI TOOL

From prompting to context

Fabricated references are rising fast

Five checks for every AI-surfaced citation

Exists

Matches

Read

Supports

Standing

Open chat

Source-grounded (NotebookLM)

Show, tell, translate

A picture can win the argument

AI and figures: four rules (Elsevier, June 2026)

Explanatory images

Data charts

Primary images

Graphical abstracts

The figure nobody read

Gemini in Colab: generate the chart, then audit it

Visitors give you seconds, not minutes

AI as a rehearsal partner

One click makes slides, a video, an infographic. Now what?

Good as a draft

Not a deliverable until checked

LLMs can make leaflets readable

From finding to a BM leaflet, and back again

Same number, two headlines

Scientific
Communication
using AI

A myth
about an error.

Terima kasih
Thank you