Session A ended with a verified body of evidence. Session B turns it into things that leave the room: a defensible framework, a proposal outline, an honest figure, and a disclosure statement you can sign.
Modul 2 (2)
Argument, conceptual framework, methodology and academic polishing — ending in a grant proposal outline.
Modul 3
Analysis plans, cleaning logs, Advanced Data Analysis, and figures that do not mislead.
Closing
Privacy, authorship, disclosure that survives review, and a 30-day plan.
Session B · how this file works
A separate file, the same contract
Session B is self-contained. Your Session A result is not needed to start, and can be carried over if you kept it.
What is assessed
Sixteen components across three blocks, listed below with the reason each one exists. Your scorecard stays hidden until you open it at the end — work first, read it after.
Where your work lives
In this browser tab, and nowhere else. Closing it clears everything. At the end you can copy a short code or download a file to hand back — that is the only way anything leaves this page.
Keys you will want
O overview · S ledger · N facilitator notes · T timer · B language · D light or dark · ? full list
Carrying your Session A result forward
Optional. If you copied your Session A code at the end of the first half-day, paste it here to see both sessions in one view. If you did not, nothing is lost — Session B stands on its own.
Session B · what this session covers
From the Session A matrix to a proposal and a figure
Everything today builds on the literature matrix and problem statement you produced in Session A — but if you were not there, each lab stands alone.
Block
What we cover
You leave with
Modul 2 (2) — proposal writing
The five-link argument, a conceptual framework you can defend, a methodology draft, academic polishing and BM–EN translation
Grant proposal outline · framework justification
Modul 3 — data analysis & research productivity
Pre-specify the analysis, log the cleaning, run an Advanced Data Analysis session, judge which figures are honest
Cleaned dataset · 2 publication-ready charts
Closing — integrity & direction
Where disclosure goes, what you may paste, and the statement you will actually submit
AI-use statement · 30-day plan
The through-lineSession A produced evidence. This block turns it into an argument, then into a proposal. The data block turns data into figures that do not mislead. The closing block is what makes all of it publishable. Nothing here is a separate topic — it is one chain, and you are holding the front of it already.
Re-entry · before anything new
Three questions from Session A, from memory
Do not open Session A. The struggle to recall is the mechanism — getting it wrong and then seeing why beats getting it right from a slide.
Q1Why does an ungrounded model produce a reference that does not exist?
Its training data contained fake references
No. The training data is overwhelmingly real. The problem is what the model does at generation time, not what it read.
It predicts what a reference should look like, and a plausible shape is all it is optimising for
Correct. Author, journal, year and DOI are each individually plausible continuations. Nothing in the process checks that the combination exists.
The reference existed but was retracted
That is a different and much rarer problem — and one Scite or a database check would catch. Fabrication is generation, not retraction.
Assessed because everything in Session B rests on knowing why the failure happens, not just that it does. CLO1.
Q2You supplied twelve sources and asked for a synthesis. What is the remaining fabrication risk?
None — it can only use what I gave it
The citations are safe. The claims attributed to them are not: a model can attach your source to a finding that source never reported.
It may attribute a claim to the wrong source, or state a finding more strongly than the paper did
Correct, and this is now the dominant error mode. Grounding fixed the reference list; it did not fix the mapping between claim and source.
It will refuse if it is unsure
Models are trained to be helpful, and a fluent answer scores better than a refusal. Do not rely on abstention.
Assessed because Module B1 asks you to draft from sources. This is the failure you must watch for while doing it. CLO1, CLO4.
Q3What made the “same first author, different paper” row so dangerous?
The DOI was malformed
It was not — it resolved perfectly. That was exactly the problem.
Most of the fields checked out, so a partial check passed it
Correct. A partial verification is worse than none, because it produces false confidence. The DOI must resolve to the named paper, not merely resolve.
The journal did not exist
The journal was real. Fabrications that fail obviously are not the ones that reach print.
Assessed because partial verification is the habit that lets errors survive peer review. CLO3.
Session B · block 1 of 3
Modul 2 (2) — research: proposal writing
Session A produced verified evidence. This block turns it into an argument, a framework and a methodology — the three things a proposal is actually made of.
What we cover
The five-link argument chain, adversarial review, conceptual frameworks with warrants, methodology drafting, and academic polishing across BM and English.
You will build
A defensible problem statement, a construct-and-relationship table you could draw, a methods draft, and a grant proposal outline.
Where authorship is lost
This is the block where it happens, which is why it ends on the ownership boundary rather than starting there.
CLO4
Modul 2 (2) · argument
A paper is five claims in a chain
Click each link. If any one is weak, no amount of rewriting the sentences will save the paper — and this is the single most common reason a draft feels wrong without the author knowing why.
01
Something matters
02
Something is unknown
03
This is knowable
04
Here is what we found
05
Therefore
Where drafts actually breakEditors rejecting without review most often cite a mismatch between the question asked and the method used — link 03. It is a structural fault, which is why line-editing never fixes it and why authors so often cannot see it in their own work.
Modul 2 (2) · lab
Build your five links, then stress-test them
Level 1 drafts the chain. Level 2 attacks it. Run both — the second is where the value is.
Lab B1.1 · the chain and its criticDo this nowBring your own topic. Placeholders are marked in square brackets.
Give it your gap sentence from Session A. If you do not have one, use any topic you know well.
ROLE: Senior researcher who reviews for a general medical journal.
TASK: Turn my raw material into a five-sentence problem statement.
STRUCTURE — exactly one sentence per link:
1. SIGNIFICANCE: a fact a hostile reader would still accept
2. GAP: what is unresolved AND why existing work could not resolve it
3. FEASIBILITY: why my design can close that specific gap
4. AIM: what I will produce, stated as an outcome not an activity
5. SO WHAT: whose decision could change, named concretely
CONSTRAINTS:
- No citations. I will attach my own.
- Do not use the words "novel", "gap in the literature", or "further research is needed"
- If link 3 does not actually close link 2, say so instead of writing sentence 3
MY MATERIAL:
Topic: [ ]
What I think is unresolved: [ ]
My planned design and measure: [ ]
The last constraint is the important one
Expected output
What a usable output looks like:
1. Medication non-adherence accounts for a substantial share of preventable hospital readmission in older adults with polypharmacy. 2. Pharmacist-led review improves adherence in trials, but adherence was measured by self-report in most of them, a method known to overestimate; whether the effect survives objective measurement is unresolved. 3. Linking review to dispensing-refill records allows adherence to be measured without relying on patient report, addressing precisely that limitation. 4. This study will estimate the effect of structured pharmacist review on refill-based adherence over twelve months. 5. A positive result would give hospital pharmacy directors a defensible basis for funding review clinics; a null result would redirect attention to the measurement artefact.
Then notice what it refused to do. If your design cannot close your gap, a well-constrained model will say so at step 3 rather than writing a sentence that papers over the break. That refusal is the most valuable output on this slide.
Link 2
“AND why existing work could not resolve it” is the clause that separates a gap from a complaint. Without it you get “few studies have examined”.
Banned words
Blocking “novel” and “further research is needed” forces specificity. These phrases are where vague thinking hides, in human writing as much as machine writing.
Escape clause
Explicitly permitting the model to refuse turns it from a text generator into a diagnostic. Always give a model a way to tell you no.
Five sentences you can defend beat five paragraphs you cannot.
Paste the five sentences you just produced. This is the step most people skip and most need.
ROLE: The most demanding reviewer on the panel. You are fair, but you reject work that does not hold together.
TASK: Attack the problem statement below. Do not improve it. Do not be encouraging.
FOR EACH of the five links, report:
- The single strongest objection a reviewer would raise
- Whether that objection is fatal, serious, or cosmetic
- What evidence or change would neutralise it
THEN answer one question directly: if you had to reject this on ONE sentence, which sentence, and why?
Do not soften your language. Do not end with encouragement.
STATEMENT:
[paste your five sentences]
Expect this to sting
Expected output
The shape of a useful attack:
Link 2 — SERIOUS. The claim that self-report overestimates adherence is asserted, not evidenced. If the overestimation is non-differential between arms, it may bias toward the null rather than inflate the effect — which would weaken rather than support the rationale. Neutralised by citing a validation study reporting the direction of bias.
Link 3 — FATAL. Refill records measure collection, not ingestion. Substituting one imperfect proxy for another does not close the stated gap; it relocates it. Neutralised only by acknowledging this explicitly and reframing the aim as a comparison of measurement methods.
If I had to reject on one sentence: sentence 3. It claims to solve a measurement problem with a different measurement problem, and no reviewer in this field will miss it.
Now the important part. That third-sentence objection is real, and it is better to meet it here than in a decision letter. Note that the model also caught something more subtle in link 2 — a bias-direction error — which is the kind of thing you would want a co-author to catch.
“Do not improve it”
Without this, the model rewrites your statement into something better and you learn nothing about the flaw. Separating critique from repair is the whole technique.
Fatal / serious / cosmetic
Forcing a severity grade stops the model listing twenty equal-weight quibbles. You need to know what to fix first.
“Do not end with encouragement”
Models are trained toward agreeableness; without this line you get a real critique followed by a paragraph that undoes it. Sycophancy is the default and must be explicitly suppressed.
A critique you can act on is worth more than a draft you cannot defend.
Modul 2 (2)
A framework is an argument with arrows
Every box is a construct you must be able to measure. Every arrow is a claim you must be able to defend. A diagram with unlabelled arrows is decoration.
ConstructThe thing itself
→
OperationalHow you measure it
→
RelationshipDirection and sign
→
WarrantTheory or evidence for it
→
TestWhat would falsify it
Lab B2.0 · framework and its criticDo this nowUse the problem statement from Lab B1.1, or a proposal you have in progress.
ROLE: Doctoral examiner who reads conceptual frameworks for a living.
TASK: Build a conceptual framework from my problem statement below.
RETURN, in this order:
1. A CONSTRUCT TABLE — one row per construct: name | conceptual definition |
how I will operationalise it | one named instrument or data source
2. A RELATIONSHIP TABLE — one row per arrow: from | to | direction and expected
sign | the theory or empirical warrant for it | what would falsify it
3. A TEXT DESCRIPTION of the diagram, precise enough that I could draw it
4. THE MODERATORS AND CONTROLS I have not mentioned but should have
RULES:
- Any construct I cannot measure with a named instrument gets flagged, not smoothed over
- Any arrow without a warrant gets flagged as an assumption, not written as a finding
- Do not invent citations. Where a warrant needs a source, write [WARRANT NEEDED]
- Tell me plainly if my framework is really two frameworks that should be separate studies
MY PROBLEM STATEMENT:
[ paste your five sentences from Lab B1.1 ]
MY DISCIPLINE AND SETTING: [ ]
Every arrow needs a warrant
Expected output
Two rows from the relationship table:
| Structured review → Adherence | Positive | Warrant: repeated-contact interventions improve adherence across chronic disease [WARRANT NEEDED — cite a trial in your population] | Falsified if refill-based adherence does not differ at 12 months |
| Polypharmacy → (moderates) review → adherence | Weakens the effect | ASSUMPTION — no warrant supplied. You have stated this as though it were established. Either find evidence or reframe it as an exploratory moderation. | Falsified if the interaction term is null |
And the structural warning: “Constructs 3 and 6 — ‘health literacy’ and ‘pharmacist communication quality’ — are each measured by a different instrument, on a different population, at a different time point from your primary outcome. This is a second study. Including it will weaken the first without adequately powering the second.”
That last paragraph is the single most useful thing on this slide. Overloaded frameworks are the most common reason a proposal reads as unfocused, and authors almost never see it in their own work.
Named instrument
Forcing a named instrument per construct is what separates a framework from a mind map. If you cannot name one, you have a concept, not a construct.
“What would falsify it”
An arrow that nothing could falsify is not a hypothesis. This column quietly removes the arrows that were only there to look complete.
“Two studies” clause
Permission to tell you the proposal is overloaded. Reviewers say this constantly; getting it now costs you nothing.
A framework you can falsify is a framework you can fund.
Modul 2 (2) · methods & language
Draft the methods, then polish without losing the meaning
Two tabs. The second one is where most of the risk in this whole workshop actually sits.
Lab B2.0b · methods and polishDo this now
ROLE: Methodologist reviewing a grant application for a national funder.
TASK: Draft the methods section from my framework and constraints below.
STRUCTURE: Design | Setting and population | Sampling and size | Variables and
instruments | Procedure | Analysis | Ethics | Limitations
RULES:
- Write in past-tense-ready form so it converts to a report later without rewriting
- State the sample size logic explicitly. If I have not given you the parameters
needed for a calculation, list exactly which ones are missing rather than inventing them
- For every instrument, note whether validation in MY population exists or is assumed
- The limitations paragraph must name the limitation a reviewer will raise FIRST,
not the safest one
- Do not cite anything. Mark evidence I must supply as [MY SOURCE]
MY FRAMEWORK: [ paste construct and relationship tables ]
DESIGN I INTEND: [ ]
ACCESS AND CONSTRAINTS: [ site, timeframe, budget, ethics status ]
It should refuse to guess your sample size
Expected output
The sampling paragraph you want to see:
“Sample size cannot be calculated from what you have supplied. To compute it I need: the expected adherence proportion in the control arm, the smallest difference you would regard as clinically meaningful, alpha, power, and your expected attrition. You have supplied none of these. I have not assumed values, because an assumed effect size is the most common way an underpowered study is justified after the fact.”
And the limitations paragraph: “The limitation a reviewer will raise first is not your single-site design — it is that refill records measure collection rather than ingestion, which means your primary outcome does not directly measure the construct in your framework. Address this in the methods, not the discussion.”
Compare that with what an unconstrained prompt returns: a plausible n of 384, a citation-free reference to Cochran’s formula, and a limitations paragraph about generalisability. Fluent, complete, and useless.
“List what is missing”
Turns a guess into a checklist. You now know exactly what to bring to your statistician, which is a better outcome than a number you could not defend.
“Validation in MY population”
An instrument validated elsewhere is an assumption, not a method. This flag catches it before a reviewer does.
First objection, not safest
Limitations paragraphs are usually written to look humble. This instruction makes them useful instead.
A methods section that names what is missing is stronger than one that fills the gap.
Language help is the most common academic use of AI and the least examined. This prompt makes every change visible.
ROLE: Bilingual academic editor working between Bahasa Malaysia and English
in the health sciences.
TASK: Improve the passage below WITHOUT changing any claim.
RETURN THREE THINGS:
1. The revised passage
2. A CHANGE TABLE — original phrase | revised phrase | reason (grammar /
concision / register / term choice)
3. A MEANING-DRIFT WARNING — any place where the revision made a claim
STRONGER or WEAKER than the original, flagged explicitly
HARD RULES:
- Never remove a hedge. If the original says "may be associated with", it stays.
- Never add a hedge either. If I committed to a claim, I meant to.
- Keep technical terms in their established form; do not "simplify" a term of art
- If a sentence is unclear because the underlying THINKING is unclear, say so
instead of smoothing the prose — tell me what decision I have not made
- If translating: give the standard Malaysian academic term, and where none is
settled, give the English term with a Malay gloss rather than coining one
PASSAGE: [ ]
DIRECTION: [ polish in English / BM to EN / EN to BM ]
The drift warning is the whole point
Expected output
A change table row: | “Keputusan menunjukkan bahawa intervensi ini berkesan” | “The results indicate that the intervention improved adherence in this sample” | Term choice: berkesan is a verdict; improved adherence in this sample is what your data supports |
A drift warning: “STRENGTHENED. Your original said mungkin menyumbang kepada. My first revision rendered this as contributes to, which removes your hedge and commits you to causation. I have restored it as may contribute to. Check every instance — polishing into English routinely strengthens Malay hedges, because Malay hedging is often carried by particles that have no direct English equivalent and simply disappear in translation.”
And the refusal: “Sentence 4 is not unclear in its grammar. It is unclear because you have not decided whether adherence is your outcome or your mechanism. Both readings are grammatical. Decide, and the sentence will write itself.”
That last one is why this tab exists. The most dangerous language editing is the kind that makes unclear thinking read fluently.
“Never remove a hedge”
Models default to confident prose because confident prose reads better. In a methods or results section that default is a factual change, not a stylistic one.
BM→EN drift
A specific, checkable risk for this cohort: Malay hedging particles often have no English equivalent and vanish in translation, quietly upgrading a cautious claim into a definite one.
Unclear thinking clause
Prevents the most common harm from language tools: fluent prose that conceals an undecided argument, which then survives all the way to peer review.
Polishing must change how it reads, never what it claims.
Modul 2 (2) · deliverable
Assemble the outline while the parts are still fresh
Every row already exists somewhere in what you built in Session A and in this block. This slide is where you put them in one place and see what is missing.
Title — the claim, not the topic
Problem statement — from Lab B1.1
Literature basis — from the Session A matrix
Conceptual framework — from Lab B2.0
Methodology — from Lab B2.0b
Expected outcomes & who acts on them
The row that is usually empty
The last one. Most proposals describe what will be produced and never say whose decision it informs. Funders read that row first, because it is the only one that tells them what the money buys.
Before you submit anything
Every citation in this outline must have entered through your reference manager, and every framework arrow must have a warrant that is not [WARRANT NEEDED]. Those two checks take an afternoon and remove almost every avoidable rejection.
Modul 2 (2) · authorship
The line is not where most people think it is
It is not about how much text the machine touched. It is about who made the claims and who can defend them.
You remain the author when AI…
• Critiques an argument you constructed • Reorganises material you selected, in an order you approve • Improves the clarity of sentences whose claims you already own • Generates alternatives you then choose between on stated grounds
In every case the intellectual decisions — what to claim, at what strength, on what evidence — were made by you and remain defensible by you.
You have surrendered it when AI…
• Decides what your findings mean • Selects which evidence is worth including • Produces prose containing claims you have not independently verified
The diagnostic question is brutally simple and worth asking of every paragraph: if a reviewer challenged this sentence, could I defend it from the primary sources, without going back to the chat? If not, it is not yours yet — regardless of who typed it.
Every claim in my draft traces to a source I have opened and read
I can state, for each paragraph, what decision I made and why
No sentence survives that I would not be able to defend in a viva
My AI use is recorded well enough that I could write the disclosure from memory
Why the viva question is the right testExaminers and reviewers do not ask whether you used a tool. They ask you to explain a choice on page eleven. The candidate who cannot is in difficulty whether or not any AI was involved — which is precisely why this test is the durable one.
Session B · block 2 of 3
Modul 3 — data analysis & research productivity
The proposal block ended with a method. This block runs it: plan the analysis before you can see the answer, clean with a receipt, and chart honestly.
What we cover
Pre-specified analysis plans, cleaning logs, Advanced Data Analysis and workflow automation, and the difference between a defensible figure and a misleading one.
You will build
An analysis plan, a cleaning log a colleague could rerun from, a cleaned dataset, and two publication-ready charts with honest captions.
The habit that carries
Ask the tool to report what it changed, not just what it produced. Code that drops 23 rows looks identical to code that drops none.
CLO4 · CLO5
Modul 3 · data
Decide the analysis before you can see the answer
AI makes it trivially easy to run twenty analyses and report the one that worked. The defence is not willpower — it is a plan written down beforehand.
PlanWritten before data
→
CleanLogged, every rule
→
AnalyseAs pre-specified
→
ExploreLabelled as such
→
ReportBoth, honestly
Lab B2.1 · analysis plan generatorDo this nowUse your own study, or the pharmacist-review example from B1.
ROLE: Biostatistician reviewing a protocol before data collection begins.
TASK: Draft a pre-specified analysis plan from the details below.
MUST INCLUDE:
- Primary outcome and the ONE primary test, named precisely
- Secondary outcomes, explicitly marked as secondary
- How missing data will be handled, decided NOW not later
- Which assumptions the primary test requires, and what I will do if each fails
- Any subgroup analysis, declared in advance as confirmatory or exploratory
MUST ALSO DO:
- Tell me which of my stated outcomes is underpowered given my n
- Flag anything I have listed that is really a post-hoc analysis in disguise
- Do NOT suggest additional analyses to make results more likely to be significant
MY STUDY:
Design: [ ]
Sample size and how it was determined: [ ]
Primary outcome: [ ]
Other outcomes I want: [ ]
The last instruction is not optional
Expected output
Expect three sections, and pay attention to the third:
1. Pre-specified. Primary outcome, one named test, assumption checks with a stated fallback for each, and a missing-data rule decided in advance.
2. Declared exploratory. Subgroups you listed that the study cannot confirm, relabelled honestly. These stay in the paper — they are just labelled for what they are.
3. Warnings. This is the part worth the exercise: — “Your third outcome is powered at roughly 30% for the effect size you describe. A null result there will be uninformative, and reporting it without this caveat would be misleading.” — “The subgroup you list as planned was not in your original hypothesis. Declare it exploratory or drop it.”
A model asked to be a helpful assistant will not volunteer either of those sentences. A model told explicitly not to help you find significance will.
One primary test
Forcing a single named primary test at plan stage is what makes the multiplicity problem visible. Two primary outcomes is usually one outcome and one wish.
Missing data, now
Deciding after you have seen which cases are missing is where a great deal of unintentional bias enters. The rule costs nothing to write in advance.
The negative instruction
“Do not suggest analyses to make results significant” blocks the model’s default helpfulness, which in a statistics context is indistinguishable from coaching you toward p-hacking.
The plan is cheap now and impossible to write honestly later.
Modul 3 · cleaning
Every cleaning decision is an analytic decision
If you cannot reproduce your cleaning, you cannot reproduce your result — and neither can anyone else. Fill this in as you work, not afterwards.
Decision
Rule applied
Rows affected
Why, in one sentence
Every exclusion rule was written down before I saw its effect on the result
I can state the n at every stage, from raw file to analysed sample
Any AI-generated cleaning code was read line by line and tested on a known subset
A colleague could rerun my cleaning from this log alone
Using AI for cleaning code, safelyAsk for the code and a printed summary of what it changed — rows dropped, values recoded, counts before and after. Code that silently drops 40 rows looks identical to code that drops none. The summary is the check, and it takes one extra sentence in the prompt.
Modul 3
Upload, clean, chart — and check what it dropped
Advanced Data Analysis writes and runs code on your file. That is genuinely powerful and genuinely dangerous, and the difference is entirely in what you ask it to report back.
Before you upload anythingEverything in the privacy drill later this session applies here first. Identifiable data, unpublished collaborator data and interview transcripts do not go into a consumer account. If your file has a name column, remove it before upload — not after.
Lab B3.0 · cleaning and chartsDeliverable × 3Use a real file if you have one; a small extract is enough. Otherwise follow the shape of the output.
ROLE: Data analyst preparing a dataset for a peer-reviewed submission.
TASK: Clean the attached file and produce an auditable record of every change.
RETURN, IN THIS ORDER:
1. A PROFILE FIRST — rows, columns, type of each, count and pattern of missing
values, and anything that looks wrong. Do NOT change anything yet.
2. A PROPOSED CLEANING PLAN — one line per rule, with the number of rows each
rule would affect. Wait for my approval before running it.
3. After I approve: the cleaned file, plus a CHANGE LOG table —
rule | rows affected | before n | after n
4. Any row you could not classify, listed individually rather than dropped
HARD RULES:
- Never drop a row silently. If a rule removes rows, report the count in the log.
- Do not impute missing values unless I ask. Report the missingness pattern instead.
- Do not recode a category because it looks like a typo. List suspected typos and let me decide.
- Show me the code you ran, not just the result.
MY FILE: [ attach ]
WHAT EACH COLUMN MEANS: [ ]
Profile first, change nothing
Expected output
The change log is the deliverable:
| Trim whitespace in site | 412 | 412 | 412 | | Parse date_dispensed as date | 412 | 412 | 412 — 7 failed to parse, listed below, not dropped | | Drop rows with no outcome value | 23 | 412 | 389 |
And the part that matters: “Suspected typos in site: ‘HKL’ (n=190), ‘H.K.L’ (n=8), ‘hkl’ (n=3). These are probably one site, but merging them is a decision about your data, not a formatting fix. Confirm before I merge.”
“Missingness is not random: 19 of the 23 rows with no outcome come from one site, collected in the same fortnight. Dropping them will bias your comparison. Consider reporting this rather than deleting it.”
That second paragraph is the entire value of the exercise. Unconstrained, the same tool cleans your file in one step, returns a tidy table, and never mentions that the missing rows clustered in one site — because you did not ask, and the output looked complete either way.
Profile before change
Splitting inspection from action is the single most useful habit with a code-running tool. It also gives you a before-state you can quote in your methods.
“Never drop silently”
Code that drops 23 rows looks identical to code that drops none. The count is the only thing standing between you and an unreported exclusion.
“List typos, do not fix”
Merging categories is an analytic decision wearing the costume of a formatting fix. Keep the decision.
If you cannot reproduce the cleaning, you cannot reproduce the result.
Ask for the figure your data deserves, not the one that looks best.
ROLE: Scientific figure designer who follows journal reporting standards.
TASK: From the cleaned dataset, produce TWO figures.
FIGURE 1 — the primary comparison.
FIGURE 2 — the distribution behind that comparison.
REQUIREMENTS FOR BOTH:
- Show the data, not only the summary: individual points, or a box/violin,
overlaid on any summary statistic
- Report n on the figure itself, per group
- Axis labels with units. No abbreviations that are not defined in the caption.
- A perceptually uniform, colourblind-safe palette. Never a rainbow scale.
- Do not truncate a y-axis on any bar geometry. If truncation aids readability,
switch to points with intervals instead and tell me you did.
- Export at 300 dpi, and give me the code
THEN:
- Write the caption for each: what it shows, n, what the error bars represent,
and what it does NOT establish
- Tell me which of the two figures is more likely to mislead a fast reader, and why
CLEANED DATA: [ from Tab 1 ]
MY PRIMARY OUTCOME AND GROUPING: [ ]
Show the data, not only the summary
Expected output
The caption you want: “Figure 1. Refill-based adherence at 12 months by review group. Points are individual patients; bars show group means with 95% confidence intervals. Intervention n=194, control n=195. This figure does not establish a causal effect: allocation was not randomised and baseline polypharmacy differed between groups.”
And the honest warning: “Figure 1 is the more likely to mislead. The confidence intervals overlap substantially, but the mean markers sit far apart at this axis scale, and a fast reader reads the gap between markers rather than the overlap between intervals. Figure 2 shows why: both distributions are bimodal, and the difference in means is driven by the size of the lower mode rather than by a shift in the whole group. Consider making Figure 2 the primary figure.”
That recommendation — demote your headline figure — is one you would rarely give yourself. It is also exactly what a good co-author would say, and it is far cheaper to hear now than in a reviewer’s report.
“Show the data”
The single strongest improvement most figures could make. If it makes your result look weaker, the result was weaker and you have learned it before a reviewer did.
“What it does NOT establish”
A caption line that pre-empts the most common overreading of your own figure. Reviewers notice its absence.
“Which one misleads”
Turns the tool from a chart generator into a reviewer. This is the same refusal-permission technique as every other lab in these two days.
A figure that shows its own weakness is stronger than one that hides it.
Weissgerber T.L. et al. (2015). Beyond bar and line graphs: time for a new data presentation paradigm. PLOS Biology, 13(4), e1002128. https://doi.org/10.1371/journal.pbio.1002128
None of these is fraud. Each one has appeared in a journal this year. Mark each, then read why two of them split expert opinion.
A bar chart of means with the y-axis starting at 40 rather than zero, with the axis break clearly marked.
Misleading — for bars. A bar encodes magnitude by its area, so truncating the axis distorts the comparison no matter how clearly the break is labelled. The honest fix is to change the geometry, not the annotation: use points with confidence intervals, where a truncated axis is entirely acceptable.
A line chart of the same data with a truncated y-axis, showing change over twelve time points.
Defensible. A line encodes change, not magnitude, so a zero baseline often hides the signal entirely. This is the exception that proves the rule: the question is never “should the axis start at zero” but “what does this mark encode”.
Plotting individual data points over the bars, revealing that one group has three extreme values driving the mean.
Defensible, and better than defensible. Showing the distribution is the strongest single improvement most figures could make. If it makes your result look weaker, the result was weaker and you have just learned something before a reviewer did.
Choosing the colour scale after seeing which one makes the treatment effect look clearest.
Misleading. This is analytic flexibility relocated into the figure. Choose a perceptually uniform scale on principle before plotting — and never a rainbow scale, which manufactures apparent boundaries where the data has none.
An AI-generated figure that renders beautifully, from a script you have not read.
Never. Plotting code silently drops non-finite values, reorders factor levels alphabetically, and applies default aggregations. A beautiful figure of the wrong subset is the most convincing error you can publish, and nothing about its appearance will warn you.
Reporting the exact p-value as 0.061 and describing the result as showing no significant difference.
Defensible — and notice how uncomfortable that is. Reporting the exact value is correct practice. The description is conventional, though “no evidence of a difference” would be more accurate than “no difference”. What would be misleading is calling 0.061 a trend toward significance, or rounding it to 0.06 and hoping. The honest failure mode here is under-reporting, not over-claiming.
Rows 1 and 2 are the same dataThe identical axis choice is misleading in one geometry and necessary in the other. This is why “always start at zero” is bad advice: the rule is about what the mark encodes, and a rule you apply without understanding will eventually be applied wrongly.
Session B · block 3 of 3
Closing — integrity, ethics & direction
Everything built over these two sessions has to survive contact with a reviewer, an ethics committee and a student. This block is what makes it publishable.
What we cover
Where each kind of AI use is declared, what may and may not be pasted into a tool, and how to redesign assessment rather than police it.
You will build
An AI-use statement you could paste into a submission today, and a 30-day implementation plan with dates attached.
The one sentence
Never put a claim into the world that you have not verified at its source. Everything else is machinery for keeping that habit.
CLO5
Closing · integrity
What the standards now require
The most recent ICMJE revision created a standalone section on AI — the first time it has done so. The rules are less onerous than most people fear and more specific than most people realise.
Authorship is closed
AI tools cannot be authors, and cannot be cited as authors, because they cannot take responsibility for accuracy, integrity and originality — and responsibility is what authorship means. Every publisher already held this position; ICMJE removed the remaining ambiguity.
Placement is specified
Writing assistance is declared in the acknowledgements. Use in data collection, analysis or figure generation is declared in the methods. These are different sections because they are different kinds of claim — one about prose, one about results.
Responsibility does not transfer
Humans are accountable for everything submitted, including text and images the tool produced. “The AI generated it” is not a defence — it is a description of how the error entered, and it makes the disclosure question worse rather than better.
If you used AI to…
Declare it in
Say what, specifically
Improve grammar or clarity of your own text
Acknowledgements
Tool, version, and that content was author-generated
Draft or restructure prose
Acknowledgements
Which sections, and that you verified every claim
Screen, extract, classify or code data
Methods
Stage, tool, and how human oversight worked
Write analysis code or generate figures
Methods
What was generated and how it was validated
Translate your own text
Acknowledgements
That meaning was verified by a competent speaker
I know which of my uses go in methods and which in acknowledgements
I keep a running note of tool and purpose while I work, not afterwards
I have checked my target journal’s own policy, which may be stricter than ICMJE
I would be comfortable if my disclosure statement were read aloud at my viva
What changed in the latest revision
ICMJE consolidated its scattered AI guidance into a dedicated Section V, alongside tightened data-access requirements for industry-funded studies and stricter trial-registration checks. If your research group formed its habits before 2024, this is the moment to audit them — the requirements were less formal when most supervisors learned them.
Assume a consumer AI account with default settings. Mark each, then read the two that split the room.
A de-identified dataset, with the codebook pasted separately in the same conversation.
Never. The separation is illusory — both sit in one context window, so you have simply re-identified the dataset in a place you do not control. If the codebook is needed, the data should not be there.
A manuscript you are peer-reviewing, to help you write a clearer review.
Never. Confidential unpublished work belonging to someone else. Reviewers are explicitly instructed not to upload manuscripts to systems that cannot guarantee confidentiality — and this is a breach of the author’s trust regardless of any policy.
Your own published paper, asking for a plain-language summary for a patient leaflet.
Safe. Published, yours, and no confidential content. Verify the simplification did not distort a clinical claim — plain-language rewriting reliably strengthens hedged statements, which matters more in a patient leaflet than anywhere else.
A grant proposal draft, before submission, in a consumer AI account.
Never on a consumer account. Unpublished intellectual property, often with collaborators’ unpublished ideas and budget figures inside it. An institutional or enterprise account with a no-training agreement changes this answer — know which kind of account you are using before you paste.
A single anonymised clinical vignette, already published as a teaching case with consent.
Safe — and worth pausing on. Published with consent for teaching, so the confidentiality question is settled. But note that a rare-disease vignette with a specific presentation can be identifying even when every name is removed. “Anonymised” is a property of the whole description, not of the name field.
Your own interview transcripts from a qualitative study, with names replaced by codes.
Never, unless your ethics approval covers it. This is the row that surprises people. Participants consented to a stated set of uses; processing on a third-party system was almost certainly not among them. Qualitative transcripts are also densely re-identifying — a participant describing their workplace and role is often identifiable to a colleague in three sentences.
The question that resolves most of theseNot “is this anonymous?” but “did the person it belongs to agree to this use?” Rows 2, 4 and 6 all fail that test while passing a naive anonymity check — and it is the consent question, not the identifiability question, that an ethics committee will ask you about.
Closing · deliverable
Write the statement you will actually submit
Four fields. Fill them for a piece of work you have in progress right now, then run the lab to turn them into two sentences.
Tool and version used
What it did, in a verb
How you verified the output
Where this goes in the paper
Lab B4.1 · statement generatorDeliverablePaste your four fields. Keep the output; it is the deliverable.
ROLE: Research integrity officer reviewing a submission.
TASK: Write my AI-use disclosure in TWO sentences, and tell me where it belongs.
RULES:
- Name the tool and what it did, in plain declarative language
- State the verification step explicitly — this is the sentence that carries the weight
- No defensive hedging, no apology, no "merely" or "only"
- Assign it to acknowledgements (writing assistance) or methods (data, analysis, figures) per ICMJE
- If my described use would NOT satisfy a reviewer, say so plainly instead of writing the statement
MY USE:
Tool and version: [ ]
What it did: [ ]
How I verified: [ ]
Keep this output
Expected output
Acknowledgements version: “Claude Opus 4.5 was used to improve the clarity and structure of the discussion section. All claims, citations and interpretations are the authors’ own, and every reference was verified against the primary source.”
Methods version: “Title and abstract screening was assisted by Rayyan’s relevance ranking. All include and exclude decisions were made independently by two human reviewers, with disagreements resolved by a third.”
And the refusal you might get: “You have described the tool as having drafted the results interpretation, and your verification step is that you read it and it seemed right. That will not satisfy a reviewer, because reading for plausibility is what the tool is optimised to pass. Either verify each claim against your data and rewrite this, or the interpretation is not yours to submit.”
Twenty-nine words of unwelcome truth, delivered before submission rather than after. That is the whole value of the last rule in the prompt.
Two sentences
Length is a tell. A long disclosure reads as anxiety; a short declarative one reads as command of your own process.
No hedging
“I only used it a little” invites the question it is trying to avoid. State what happened.
The refusal clause
The same technique as Lab B1.1: giving the model permission to say no converts it from a writing tool into a check on your process.
If the statement is hard to write, the practice needs changing — not the wording.
Checkpoint B2 · immediate feedback
Four judgements before we close
These are the situations you will meet, not the ones that make a clean example.
Q1A co-author used AI heavily and did not tell you. You find out at proof stage. What is the first thing you do?
Add a disclosure line and proceed
Necessary, but not first. A disclosure that describes an unverified process discloses the wrong thing accurately.
Verify the claims and references in the affected sections before deciding anything else
Correct. Your name is on it and you are accountable for its accuracy regardless of who produced the text. Disclosure follows verification; it cannot substitute for it.
Remove your name from the paper
Possibly the right ending, but not the right beginning. You do not yet know whether anything is actually wrong.
Assessed because accountability is joint and does not depend on who typed. CLO5.
Q2Which disclosure would trouble a reviewer most?
“AI was used to draft the discussion; all claims were verified against primary sources.”
This is a clear, ordinary disclosure. Extensive use plus stated verification is not a problem.
“AI was used minimally for language polishing only.”
Correct, and this surprises people. “Minimally” and “only” are unverifiable defensive words that describe nothing. A reviewer reads them as either carelessness or concealment, and neither helps you.
“AI was used for abstract screening; two humans made all inclusion decisions.”
Exactly the right form: stage, tool role, and where human authority sat.
Assessed because vague disclosure is more damaging than extensive disclosed use. CLO5.
Q3A student submits work you are confident is AI-generated. Your evidence is a detector score and your impression of the writing. What is defensible?
Refer to a misconduct panel
Not on this evidence. Both inputs are unreliable, and the writing-impression judgement is known to misfire on second-language writers.
Ask the student to talk you through three specific choices in the work
Correct. This is fair, it is diagnostic, and it tests the outcome you claimed to assess. If they can defend the work, the question is answered; if they cannot, you now have something real.
Mark it down for style
Punishing a suspicion through a criterion you did not declare is less defensible than an open conversation, and it is invisible to appeal.
Assessed because this decision is made under pressure, and the wrong one harms a named student. CLO5.
Q4Across both sessions, which single habit prevents the most failures?
Using only grounded tools
Helps a great deal with fabricated references, and does nothing about misattributed claims, overstated findings, or privacy.
Never generating a claim you have not verified at its source
Correct. It is one habit, and it defeats fabrication, misattribution, stale guidance and confident wrongness at once — because all four are the same failure wearing different clothes.
Disclosing everything
Necessary and insufficient. An honest account of an unverified process is still an unverified process.
Assessed because it is the sentence worth leaving with. CLO1–CLO5.
Closing · transfer
Four commitments, each with a date
Not aspirations. Things small enough that you will actually do them, attached to a day.
This week — verify
This fortnight — build
This month — teach or write
This month — institutionalise
If you do only one
The first. It takes an afternoon, it costs nothing, and it is the one that has prevented an actual correction notice for someone in every cohort that has run this workshop.
If you supervise
Make the fourth one first. A group norm agreed in a twenty-minute meeting prevents more problems than any amount of individual good practice, because it makes disclosure ordinary rather than confessional.
End of half-day 2
Your Session B scorecard
Same rule as Session A: the lowest rows are the instruction, not a judgement.
Below 40%
Return to Lab B1.1 and run both levels on a real draft. The adversarial pass alone moves most people a full band, and it takes twelve minutes.
40–64%
Open the ledger and find the judge rows. Those two drills are where the durable skill sits, and they can be redone in ten minutes each.
Above 65%
Teach one piece of this to someone else within a fortnight. Explaining the grounded/ungrounded distinction to a colleague is the fastest way to find out whether you have it.
What the score is notIt is not a measure of how much you know. It is a record of what you built while you were here. Someone who read everything carefully and clicked nothing has a low score and may have learned a great deal — but they have no artefact, and the artefact is what survives the week.
Both sessions
Everything, in one view
Press S at any time to reopen this ledger, in either language.
Session B · resources
Sources, standards and where to go next
Standards first — those are the ones with consequences attached.
Simmons J.P. et al. (2011). False-positive psychology: undisclosed flexibility in data collection and analysis. Psychological Science, 22(11), 1359–1366. https://doi.org/10.1177/0956797611417632
Weissgerber T.L. et al. (2015). Beyond bar and line graphs: time for a new data presentation paradigm. PLOS Biology, 13(4), e1002128. https://doi.org/10.1371/journal.pbio.1002128
Perkins M. et al. (2024). The AI Assessment Scale: a framework for AI use in education. Journal of University Teaching & Learning Practice, 21(6). https://doi.org/10.53761/q3azde36
Weber-Wulff D. et al. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26. https://doi.org/10.1007/s40979-023-00146-z
Walters W.H. & Wilder E.I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5
Two papers to hand your committeeLiang and Weber-Wulff together are the evidence base for why detector scores cannot support a misconduct finding. If your institution is drafting policy this year, these are the citations that will do the most work in the room.
Close
One sentence to leave with
Never put a claim into the world that you have not verified at its source. Everything else in these two sessions — grounding, PICO, PRISMA, the matrix, the rubric, the disclosure — is machinery built to make that one habit easier to keep.
The deck is the handout
This file works offline, in both languages, in light and dark. Your score does not persist between openings — but the prompts, the labs and every DOI do.
Working through it alone
Press O for the overview and jump to any module. The labs stand on their own; the two judge drills are the ones worth redoing cold, weeks later.
If you teach this
Every slide carries facilitator notes — press N. The assessment map on slide five explains what each component is evidencing and why, which is the part worth adapting rather than copying.