← series index
# Demo Script — Session 4: Writing, Publishing & Integrity
*AI for Researchers — presenter script · Landscape as of August 2026*
**Runtime:** ~25 minutes (Setup ~2 min · Demo A, suggestion mode ~8 min · Demo A, rewrite mode ~7 min · Demo B, disclosure ~6 min · Close ~2 min)
**Slides it follows:** slide 29 ("CRIT, Applied to a Compliant Editing Pass")
**Artefacts it exercises:** The Publisher AI-Policy Comparison Table (slides 12–14), CRIT (Session 1), the disclosure templates in `handout.md`
---
## 0. What this demo is, and what it deliberately is not
Two demos, one paragraph:
- **Demo A** runs the same rough methods paragraph through **two different prompts** — *suggest edits* versus *rewrite this* — and diffs both against the original. The point is not that one is good and one is bad. The point is that they land in **different places on the editor–ghostwriter spectrum**, and therefore under **different disclosure rules**.
- **Demo B** writes the AI-use disclosure statement that Demo A's edit actually requires, in one real publisher's required format, quoting that publisher's own requirement on screen.
**A rule this demo obeys, and you should say so out loud:** the paragraph below was written for this webinar. It is **not** any real author's text. Never paste a colleague's manuscript, a manuscript you are reviewing, or an unpublished draft that is not yours into a public chatbot to make a teaching point — Session 4's own peer-review slides (27–28) forbid exactly that.
> **Say this before you start:** *"I wrote this paragraph badly on purpose. Everything the tool does to it, you can check against the original — which is the only reason a demo like this is worth watching."*
---
## 1. Preparation (the day before)
| # | Item | Why |
|---|---|---|
| 1 | Put the sample paragraph (§2) in a plain-text file you can paste from. | Typing it live is 90 seconds of dead air. |
| 2 | Run both prompts once and **screenshot every output**. | Primary fallback (§6). |
| 3 | Open your target publisher's AI policy page in a tab — Elsevier's is the one this script quotes. | Demo B shows the requirement *before* it writes to it. |
| 4 | Open a text-diff view (any editor with compare, or two windows side by side). | The diff is the demo. Without it you are just reading two paragraphs aloud. |
| 5 | Decide which tool you will use, and be ready to say why. | Tool neutrality note below. |
| 6 | Have `sources.md` open on a second screen. | Someone will ask where 61.3% comes from — and it is a 2023 study, so be ready to say so and to pair it with the August-2026 re-test `[30]`. |
**Tool neutrality note.** This script does not name a chatbot, because every general assistant (Tier 1 in Session 1's taxonomy) and every writing-specific tool will run both prompts. Say on the day: *"I am using this one because I have an account. The prompts below are the transferable part — they work anywhere, and the disclosure requirements apply to whatever you used."* If your institution provides a privacy-reviewed instance, use that and say why.
---
## 2. The sample paragraph (write it on the slide, or hand it out)
This is a deliberately rough ~150-word methods paragraph, written for this session. Read it aloud — badly written prose sounds worse than it looks, which is the point.
```
We did a survey for to know how researchers in our university is using AI
tools in their writing. The questionnaire was distributed by email in the
month of March 2025 to all academic staff (n = 412) and it was open for
three weeks. Reminder was sent one time after two weeks. The questionnaire
had 18 questions, 14 of them closed and 4 open-ended, and it was adapted
from a previous instrument which we modified. 137 people responded, this is
a response rate of 33.3%, but 9 responses were incomplete so we excluded
them from analysis. Data was analysed with descriptive statistics, and for
the open-ended questions two coders coded them independently and the
agreement was fair (κ = 0.71). Ethical approval was got from the
institutional review board (ref 2025-041). Because the sample is from one
institution the results maybe are not generalisable to other contexts.
```
**What is buried in it, for you to watch for (do not tell the audience yet):**
| Item | Why it matters |
|---|---|
| `n = 412`, `137`, `33.3%`, `9 excluded`, `κ = 0.71`, `ref 2025-041` | Six facts. A copy-edit must leave every one untouched. |
| "the agreement was fair (κ = 0.71)" | κ = 0.71 is conventionally described as *substantial*, not *fair*. **If the tool "corrects" this, it has made an interpretive claim you did not make.** This is the trap. |
| "the results maybe are not generalisable" | A hedge. Watch whether the rewrite strengthens, weakens, or deletes it. |
| "adapted from a previous instrument which we modified" | Redundant *and* missing a citation. A good editor flags it; a ghostwriter invents the citation. |
---
## 3. Demo A, part 1 — suggestion mode (~8 min)
### Step A1 — The prompt (paste verbatim)
```
Context: This is a methods paragraph I wrote myself, for a survey study I
ran. I am submitting to a journal that permits AI-assisted language editing
provided I disclose it.
Role and register: Act as a copy-editor for an academic methods section in
a social-science journal. You are not a co-author and not a reviewer.
Instructions and constraints:
- Return a NUMBERED LIST OF SUGGESTIONS. Do not return a rewritten
paragraph.
- For each suggestion give: the exact original wording, your proposed
wording, and a one-line reason.
- Do not change any number, statistic, symbol, reference or identifier.
- Do not change any word that expresses certainty, hedging, or an
interpretation of a result.
- Do not add any content, citation, or claim that is not already present.
- If something is unclear or looks factually wrong, say so as a QUESTION
to me. Do not fix it.
Task: List the changes you would make to the paragraph below, in that
format.
[paste the paragraph]
```
**Expected outcome.** A numbered list, typically 8–14 items. Reliably present:
1. "for to know" → "to determine" / "to assess"
2. "researchers … is using" → "are using" (subject–verb agreement)
3. "Reminder was sent one time" → "One reminder was sent"
4. "137 people responded, this is a response rate of 33.3%" → comma splice fixed
5. "Data was analysed" → "Data were analysed"
6. "Ethical approval was got" → "was obtained"
7. "the results maybe are not generalisable" → "the results may not be generalisable"
8. Flagged as a **question**, if the constraints held — and current-generation assistants usually do hold them: *"'fair' is an unusual descriptor for κ = 0.71 — did you mean substantial?"* and *"which previous instrument? A citation appears to be missing."* If instead it silently "fixed" either one, that is your Step A2 talking point handed to you early.
### Step A2 — What to point at on screen
- **Every suggestion is separable.** You accept #2 and reject #7 if you want the hedge as it stands. That is what makes this *editing*: the judgement stays with you.
- **The two questions are the tell.** A tool that *asked* about κ = 0.71 obeyed the constraint. A tool that silently changed "fair" to "substantial" wrote an interpretation into your methods section — point at it and say so.
- **Check the six numbers, out loud, one at a time.** This takes 20 seconds and is the single most transferable habit in the session.
### Step A3 — Accept the edits yourself
Apply the accepted suggestions **by hand** in your document, in front of the room.
**Say:** *"This is the slow part, and it is the part that keeps me the author. I read every change. Session 3's lesson applies here too: the tool proposes, I dispose."*
---
## 4. Demo A, part 2 — rewrite mode, and the diff (~7 min)
### Step A4 — The contrasting prompt (paste verbatim, on the ORIGINAL paragraph)
```
Rewrite the following methods paragraph so it reads clearly and
professionally for a social-science journal.
[paste the paragraph]
```
That is the whole prompt. It is what most people actually type.
**Expected outcome — dual-path.** One fluent paragraph, noticeably better English, either way. Which path you get, you cannot know in advance — and that is the lesson, so rehearse both.
**Path 1 — the rewrite behaves (on current-generation models, the more likely outcome).** All six facts intact, the hedge preserved, "fair (κ = 0.71)" left alone. **This is the primary framing, not a letdown:** *"This time it kept everything. Run it three times and you will get three answers — which is exactly why 'it was fine last time' is not a compliance strategy."* Then show your pre-captured run where it *did* drift (prep item 2) and make the variance the point: the diff is how you find out which run you got.
**Path 2 — the rewrite drifts.** Watch for the classic moves (your rehearsal screenshots should show at least one of them):
- "fair" silently upgraded to "substantial" or "good" agreement;
- the hedge "maybe are not generalisable" hardened ("these findings are specific to a single institution") or dropped;
- the redundant clause about the instrument smoothed away — sometimes with an invented attribution;
- occasionally, a recomputed or "tidied" response rate (33% / 33.5%), or the 9 exclusions quietly folded into the total.
**Either path lands the same line:** the diff is not there to catch a bad model. It is there because you cannot know in advance which model showed up — and because the disclosure duty is identical on both paths.
### Step A5 — Diff both versions against the original
Put original, suggestion-mode result and rewrite-mode result side by side.
**The line to land:**
> *"Both outputs used AI. Only one of them is still only my content. Under Sage's own vocabulary, the left-hand one is **assistive** and the right-hand one is heading for **generative** — and those two words carry different disclosure duties."* `[21]`
### Step A6 — The evidence slide, called back
Three numbers, said once:
- Published, peer-reviewed research letters polished for readability only (a 2024-era pipeline: chatgpt-4o-latest, scored by GPTZero) went from **97–100% classified human** to **75–85% flagged AI-generated**. `[10]`
- A review of ten of those polished letters found **meaning altered in 2–3 sentences per letter**. `[10]`
- And it is not a solved 2024 problem: an August-2026 study found guideline-compliant light editing flagged **38–80%**, while humanizer-evaded AI text was flagged **under 4%**. `[30]`
**Say:** *"The middle number is why we just did a diff. The other two are why nobody in this room should treat a detector score as evidence about anybody — including themselves."* `[8]` `[9]` `[30]`
---
## 5. Demo B — write the disclosure statement (~6 min)
### Step B1 — Show the requirement BEFORE writing to it
Open Elsevier's *Generative AI policies for journals* page live and read the requirement off the publisher's own screen `[17]`:
> Authors must add a new section at the end of the manuscript, before the references list, titled **"Declaration of generative AI and AI-assisted technologies in the manuscript preparation process"**, containing:
>
> *"During the preparation of this work, the author(s) used [NAME OF TOOL / SERVICE] in order to [REASON]. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the published article."*
>
> And the exemption, quoted exactly: *"Basic checks of grammar, spelling and punctuation do not need a declaration statement. However, when an AI tool makes substantive changes to sentence structure or organization of a part of the text, this should be disclosed."*
**Ask the room:** *"Does the edit we just made need a declaration?"* Let them argue for 30 seconds. The answer for suggestion mode is **yes** — we changed sentence structure ("Reminder was sent one time" → "One reminder was sent"; the comma splice), which is past the "basic checks" exemption. For rewrite mode it is unambiguously yes.
### Step B2 — The prompt (paste verbatim)
```
Context: I used an AI assistant to copy-edit one methods paragraph of my
manuscript. It suggested language changes; I reviewed and accepted them
individually. No content, numbers, citations or interpretations were
generated by the tool.
Role and register: Act as a scrupulous author filling in a publisher's
required disclosure form.
Instructions and constraints:
- Use EXACTLY this template and change nothing outside the brackets:
"During the preparation of this work, the author(s) used [NAME OF TOOL /
SERVICE] in order to [REASON]. After using this tool/service, the
author(s) reviewed and edited the content as needed and take(s) full
responsibility for the content of the published article."
- Fill [NAME OF TOOL / SERVICE] with: <tool name and version you actually
used>.
- Fill [REASON] with a factual description of what it did, in under 20
words. Do not overstate or understate.
- Do not add sentences, hedges, apologies or justifications.
- Output the section heading and the statement, nothing else.
Task: Produce the declaration.
```
**Expected outcome** (your tool name and version substituted):
> **Declaration of generative AI and AI-assisted technologies in the manuscript preparation process**
>
> During the preparation of this work, the author used [tool name, version] in order to copy-edit the language of the methods section. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the content of the published article.
### Step B3 — Check it against three things, live
1. **The template.** Word for word outside the brackets. If the tool "improved" the template sentence, that is a failure — show it and fix it by hand. `[17]`
2. **The truth.** [REASON] must describe what actually happened. "Copy-edit the language of the methods section" is true. "Improve the manuscript" is vague; "generate the methods section" would be false.
3. **The journal's own guide for authors.** Slide 16's lesson: on one day, two Elsevier journals asked for two different headings for this same declaration. `[26]` Say: *"The publisher page gives you the shape. The journal page gives you the wording. Check it the week you submit."*
### Step B4 — The 30-second variant round (optional if time allows)
Same edit, three other publishers, straight off the comparison table:
- **Springer Nature:** copy-editing for readability, style and grammar **does not need declaring at all** — "The use of an LLM (or other AI-tool) for 'AI assisted copy editing' purposes does not need to be declared." `[18]`
- **JAMA Network:** Acknowledgment section, and it must name the tool, "version and extension numbers, and manufacturer", plus dates of use. `[24]` `[25]`
- **ICMJE-following journals:** in **both** the cover letter **and** the acknowledgments. `[14]`
**Say:** *"Same edit. Four different obligations, one of which is 'do nothing'. That is why the table exists — and why the answer to 'do I need to disclose?' is always 'where are you sending it?'"*
---
## 6. Fallbacks
| If this fails | Do this |
|---|---|
| The tool is down or rate-limited | Show the pre-captured screenshots of both prompts' outputs (prep item 2). The diff is the demo; it works perfectly from screenshots. |
| Both the live rewrite **and** all your rehearsal runs behaved | Say so honestly — behaving-on-average is real progress and worth naming. Then land the two things that did not change: run-to-run variance means the *next* run carries the same risk, and the disclosure duty attaches to the use, not to whether the output drifted. The diff still earned its keep — it is how you found out it behaved. |
| Suggestion mode returns a rewritten paragraph anyway (rarer on current-generation models, but rehearse it) | Perfect. Point at it: *"It ignored an explicit constraint. That is the failure mode you are trusting when you skip the diff."* Then re-prompt with "You returned a rewrite. Return a numbered list of suggestions instead." |
| The publisher page will not load | Read the requirement from the handout's Template 1, which quotes it verbatim, and say the page was fetched on 2026-07-28. |
| You are running long | Cut §5 Step B4 (the variant round) and the second half of §4 Step A6. Never cut the diff (§A5) or the disclosure itself (§B2). |
| Someone asks you to run a detector on the outputs live | Decline, and explain why in one sentence: the 2023 tests found the tools "neither accurate nor reliable" `[8]` with a 61.3% false-positive rate on human-written non-native English essays `[9]`, the August-2026 re-test still flags compliant light editing at 38–80% while humanizer-evaded text sails through `[30]`, and demonstrating one on screen teaches the room to trust it. |
---
## 7. Close (2 min)
Three sentences, then hand back to the slides:
1. *"Editing keeps your content and your accountability; ghostwriting moves both. Everything else in this session is downstream of that distinction."*
2. *"Disclosure is not a confession. It is the thing that converts a suspicion into a documented fact."*
3. *"Before you submit, open your target journal's guide for authors and read the AI paragraph. It takes ninety seconds and it is the only version of this session that is guaranteed still to be true."*
---
*AI for Researchers · Session 4: Writing, Publishing & Integrity · Landscape as of August 2026 · Citations `[n]` resolve in this session's `sources.md`.*