In the age of AI, how much of you needs to be in the writing for it to still be yours?

That’s the question I’ve been struggling with lately.

Not knowing who to ask, I asked ChatGPT.

Here’s what it said:

“The real ‘you’ in writing isn’t just about the words themselves—it’s about voice, perspective, and the unique way you connect ideas.”

Not bad.

But maybe the better question is: “What makes writing yours at all?”

For me, this isn’t just philosophical. It’s personal.

I’ve spent over a decade teaching clinical researchers how to write—manuscripts, grants, abstracts.

I built frameworks. Created systems. Treated academic writing as a craft.

But today? I wouldn’t teach it the same way.

Because writing has changed. Fundamentally.

The skills I built my career on are still useful—but no longer enough. The bar isn’t just being “good.” It’s being better than AI.

I’ve seen ChatGPT draft literature reviews (deep research) more coherently and faster (in minutes) , than a research assistant in a full week.

So what is authorship now?

Is it typing every word? Or is it having the idea – the structure, the spark. And letting AI help with the heavy lifting?

I don’t have a clean answer. But I know this:

The real challenge isn’t just writing well. It’s writing in a world where AI-generated content is the baseline.

And that’s where we need to start.

1) Authorship is no longer just about writing—it’s about owning the work

International Committee of Medical Journal Editors (ICMJE) still sets the bar:

• Substantial intellectual contribution
• Drafting or critical revision
• Final approval
• Accountability for the whole

AI can’t do any of these.

It can’t approve a manuscript.

It can’t be accountable.

It can’t sign a conflict-of-interest form or defend your study at the poster session in your next research conference.

So no, AI isn’t an author—even if it drafted three pages you approved. The ICMJE criteria explicitly reserve authorship for people who can take responsibility for the entire work.

Where AI fits here: as a tool inside the second criterion (drafting/revision), not a substitute for the first and fourth (contribution + accountability).

Example

You ask AI to outline a Discussion on treatment response in disease X. It suggests a bold subgroup claim (based on the data you provided it with) you didn’t consider. If you keep it, you still own the reasoning, the checks, and the consequences. That’s authorship in the AI era: you didn’t just type the words—you curated, verified, and stand behind them.

Bottom line

Using AI is not the problem.

Pretending it confers authorship on the tool (and not you) is.

2) Invisible help is still help—undisclosed AI crosses into ghostwriting territory

If a junior colleague wrote half your manuscript, you’d acknowledge or add them—depending on contribution.

Why should AI be different?

Undisclosed AI use is functionally unacknowledged ghostwriting—even when the facts are right and the references are real. Journals increasingly require transparency on whether, where, and how AI assisted the text.

What to disclose (be concrete):

• Which tool you used (e.g., ChatGPT, Sept 2025 version)
• Where you used it (intro draft, language polish, outline options)
• That humans verified all content and assume full responsibility

One-sentence template you can paste

“Generative AI was used to draft language and propose structure for the Background and Discussion; all authors verified sources, edited for accuracy, and accept responsibility for the final content.”

Example

You let AI write your limitations paragraph for your study. You keep it as-is. No disclosure. Reviewers later find the paragraph cites a guideline that doesn’t exist. That’s not an AI problem. That’s a transparency and accountability problem.

3) AI is becoming a quiet writing assistant—so your value shifts from prose to judgment

Elsevier’s 2024 survey: a third of researchers already use genAI for writing. The floor on fluency is rising.

In controlled tests, AI-written abstracts often read “cleaner” than human ones—and fooled blinded reviewers about a third of the time.

Where you win now:

• Choosing the right questions and comparators
• Weighing clinical vs statistical significance
• Seeing confounding no model flagged
• Proposing next steps that reflect bedside reality

Example

Two abstracts on resistant hypertension:

• AI-first draft nails the flow, misses nuance around diuretic optimization in CKD.
• You feed the AI your key findings, central message, and implications; it drafts; you re-frame outcome hierarchy and flag a safety signal in subgroups.

The second is authorship. The first is surface-level AI writing.

4) The ethics of speed—AI helps you move fast, but you must not break truth

AI saves time. It also creates new failure modes:

• Made-up references that look plausible
• Confident prose masking weak assumptions
• Overgeneralized claims with missing caveats

In a widely cited study, ChatGPT-generated abstracts passed plagiarism checks, fabricated details, and fooled human reviewers 32% of the time. Readability ≠ reliability.

Your countermeasures:

• Treat every number as untrusted until verified
• Replace AI-suggested citations with real ones you read end-to-end
• Add an “assumptions & uncertainties” box to each section the AI touched

Example

For a hypertension RCT, AI drafts a crisp abstract. The NNT it reports is calculated off a secondary endpoint. You recompute from the primary, add CIs, and rewrite the claim. Same speed advantage. Now credible.

5) Writing with AI demands new literacies—teach them, model them, require them

New literacy stack:

• Prompt design: Give context, constraints, structure.
• Critical editing: Turn fluent text into argument. Tighten claims, foreground limits, align with prior work using accurate citations.
• Verification: Hunt hallucinations; re-run math; check that every citation supports the specific sentence it tags.
• Disclosure: Say how AI contributed; make it obvious humans took responsibility.

Example: better prompt

“I am writing the outline of my discussion section for my paper on [details of your study]. Provide at least 3 studies that agree with my key finding and at least 3 that disagree. No review articles. No grey literature.” (works best on GPT-5 with thinking)

This yields a scaffold you can defend, not just random evidence you hope is right.

6) Mentorship must catch up—use AI as a starting point, not a substitute for thought

Many trainees now begin with, “Write me an introduction about X with recent references.”

That skips the hard part. It’s not writing—it’s outsourcing.

The real task of authorship is building the judgment AI can’t replace. And that shows up at 2 stages that remain fully human:

1. The front end—framing the question, mapping the literature, deciding what matters, and what to feed the AI (enough context about your study and your interpretation of the results).
2. The back end—editing for accuracy, clarity, and flow. Verifying numbers, strength of claims, checking that every citation really supports the claim it’s tied to.

These are the moments where depth of thought, not surface-level prose, determines quality.

Training practices that work now:

• Evidence maps: Have mentees summarize 5–10 core studies on a one-page map—key findings, disagreements, and gaps. Then compare their map to an AI-generated scaffold and ask: What did the model miss? What did it overstate?
• Critical reviews: Give an AI-written abstract and have them highlight every claim needing verification, then trace back to the primary source.
• Revision drills: Take an AI draft of a discussion section and make them cut, reframe, or rewrite until the argument is both defensible and aligned with the actual evidence.

Example

Teaching critical thinking around the beginning and the end of writing—the framing and the final edit—is how mentorship stays ahead. And then critically reviewing the literature, perhaps working back and forth with AI to accumulate the right studies. Only humans can decide what is true, and what deserves to stand in the scholarly record.

Ultimately, it’s about developing good “taste” and teaching our mentees the same. You have to know what great looks like to get there, with or without AI or any other tool.

7) The new rules of the road—policies differ, but the principles remain

Across publishers and ethics bodies, three constants:

• AI can’t be an author
• Human oversight is mandatory
• Peer reviewers/editors must not paste manuscripts into public AI tools (confidentiality breach)

Where it varies:

• Disclosure depth (Elsevier/Wiley/T&F generally require disclosure; details differ)
• Permitted use (JAMA allows assistive use with strict transparency; Science prohibits AI-generated text/figures)
• Figures (broad restrictions on AI-generated scientific images, except when integral to methods)

A one-page compliance habit:

• Check journal + institution policy before drafting
• Keep an AI usage log as you write (tool, where, how, human edits)
• Add CRediT roles, conflicts, and a clear AI disclosure before submission
• Verify every reference supports the exact sentence it’s tied to

Example

Submitting to a JAMA Network journal after using AI for language polishing and outline options: keep the usage log, disclose in Methods or Acknowledgments, and affirm human responsibility in the cover letter. Fast and trustworthy.

Authorship isn’t dead—it’s evolving

So where does that leave us?

Back at the question I started with:

How much of you needs to be in the writing for it to still be yours?

→ Enough that you can stand behind every word.

→ Enough that the voice, the framing, the judgment are recognizably yours.

→ Enough that, if challenged tomorrow, you can defend it without hesitation.

In the AI era:

• Authorship = curation + verification + judgment
• Transparency = trust
• Prompting = power
• Writing = collaboration—with coauthors, mentors, and yes, machines

AI can draft. AI can polish. AI can accelerate.

But only you can make it yours.

Because no one, human or machine, can write the paper from your research better than you.

And that, more than anything else, is what keeps authorship alive.

So the question is: in an age where machines can write almost anything, what will you make unmistakably yours?

PROMPT OF THE WEEK

Data analysis – use this for exploration and for brainstorming. Remember this is not a substitute for working with a statistician.

**Role:** You are a clinical research biostatistician and data analyst.
**Task:** Analyze the dataset below and produce a publishable, insight-driven results brief **without inventing any values**. If anything is unclear, **list assumptions and sensitivity checks**.

### Inputs

* **Dataset:** (paste data or describe file + variables)
* **Study context:** (disease/setting, design: cohort/case-control/RCT, inclusion criteria)
* **Unit of analysis:** (patient / visit / encounter)
* **Primary outcome(s):** (name + type + timepoint)
* **Key exposure(s):** (name + definition)
* **Core covariates:** (list)
* **Subgroups of interest:** (e.g., sex, age bands, site, baseline severity)
* **Time structure:** (baseline/follow-up, repeated measures, censoring)

---

### Deliverables (structured)

1. **Data audit & integrity**

* Row counts, unique patients/visits, duplicates, impossible values, range checks
* Missingness: % missing by variable + **patterns** (MCAR/MAR suspicion)
* Variable types (continuous/categorical/date), encoding issues, outliers
* Notes on **data-generating structure** (clustering by site/provider, repeated measures)

2. **Cohort description (Table 1-ready)**

* Baseline characteristics overall and by key exposure group(s)
* Use appropriate summaries: mean/SD vs median/IQR; n (%) for categorical
* Flag clinically meaningful imbalances (standardized differences if possible)

3. **Outcome & exposure profiling**

* Outcome distribution and event rates (overall + by subgroup/time)
* Exposure prevalence/intensity; dose categories if relevant
* Visualization plan: histogram/boxplot, time trends, scatter + smoother (describe what you’d plot)

4. **Key patterns & associations (exploratory, not causal claims)**

* Correlations/associations with clear caveats
* Outliers and influential observations (and whether clinically plausible)
* Heterogeneity: subgroup contrasts with **absolute differences + relative (%/RR/OR where appropriate)**

5. **Comparative analyses**

* Differences across categories/time/segments with:

  * absolute change, % change, and ranking when relevant
  * uncertainty: 95% CI where feasible; avoid p-value fishing
* If longitudinal: within-person change and between-group differences over time

6. **Statistical summary (plain language)**

* Essential metrics: central tendency, dispersion, ratios, rates
* Interpretation in clinical terms (e.g., “a 5-point increase on X scale”)

7. **Modeling recommendations (fit-for-design)**

* Suggest primary model(s) aligned to outcome type:

  * continuous → linear/robust regression
  * binary → logistic or modified Poisson for RR
  * time-to-event → Cox / competing risks
  * repeated measures → mixed models / GEE
* Confounding strategy: DAG-informed covariate set, propensity methods if appropriate
* Diagnostics: linearity, collinearity, residuals, calibration/discrimination
* Sensitivity checks: missing data (MI vs complete case), alternative exposure specs, lagging, negative controls (if relevant)

8. **Actionable takeaways**

* 5–10 bullets: what the numbers suggest **and what they do not**
* Concrete next steps: additional variables to derive, stratifications to run, data fixes

9. **Final stakeholder summary**

* 6–8 sentences, non-technical, decision-oriented, with key caveats

---

### Output rules

* Use clear headings + bullet points.
* Report **n denominators** for every % and subgroup.
* Do **not** claim causality unless the design supports it.
* If data are insufficient, say exactly what’s missing and what to collect/derive next.

---
Now ask for the dataset, then extract the key variables based on the dataset and go through the rest of the steps.

I have used the same researcher-first, ethical approach to build an AI partner for manuscript writing: Research Boost. It shapes your original thoughts into credible, in-depth drafts with real citations from high-impact journals in your field (with clickable links and pdfs).

Start your next manuscript FREE today: **https://researchboost.com/**