You didn’t train for years to become a professional approver of confident AI text.
A normal research day can now look like this:
Email from a collaborator.
“Summarize this thread.”
Reviewer 2 comment.
“Draft a response.”
New paper drop.
“Give me the key points.”
Results section.
“Turn these bullets into prose.”
By 4 pm, you’ve “done work” all day.
→ But you barely touched the work.
→ You touched the outputs.
That’s the quiet risk of AI in academic research right now.
Not that AI writes badly.
It writes smoothly enough that you stop noticing when your own thinking goes missing.
You become a professional validator of a machine’s confidence.
And in research, confidence without thinking is how bad science looks polished.
Here are 5 strategies I use to stop AI from killing my critical thinking in research:
1️⃣ Make AI argue with you
This is how I have seen most prompt AI:
“Write my discussion.”
“Make this sound stronger.”
“Add why this matters.”
Of course it sounds good.
You asked for agreement.
You asked for confidence.
You asked for a clean story.
That is the exact opposite of what your critical thinking is.
If you want to stay sharp, you need friction.
You need pushback.
You need someone in the room who says, “Hold on. Prove it.”
I use AI as an adversary first. Not as a writer first.
Prompts I actually use:
- “What’s the strongest counterargument to my conclusion.”
- “List 5 alternative explanations that do not require my hypothesis to be true.”
- “If you were Reviewer 2, what would you attack first, and why.”
- “What is the most plausible confounder that could explain this association.”
- “What would someone who dislikes my method say.”
Then I go one layer deeper:
“Write the harshest one-paragraph critique of my study, but make it fair.”
Now you’re not collecting text.
You’re collecting angles.
You’re forcing your brain to defend its reasoning.
That’s where critical thinking lives.
EXAMPLE:
Let’s say your paper shows an association between an inflammatory condition and cardiovascular risk.
The easy AI move is a discussion that says:
“This supports the inflammatory hypothesis and highlights the need for early intervention.”
Clean.
Polished.
Totally plausible.
Now do the hard move.
Ask AI for alternative explanations.
→ Measurement bias.
→ Selection bias.
→ Confounding by indication.
→ Differences in healthcare utilization.
→ Residual confounding you cannot adjust away.
This is when you will have a real discussion.
Not a marketing paragraph.
If AI can’t challenge your idea, it’s not helping you think.
It’s just helping you ship (mediocrity).
2️⃣ Keep one “hands-on” step in every task
This is the simplest rule I follow.
Every time I use AI, I keep one critical step that is fully mine (for each task).
One step that forces direct contact with the material.
Because when you stay hands-off the whole time, your can never form a real mental model.
You just watch a stream of polished conclusions go by.
Pick one part you do the hard way.
- Read one primary paper start to finish. No summary first.
- Write the first ugly paragraph yourself. Not pretty, just true.
- Sketch the causal story on paper. Exposure, outcome, what could mess it up.
- Do the first pass of the table by hand. Even if the formatting is awful.
The hands-on step forces ownership.
It forces you to feel the gaps.
It forces you to notice uncertainty.
AI smooths uncertainty away.
(Premature coherence is a big problem with almost all AI models. ChatGPT falls for this the most.)
Science is built on noticing uncertainty.
A practical way to use this today:
Take the next task you were going to send to AI.
Add one manual checkpoint.
If it’s a paper summary, you read the abstract and limitations yourself first.
If it’s an R analysis, you try writing the code yourself before AI touches anything.
If it’s a discussion, you write your interpretation in plain language first.
Then let AI help you tighten.
This is how you keep your brain in the loop.
Pro Tip: My typical workflow is always analog first → then only digital. For example, I spend hours sketching out my grant aims on a piece of paper before moving to digital.
3️⃣ Force the assumptions into daylight
AI loves to sound certain.
Premature coherence again is a real issue.
It will produce a clean answer even when the question is messy.
It will fill missing steps like they were always obvious.
It will connect dots that should never be connected.
But science does not work like that.
So I make assumption hunting a default step.
I ask:
- “What assumptions are you making.”
- “What would make this answer wrong.”
- “What information are you missing.”
- “Where are you guessing.”
- “List the top 10 hidden assumptions in this argument.”
Then I ask for an assumption audit I can scan:
- Assumption
- Why it matters
- How to test it
- How to weaken the claim if it fails
This matters everywhere.
In stats, assumptions hide in model choices.
Linearity.
Missingness.
Independence.
Generalizability.
In interpretation, assumptions hide in storytelling.
We assume causality when we only have association.
We assume mechanism when we only have correlation.
We assume clinical relevance when we only have statistical significance.
AI will happily write that story.
Your job is to stop it.
Coherence is not accuracy.
Coherence is just a well-written chain of words.
So when the output sounds too clean, I slow down.
Then I force the assumptions into daylight.
4️⃣ Treat outputs like a draft from a new trainee
When a trainee hands you something clean, you still check the data.
Same rule here.
I treat every AI output like a first draft from a smart intern.
Helpful.
Fast.
Not accountable.
That mindset keeps you safe.
And it keeps you sharp.
Here’s my basic verification checklist:
Numbers. Re-calculate. Even simple ones.
If you can’t reproduce it quickly, it doesn’t belong.
Citations. Open and verify.
Does the paper actually say that.
Or is it just vaguely related.
Claims. Trace back to the source.
What part of the paper supports the claim.
If it’s not supported, delete it or rewrite it as uncertainty.
Language. Downgrade certainty.
AI loves absolute words.
“Demonstrates.”
“Proves.”
“Confirms.”
Most research does not prove anything.
So I rewrite into honest science language.
“Suggests.”
“Is consistent with.”
“May indicate.”
Not timid.
Correct.
Then I do 2 passes.
Pass 1: logic and structure.
Does the argument follow the data.
Pass 2: facts and references.
Can I trace every strong claim to something real.
If I can’t point to where it came from, it doesn’t belong.
That single rule protects your reputation.
5️⃣ End with a 60-second ownership check
This is my favorite habit because it’s small and powerful.
Before you paste anything into the draft, pause.
Write this, in your own words:
- What do I believe.
- Why do I believe it.
- What would change my mind.
Three sentences.
No AI.
It sounds almost silly.
But it solves the real problem.
Not writing speed.
Thinking ownership.
It forces you to take a stance.
It forces you to name your evidence.
It forces you to admit what would falsify your story.
That is science.
It also protects your critical thinking.
If AI drafts everything, you will forget what you “wrote.”
You’ll remember the conclusion.
Not the reasoning.
Then 6 months later someone asks you a direct question about your own paper, and you feel that gap.
You know the feeling.
You wrote it.
But you don’t fully own it.
The ownership check prevents that.
Because it makes you rehearse your reasoning in your own words.
Here’s the bar I use:
If I can’t explain the claim out loud in plain language, I’m not ready to put it in the manuscript.
If I can’t explain what would change my mind, I’m not thinking like a scientist.
That 60 seconds keeps me honest.
A simple way to implement this
Try this sequence on your next task.
Step 1: You write the ugly version first.
Two to five bullets.
Plain language.
No polish.
Step 2: AI argues with you.
Counterarguments.
Alternative explanations.
What you might be missing.
Step 3: You adjust your stance.
You decide what holds.
What weakens.
What you will not claim.
Step 4: AI helps you express it clearly.
Now the writing support is safe.
Because the thinking already happened.
This is how you use AI without losing what makes you the expert.
Use AI not just as an assistant that obeys.
As a training partner that pushes back.
AI will keep getting better.
The question is whether we’ll use it to become sharper.
Or mindlessly outsource what uniquely makes us the ‘expert’.
PROMPT OF THE WEEK
Draft a powerful “statistical power” section
Use below prompt to critically think and craft your statistical power approach.
(Make sure you turn on “Thinking”).
You didn’t train for years to become a professional approver of confident AI text.
A normal research day can now look like this:
Email from a collaborator.
“Summarize this thread.”
Reviewer 2 comment.
“Draft a response.”
New paper drop.
“Give me the key points.”
Results section.
“Turn these bullets into prose.”
By 4 pm, you’ve “done work” all day.
→ But you barely touched the work.
→ You touched the outputs.
That’s the quiet risk of AI in academic research right now.
Not that AI writes badly.
It writes smoothly enough that you stop noticing when your own thinking goes missing.
You become a professional validator of a machine’s confidence.
And in research, confidence without thinking is how bad science looks polished.
Here are 5 strategies I use to stop AI from killing my critical thinking in research:
## 1️⃣ Make AI argue with you
This is how I have seen most prompt AI:
“Write my discussion.”
“Make this sound stronger.”
“Add why this matters.”
Of course it sounds good.
You asked for agreement.
You asked for confidence.
You asked for a clean story.
That is the exact opposite of what your critical thinking is.
If you want to stay sharp, you need friction.
You need pushback.
You need someone in the room who says, “Hold on. Prove it.”
> I use AI as an adversary first.
Not as a writer first.
>
Prompts I actually use:
- “What’s the strongest counterargument to my conclusion.”
- “List 5 alternative explanations that do not require my hypothesis to be true.”
- “If you were Reviewer 2, what would you attack first, and why.”
- “What is the most plausible confounder that could explain this association.”
- “What would someone who dislikes my method say.”
Then I go one layer deeper:
“Write the **harshest one-paragraph critique** of my study, but **make it fair**.”
Now you’re not collecting text.
You’re collecting angles.
You’re forcing your brain to defend its reasoning.
That’s where critical thinking lives.
**EXAMPLE:**
Let’s say your paper shows an association between an inflammatory condition and cardiovascular risk.
The easy AI move is a discussion that says:
“This supports the inflammatory hypothesis and highlights the need for early intervention.”
Clean.
Polished.
Totally plausible.
Now do the hard move.
Ask AI for **alternative explanations**.
→ Measurement bias.
→ Selection bias.
→ Confounding by indication.
→ Differences in healthcare utilization.
→ Residual confounding you cannot adjust away.
This is when you will have a real discussion.
Not a marketing paragraph.
If AI can’t challenge your idea, it’s not helping you think.
It’s just helping you ship (mediocrity).
---
## 2️⃣ Keep one “hands-on” step in every task
This is the simplest rule I follow.
Every time I use AI, I keep one critical step that is fully mine (for each task).
One step that forces direct contact with the material.
Because when you stay hands-off the whole time, your can never form a real mental model.
You just watch a stream of polished conclusions go by.
Pick one part you do the hard way.
- Read one primary paper start to finish. No summary first.
- Write the first ugly paragraph yourself. Not pretty, just true.
- Sketch the causal story on paper. Exposure, outcome, what could mess it up.
- Do the first pass of the table by hand. Even if the formatting is awful.
The hands-on step forces ownership.
It forces you to feel the gaps.
It forces you to notice uncertainty.
AI smooths uncertainty away.
(**Premature coherence** is a big problem with almost all AI models. ChatGPT falls for this the most.)
Science is built on noticing uncertainty.
**A practical way to use this today:**
Take the next task you were going to send to AI.
Add one manual checkpoint.
If it’s a paper summary, you read the abstract and limitations yourself first.
If it’s an R analysis, you try writing the code yourself before AI touches anything.
If it’s a discussion, you write your interpretation in plain language first.
Then let AI help you tighten.
This is how you keep your brain in the loop.
**Pro Tip:** My typical workflow is always analog first → then only digital. For example, I spend hours sketching out my grant aims on a piece of paper before moving to digital.
---
## 3️⃣ Force the assumptions into daylight
AI loves to sound certain.
Premature coherence again is a real issue.
It will produce a clean answer even when the question is messy.
It will fill missing steps like they were always obvious.
It will connect dots that should never be connected.
But science does not work like that.
So I make assumption hunting a default step.
I ask:
- “What assumptions are you making.”
- “What would make this answer wrong.”
- “What information are you missing.”
- “Where are you guessing.”
- “List the top 10 hidden assumptions in this argument.”
Then I ask for an assumption audit I can scan:
- Assumption
- Why it matters
- How to test it
- How to weaken the claim if it fails
This matters everywhere.
In stats, assumptions hide in model choices.
Linearity.
Missingness.
Independence.
Generalizability.
In interpretation, assumptions hide in storytelling.
We assume causality when we only have association.
We assume mechanism when we only have correlation.
We assume clinical relevance when we only have statistical significance.
AI will happily write that story.
Your job is to stop it.
Coherence is not accuracy.
Coherence is just a well-written chain of words.
So when the output sounds too clean, I slow down.
Then I force the assumptions into daylight.
---
## 4️⃣ Treat outputs like a draft from a new trainee
When a trainee hands you something clean, you still check the data.
Same rule here.
I treat every AI output like a first draft from a smart intern.
Helpful.
Fast.
Not accountable.
That mindset keeps you safe.
And it keeps you sharp.
Here’s my basic verification checklist:
**Numbers.** Re-calculate. Even simple ones.
If you can’t reproduce it quickly, it doesn’t belong.
**Citations.** Open and verify.
Does the paper actually say that.
Or is it just vaguely related.
**Claims.** Trace back to the source.
What part of the paper supports the claim.
If it’s not supported, delete it or rewrite it as uncertainty.
**Language.** Downgrade certainty.
AI loves absolute words.
“Demonstrates.”
“Proves.”
“Confirms.”
Most research does not prove anything.
So I rewrite into honest science language.
“Suggests.”
“Is consistent with.”
“May indicate.”
Not timid.
Correct.
Then I do **2 passes**.
**Pass 1:** logic and structure.
Does the argument follow the data.
**Pass 2:** facts and references.
Can I trace every strong claim to something real.
If I can’t point to where it came from, it doesn’t belong.
That single rule protects your reputation.
---
## 5️⃣ End with a 60-second ownership check
This is my favorite habit because it’s small and powerful.
Before you paste anything into the draft, pause.
Write this, in your own words:
- What do I believe.
- Why do I believe it.
- What would change my mind.
Three sentences.
No AI.
It sounds almost silly.
But it solves the real problem.
Not writing speed.
Thinking ownership.
It forces you to take a stance.
It forces you to name your evidence.
It forces you to admit what would falsify your story.
That is science.
It also protects your critical thinking.
If AI drafts everything, you will forget what you “wrote.”
You’ll remember the conclusion.
Not the reasoning.
Then 6 months later someone asks you a direct question about your own paper, and you feel that gap.
You know the feeling.
You wrote it.
But you don’t fully own it.
The ownership check prevents that.
Because it makes you rehearse your reasoning in your own words.
Here’s the bar I use:
If I can’t explain the claim out loud in plain language, I’m not ready to put it in the manuscript.
If I can’t explain what would change my mind, I’m not thinking like a scientist.
That 60 seconds keeps me honest.
---
## A simple way to implement this
Try this sequence on your next task.
**Step 1:** You write the ugly version first.
Two to five bullets.
Plain language.
No polish.
**Step 2:** AI argues with you.
Counterarguments.
Alternative explanations.
What you might be missing.
**Step 3:** You adjust your stance.
You decide what holds.
What weakens.
What you will not claim.
**Step 4:** AI helps you express it clearly.
Now the writing support is safe.
Because the thinking already happened.
This is how you use AI without losing what makes you the expert.
Use AI not just as an assistant that obeys.
As a training partner that pushes back.
AI will keep getting better.
The question is whether we’ll use it to become sharper.
Or mindlessly outsource what uniquely makes us the ‘expert’.
### PROMPT OF THE WEEK
**Draft a powerful “statistical power” section**
Use below prompt to critically think and craft your statistical power approach.
(Make sure you turn on “Thinking”).After you run this step-by-step with ChatGPT, make sure you reproduce the exact calculation in PSPower or other power calculation softwares using the same inputs.
💬 What’s one place in your research workflow where you’ve noticed yourself switching from “thinking” to “approving”?
P.S. I built Research Boost to keep the academic writing research first, grounded on the researcher’s thinking and draft materials.
It only pulls from your own draft plus high-impact, peer-reviewed sources and verifies every citation, so your facts, numbers, and statistics stay rooted in real, verifiable sources, not AI guesses.
Try it FREE: https://researchboost.com/