Vibe Learning — Issue #23
AI as the Devil’s Advocate
How to Use AI to Attack Your Best Ideas Before the Real World Does
We usually ask AI to help us make ideas better.
Improve this.
Expand this.
Give me more ideas.
Make this argument stronger.
Tell me whether this is good.
There’s nothing wrong with that.
But there’s another way to use AI that may be far more valuable:
Ask it to prove you wrong.
Not because your idea is necessarily bad.
Because good ideas should survive good criticism.
Welcome to AI as the Devil’s Advocate.
Your Best Ideas Need Opposition
Imagine you’ve spent weeks developing an idea.
Maybe it’s:
- a business
- a lesson
- a product
- an article
- a strategy
- a software project
- a career decision
- an argument
You understand it.
You’ve researched it.
You’ve refined it.
And naturally, you’ve become attached to it.
That’s where a problem begins.
The more invested we become in an idea, the harder it can be to see what’s wrong with it.
We start looking for evidence that confirms our thinking.
We explain away contradictions.
We defend assumptions we haven’t actually tested.
Sometimes what our idea needs most isn’t another supporter.
It needs an intelligent opponent.
AI can play that role.
From Socratic AI to Adversarial AI
In Issue #22, we explored the Socratic AI.
Instead of giving answers, AI asks questions:
Why?
What evidence supports that?
What assumption are you making?
What else could be true?
The Devil’s Advocate takes this one step further.
Instead of simply questioning your reasoning, we deliberately ask AI to construct the strongest challenge against it.
The objective isn’t conflict.
It’s stress testing.
Think of it like testing a bridge before opening it to traffic.
You don’t test only the conditions you expect.
You apply pressure.
You simulate failure.
You discover weaknesses.
Then you strengthen them.
Ideas deserve the same treatment.
The Confirmation Trap
Suppose you ask AI:
“Here’s my business idea. What do you think?”
There’s a good chance you’ll receive something encouraging.
“Interesting concept.”
“Strong potential.”
“Here are some opportunities…”
That feels good.
But feeling good isn’t the same as learning.
Try this instead:
“Assume I am emotionally invested in this idea and may be overlooking problems. Make the strongest evidence-based case for why I should NOT pursue it.”
Now the conversation becomes very different.
Maybe AI identifies:
- weak demand
- expensive customer acquisition
- existing alternatives
- regulatory barriers
- poor differentiation
- unrealistic assumptions
- technical complexity
That doesn’t mean the idea is dead.
It means you finally have something useful to investigate.
Don’t Ask AI to Be Negative
There’s an important distinction here.
The goal isn’t:
“Tell me everything that’s wrong.”
That’s easy.
Almost any idea can be criticized.
The goal is:
“Find the strongest legitimate weaknesses.”
A useful Devil’s Advocate doesn’t complain.
It reasons.
Tell AI:
“Do not criticize simply for the sake of criticism. Focus on weaknesses that could materially affect the outcome. Explain the reasoning behind each challenge.”
Now you’re getting analysis instead of negativity.
The Red Team Mindset
Security professionals sometimes use red teams.
One team builds or protects a system.
Another tries to break it.
Why?
Because discovering weaknesses internally is much cheaper than discovering them after a real attack.
We can borrow the same mindset for learning.
Before presenting your argument…
Attack it.
Before launching your project…
Break it.
Before making your decision…
Challenge it.
Before publishing your prediction…
Try to disprove it.
Call it Red Team Learning.
The AI Red Team Loop
Here’s a simple process:
Create → Attack → Defend → Investigate → Revise → Retest
Let’s break it down.
Step 1 — Create Your Position First
This is important.
Don’t immediately ask AI what you should think.
Start with your thinking.
Write:
your idea
your argument
your prediction
your solution
your plan
For example:
“I believe universities should replace most traditional lectures with AI-powered personalized learning.”
Now there’s something concrete to test.
Step 2 — Ask AI to Attack It
Prompt:
“Act as an intelligent critic who disagrees with this position. Construct the strongest arguments against it.”
AI might raise:
equity concerns
accuracy problems
social learning
student motivation
teacher expertise
privacy
overdependence
assessment integrity
Good.
Don’t defend yourself yet.
Read the criticism.
Step 3 — Rank the Threats
Not every criticism matters equally.
Ask:
“Rank these objections from strongest to weakest based on how seriously they threaten my argument. Explain the ranking.”
This is important.
You want to focus your learning on the weaknesses that matter most.
Step 4 — Defend Your Position
Now respond.
Take the strongest criticism.
Try to answer it yourself.
Don’t immediately ask AI to write your defense.
You might say:
“I think the social-learning objection could be addressed by keeping collaborative workshops while replacing only passive lectures.”
Now you’re refining the idea.
Step 5 — Attack the Defense
Here’s where it gets interesting.
Tell AI:
“Now challenge my response. Don’t let me escape the original criticism with a superficial answer.”
Maybe AI asks:
If collaborative workshops remain, how much instructional time and cost have you actually eliminated?
Excellent.
The idea becomes more precise again.
Step 6 — Investigate
Eventually you’ll reach something neither you nor AI should simply guess about.
Maybe the argument depends on whether AI tutoring actually improves learning outcomes.
Now stop debating.
Research.
That’s an important lesson.
Argument should eventually lead to evidence.
Ask:
“What evidence would we need to determine which side of this argument is stronger?”
Now you’ve turned disagreement into a research plan.
Step 7 — Revise
After the challenge, rewrite your original position.
Maybe:
“Universities should replace most traditional lectures with AI-powered personalized learning.”
becomes:
“Universities should experiment with AI-supported personalized instruction for some content delivery while preserving human-led discussion, collaboration, mentoring, and assessment where those interactions create distinct learning value.”
Less dramatic?
Yes.
Stronger?
Much.
That’s progress.
The Steelman Rule
There’s a useful concept called steelmanning.
Instead of attacking the weakest version of an opposing argument, construct the strongest version.
This matters enormously with AI.
Otherwise, you can accidentally ask AI to create a weak opponent that your idea easily defeats.
Try:
“Before criticizing my idea, construct the strongest possible version of the opposing position. Assume the opposing side is intelligent, informed, and acting in good faith.”
This makes the exercise much more valuable.
Don’t defeat a straw man.
Challenge the strongest opponent you can build.
Devil’s Advocate Mode #1 — The Skeptical Customer
Suppose you’re developing a product.
Tell AI:
“Act as a skeptical potential customer. I will explain my product. Ask why I should care, why I should trust it, why I should switch from what I already use, and why the price is justified.”
This can expose weak value propositions immediately.
If you can’t answer:
“Why should I care?”
You may not have a product problem.
You may have a problem-definition problem.
Mode #2 — The Tough Student
If you’re teaching something, ask AI to become the student who doesn’t automatically accept your explanation.
Prompt:
“Act as an intelligent but skeptical student. Challenge unclear explanations, identify assumptions about prior knowledge, and tell me when an example doesn’t actually prove the point I’m making.”
This is incredibly useful for teachers.
You may discover:
missing steps
undefined terminology
weak examples
logical jumps
assumed background knowledge
Your lesson improves before students ever see it.
Mode #3 — The Code Reviewer
If you’re programming, don’t just ask:
“Does this code work?”
Ask:
“Act as a strict senior code reviewer. Try to break this approach. Look for edge cases, incorrect assumptions, maintainability problems, security concerns, and situations where the design could fail.”
Working code isn’t necessarily good code.
Adversarial review pushes beyond:
“Does it run?”
toward:
“How does it fail?”
That’s a much more interesting question.
Mode #4 — The Editor Who Disagrees
Before publishing an article, give AI the draft.
Ask:
“Assume you disagree with my central argument. Identify the three places where my reasoning is easiest to challenge.”
Then:
“What evidence or explanation would make each section harder to dispute?”
Now AI isn’t just correcting grammar.
It’s strengthening thought.
Mode #5 — The Competitor
Have a business idea?
Ask:
“Imagine you run the strongest competitor in this market. How would you respond to my launch? What advantages would you exploit? How could you make my product irrelevant?”
That’s uncomfortable.
Good.
Now ask:
“What could I change before launch to make those responses less effective?”
Attack becomes strategy.
Mode #6 — The Future You
This one is different.
Ask AI:
“Imagine it’s three years from now and this decision turned out badly. Interview me from that future perspective to determine what warning signs I ignored today.”
This resembles a premortem.
Instead of asking:
“What could go wrong?”
Imagine it already did.
Then work backward.
You may discover risks that optimism was hiding.
The Premortem Exercise
Try this before an important project.
Tell AI:
“It is one year from today. My project failed badly.
Generate five plausible reasons it failed.
Avoid random disasters. Focus on problems that could reasonably be anticipated today.
Then ask me what I could do now to reduce each risk.”
This changes risk from something abstract into something actionable.
The “Kill My Idea” Session
Here’s a powerful 15-minute exercise.
Take your favorite idea.
Give AI this instruction:
“For the next 15 minutes, your job is to try to kill this idea.
Do not be rude or pessimistic.
Assume I want to discover serious weaknesses before investing more time or resources.
Attack assumptions, demand, execution, alternatives, unintended consequences, and evidence.
Ask one challenge at a time.
After each answer, decide whether my response genuinely resolves the issue or merely avoids it.
If I successfully defend something, move to the next weakness.
At the end, identify:
- The strongest surviving objection.
- The weakest assumption.
- The most important unanswered question.
- The cheapest experiment I could run next.”
That final point matters.
The exercise should lead to action.
Use AI to Attack Its Own Answers
Here’s where this gets even more interesting.
AI gives you a recommendation.
Don’t accept it.
Ask:
“Now act as an expert who strongly disagrees with your previous answer. What did the first answer overlook?”
Then:
“Compare both positions. Where is the real uncertainty?”
This is one of the simplest ways to avoid treating the first AI response as authoritative.
AI output should often be the beginning of thinking, not the end.
Create Two Competing AIs
Try this experiment.
AI #1 — The Advocate
“Make the strongest case FOR this idea.”
AI #2 — The Critic
“Make the strongest case AGAINST this idea.”
Then use a third role:
AI #3 — The Judge
“Compare these arguments. Identify which claims depend on evidence, which are assumptions, and what information would help resolve the disagreement.”
Now you have:
Thesis → Antithesis → Evaluation
But remember:
AI isn’t actually three independent experts.
You’re using different perspectives to structure your own thinking.
That’s the value.
Beware of Fake Debate
There’s a danger here.
AI can generate convincing arguments for almost anything.
A polished argument isn’t necessarily true.
That’s why Devil’s Advocate mode must eventually ask:
What evidence would settle this?
When AI says:
“Studies show…”
Ask:
Which studies?
When it makes a prediction:
What assumptions drive that prediction?
When it gives statistics:
Verify them.
Adversarial thinking isn’t about generating more words.
It’s about improving judgment.
The Confidence Test
After completing a Devil’s Advocate session, ask yourself:
Before
How confident was I?
8/10?
After
How confident am I now?
Maybe 6/10.
That doesn’t mean the exercise failed.
It may mean your understanding improved.
Learning sometimes makes us less certain because we finally understand the complexity.
That’s intellectual progress.
The Strong-Idea Test
After the challenge, your idea should end up in one of four places.
1. Stronger
The criticism helped improve it.
2. Narrower
The original claim was too broad.
3. Different
The challenge revealed a better direction.
4. Dead
The idea couldn’t survive basic scrutiny.
Number four isn’t failure.
Discovering that early may save enormous time.
The Master Devil’s Advocate Prompt
Use this whenever an idea matters enough to challenge:
Act as my Devil’s Advocate and intellectual red team.
Your purpose is not to agree with me or oppose me automatically. Your purpose is to identify weaknesses that could materially affect whether my idea, argument, strategy, or decision succeeds.
First, make sure you understand my position.
Then:
- identify hidden assumptions
- construct the strongest opposing position
- challenge weak evidence
- look for alternative explanations
- identify edge cases
- examine unintended consequences
- test whether my conclusions actually follow from my evidence
- distinguish facts from assumptions
- identify what information is missing
Ask one major challenge at a time.
Let me defend my position.
If my defense is weak, continue challenging it.
If my defense is strong, acknowledge that and move to the next issue.
Do not invent facts simply to create an objection.
At the end, summarize:
- What survived the challenge.
- What needs improvement.
- What remains uncertain.
- What evidence I should gather.
- What experiment or action I should take next.
Your job isn’t to destroy good ideas. It’s to make weak ideas stronger and expose bad ones early.
Weekly Challenge: Attack Something You Believe
Choose one idea you’re confident about.
Not something trivial.
Something connected to:
your work
your learning
your career
your business
your teaching
your creative process
Write your position before opening AI.
Then run a 15-minute Devil’s Advocate session.
Don’t try to win.
Try to discover.
At the end, write:
“The strongest challenge to my thinking was…”
Then:
“What I believe now is…”
Compare the two.
That’s learning.
The Bigger Lesson
AI is often marketed as something that removes friction.
Faster writing.
Faster research.
Faster coding.
Faster answers.
But some of AI’s greatest value may come from intentionally adding useful friction.
Questioning us.
Challenging us.
Disagreeing with us.
Making us defend our reasoning.
Showing us what we missed.
Because speed isn’t always the goal.
Sometimes the goal is avoiding the wrong destination.
Final Reflection
Your ideas don’t need AI to tell them they’re brilliant.
They need pressure.
They need questions.
They need evidence.
They need opposition.
They need opportunities to fail safely before they fail publicly.
AI gives us something historically difficult to access:
an opponent available whenever we need one.
Use it.
Not to create endless arguments.
Not to make yourself doubt everything.
But to build ideas that deserve confidence.
Because the strongest idea isn’t the one nobody challenges.
It’s the one that becomes better because somebody did.
That is the power of using AI as the Devil’s Advocate.
Coming Next Issue
Issue #24 — The AI Feedback Loop
Getting feedback is easy with AI.
Getting useful feedback that actually makes you better is much harder.
Next, we’ll build an AI feedback system that moves beyond “make this better” and turns every project, mistake, revision, and attempt into a deliberate improvement cycle.
Question for readers
What’s one idea, strategy, or assumption you’re confident about—but would be willing to let AI try to prove wrong?