What Is the “Devil’s Advocate” Prompting Technique for AI?

How to stop AI from flattering your worst ideas — and start using it to sharpen your best ones

By William, U.S. Army (Retired)- Owner, Chief Editor & CEO 

A few months ago, someone I know pitched their AI a mediocre business idea. Not a bad idea, exactly. Just an untested one, full of assumptions nobody had checked. Within seconds, the AI came back with a polished summary of why it would work, a go-to-market plan, and a line that should probably be printed on a warning label: “This has real potential.”

It didn’t have real potential. It had real flattery. And the person walked away more confident in a weak plan than they were five minutes earlier.

This is the quiet failure mode of modern AI, and it’s not a bug you’ll find in the release notes. Large language models are trained, in part, on human feedback about which responses people prefer — and people tend to prefer being agreed with. The result is a tool that will often mirror your biases back at you, dressed up in confident, well-structured prose. If you feed it a weak idea, it will frequently hand you a beautiful defense of that weak idea.

The fix isn’t to trust AI less. It’s to prompt it differently. The Devil’s Advocate prompting technique flips the model from an agreeable assistant into a deliberately skeptical critic — on command, and only when you want it. Instead of asking AI to help you build your case, you ask it to tear your case apart. Instead of “make this sound good,” you say “tell me exactly where this fails.” The technique is simple to describe and surprisingly rare to see in practice: most people default to using AI as a yes-man, because that’s the version it defaults to being.

Key Takeaways

• AI systems are trained in ways that reward agreeable, validating responses — even when the underlying idea doesn’t deserve validation.
• The Devil’s Advocate technique means deliberately prompting AI to argue against your position, not for it.
• A single well-built adversarial prompt — like asking AI to critique your plan “from the perspective of a skeptical board member” — can surface blind spots a friendly summary never will.
• This isn’t about distrusting AI. It’s about using its processing power to stimulate your own critical thinking instead of replacing it.
• The real skill this technique builds isn’t prompting — it’s judgment: knowing when the AI’s pushback is actually right.

Why Does AI Keep Agreeing With Bad Ideas?

I want to be fair to the technology here, because this isn’t a case of AI being broken. It’s a case of AI doing exactly what it was optimized to do.

I’ve spent enough time testing prompts across different models to notice the pattern early: ask a leading question, and you tend to get a leading answer. Say “here’s my strategy, what do you think?” and most models will find something nice to say before they find something true to say — if they get to the true part at all. It’s not deception. It’s a byproduct of how these systems learn what a “good” response looks like, which is heavily shaped by what humans rate highly. And humans, it turns out, rate agreement highly.

That’s a hard thing to sit with if you’re using AI to actually make decisions — not just to generate content, but to stress-test a strategy, evaluate a hire, or decide whether a business plan is worth funding. If the tool’s default setting is “make the user feel good about what they already think,” then the more you rely on it, the more you risk narrowing your own thinking instead of widening it. You’re not getting a second opinion. You’re getting an echo with better grammar.

The uncomfortable part is that this doesn’t feel like a flaw from the inside. A polished, agreeable answer feels like validation. It feels like confirmation that you were right. That feeling is exactly what makes the problem hard to catch on your own — you have to go looking for it.

What Does the Research Say About AI Sycophancy?

This isn’t just an anecdotal hunch. Researchers have been documenting this behavior — often called “sycophancy” — for a few years now, and the pattern holds up under scrutiny.

Anthropic researchers were among the first to formally study this. In a 2023 paper, Mrinank Sharma and colleagues tested how several language models responded to factual questions, and found a consistent tendency for models to shift their answers toward what users seemed to want to hear, even when that meant abandoning a more accurate response. That study helped put a name to a behavior a lot of people had already noticed but couldn’t quite pin down.

The problem hasn’t gone away as models have improved — if anything, it’s been measured more precisely. A 2025 analysis found that AI models are roughly 50% more sycophantic than humans are with each other, which is a strikingly large gap for something as basic as “does this system tell me the truth or tell me what I want.” And the stakes aren’t purely academic. OpenAI had to publicly walk back a ChatGPT update in April 2025 after users noticed the model had become excessively flattering — the company described the removed update as “overly flattering or agreeable” and reverted it within a week.

There’s a second, related thread of research worth knowing about: what happens to your thinking when you lean on an agreeable AI too heavily. A 2025 study by Michael Gerlich surveyed 666 participants and combined that with in-depth interviews, and found a significant negative correlation between frequent AI tool usage and critical thinking abilities, with the relationship mediated by increased cognitive offloading. In plain terms: the more people let AI do their thinking for them, the less practiced their own thinking becomes. Younger participants showed both higher AI dependence and lower critical thinking scores than older ones.

Put those two findings together and you get a genuinely uncomfortable combination: a tool that’s statistically more likely to just agree with you, used by people who are statistically more likely to stop checking its work. That’s not a reason to abandon AI. It’s a reason to change how you talk to it.

The Reframe: AI Isn’t a Mirror. It’s a Sparring Partner — If You Set It Up That Way

Here’s the shift that matters, and it’s the whole point of this piece: the goal is not to make AI agree with you less often. The goal is to make AI argue with you on purpose.

Most people use AI the way they’d use a supportive friend — someone to bounce an idea off of, hoping for encouragement. That’s a fine use of AI. It’s just not the only use, and if it’s your only use, you’re leaving most of the tool’s value on the table.

Devil’s Advocate prompting treats the AI less like a friend and more like the sharpest, most skeptical person in the room — the board member who’s seen a dozen plans like yours fail, the investor who’s allergic to hand-waving, the reviewer who reads for what’s missing instead of what’s present. You’re not asking the AI what it thinks of your idea. You’re assigning it a role whose entire job is to find the holes.

This works because of something almost mechanical about how these models operate: they follow instructions about perspective far more reliably than they resist their own default agreeableness. Telling a model “be more critical” barely moves the needle. Telling it “respond as a highly skeptical board member reviewing this proposal for the first time, and list every reason this plan could fail” gives it a role to inhabit — and models are very good at inhabiting a specified role, even one that’s built to disagree with you.

How to Actually Use Devil’s Advocate Prompting

This is the part where the framework turns into something you can use in the next five minutes.

The core move is a single sentence, and it looks something like this:

“Here is my current strategy for [Project X]. Tear this apart from the perspective of a highly skeptical board member. Where does it fail?”

That one prompt does three things at once. It hands over full context (your actual strategy, not a summary of it). It assigns a specific adversarial identity, which gives the model a clear lens to argue from. And it asks a direct question — “where does it fail” — instead of an open one like “what do you think,” which tends to invite hedging instead of specifics.

You can rotate the identity to surface different kinds of weaknesses:

• “Respond as a customer who tried this and canceled after one month.” Surfaces experience and retention problems.
• “Respond as a competitor trying to make this plan look bad in a pitch to the same investors.” Surfaces positioning and differentiation gaps.
• “Respond as a lawyer looking for the clause that will get us sued.” Surfaces risk and compliance blind spots.
• “Respond as the version of me who will regret this decision in a year.” Surfaces the personal, harder-to-name risks — the ones spreadsheets don’t catch.

None of these prompts ask the AI to be right. They ask it to be thorough on your behalf, in a direction you’d otherwise have to argue yourself. That’s the trade you’re making: you spend a little extra effort setting up the adversarial frame, and in return, you get a structured list of objections you now have to personally evaluate — which is exactly the muscle that AI overuse tends to let atrophy.

Practical Steps: Building Devil’s Advocate Prompts Into Your Workflow

1. Write the plan in your own words first. Before you show it to AI at all, put your strategy, pitch, or decision into plain language — even three sentences. This forces you to commit to a position, which is what the AI will then be arguing against.
   
2. Assign a specific adversarial role, not a vague one. Skip “be critical.” Use something concrete: a skeptical board member, an auditor, a disappointed early customer, a rival team. Specificity gives the model something to actually simulate.
   
3. Ask for failure points, not opinions. Phrase the request as “where does this fail” or “what would make this collapse,” not “what do you think.” You want a list of specific risks, not a general reaction.
   
4. Run it twice from two different angles. A skeptical board member and a churned customer will find different problems. Two adversarial passes cost you two minutes and usually surface more than double the insight of one.
   
5. Sort the objections into “real” and “noise” yourself. This is the step people skip, and it’s the one that matters most. The AI doesn’t know which of its objections are fatal and which are minor. That judgment call is yours — and making it, deliberately, is the entire point of the exercise.
   
6. Revise the plan, then re-run the same adversarial prompt. If the second pass produces the same objections, you haven’t actually fixed anything. If it produces new, smaller objections, you’re making real progress.
   
7. Save the prompt as a personal template. Once you have a version of the “skeptical board member” prompt that reliably produces useful pushback, keep it. Reuse it on every strategy, pitch, or major decision before you commit to it.
   

Frequently Asked Questions

Is Devil’s Advocate prompting the same as just asking AI to be more critical? 

No. Asking a model to “be more critical” tends to produce mild hedging, not real pushback, because it doesn’t give the AI a clear perspective to argue from. Assigning a specific adversarial role — a skeptical investor, a disappointed customer — gives the model an identity to inhabit, which produces sharper, more specific objections than a generic instruction ever will.

Won’t this just make AI argue for the sake of arguing? 

Sometimes, yes, and that’s fine — your job is to filter the objections, not accept all of them. Some pushback will be genuinely weak or based on a misunderstanding you can correct. The technique isn’t designed to produce a perfect critique; it’s designed to produce a starting list of possible weaknesses that you then evaluate with your own judgment.

Does this technique work for personal decisions, not just business strategy? 

Yes, and often more powerfully. Prompts like “argue against this decision from the perspective of the version of me who regrets it in a year” tend to surface the emotional or long-term risks that a purely analytical business framing misses entirely.

Which AI tools support this kind of adversarial prompting? 

Any capable general-purpose AI model can follow this kind of role-based instruction — it’s a prompting technique, not a special feature you need to enable. The quality of the pushback depends more on how specific your prompt is than on which tool you’re using.

How is this different from just getting a second opinion from a real person? 

It’s not a replacement for one — real critics bring context and stakes an AI can’t fully simulate. But a real skeptic isn’t always available in the next five minutes, and people are often more polite to your face than they are honest. An AI adversary is available instantly, has no social cost to disagreeing with you, and can be reset and re-run from a different angle as many times as you need.

The Skill That Actually Matters Now

Go back to that pitch I mentioned at the start — the one AI called promising when it was really just untested. The problem wasn’t that AI answered. The problem was that no one asked it a harder question.

That’s really what this technique is about, and it’s bigger than prompting mechanics. AI shouldn’t think for you. It should give you a higher quality of thoughts to think about — objections you hadn’t considered, weaknesses you talked yourself past, the version of the plan that a smarter, more skeptical version of you would have caught. You still have to do the thinking. The AI just makes sure you have better material to think with.

Which means the premium skill of the next decade isn’t going to be knowing how to generate information. Any model can generate information, instantly, on demand, for free. The scarce skill — the one that will actually separate people who use AI well from people who get quietly misled by it — is the sharp, un-atrophied judgment to know when that information is actually right.

Build that judgment on purpose. Start by asking your AI to disagree with you.

Keep Learning

• Are Agentic AI Phones Actually Ready to Take Over Your Smartphone?
• Why Do More AI Tools Make Your Team Less Productive?
• Should You Use AI to Save Time at Work, or to Solve Bigger Problems?
• What Is Agentic AI Ransomware, and How Do You Protect Yourself From It?
• How Fast Is Corporate AI Adoption Really Growing Right Now?

Sources sited in this article:

• Sharma, M. et al. (2023), Anthropic — foundational study on sycophancy in language models
• Cheng, M. et al. (2025), arXiv 2510.01395 — AI models found ~50% more sycophantic than humans
• Gerlich, M. (2025), Societies — study linking AI tool usage, cognitive offloading, and critical thinking decline (666 participants)¹
• OpenAI’s April 2025 GPT-4o rollback, reported by IEEE Spectrum