Most people use AI to get an answer faster. Using it as a thinking partner means the opposite: handing it a conclusion you've already reached and asking it to take the argument apart.
Asking AI “what should I do” and asking AI “argue against what I’ve decided” produce two completely different effects on your thinking, even though they use the same tool. The first ends the thinking, because a confident-sounding answer feels like a finished one. The second extends it, because now you have a specific objection to test rather than a vague feeling of certainty. This piece breaks down three types of questions that do the second thing, works them through on a hiring decision, a pricing memo, and a strategy memo, and gives you a short self-check for the moment sharpening turns into quietly letting AI decide for you.
A hiring manager has two finalist profiles open on a Friday afternoon, both strong, both due for discussion at Monday’s debrief. She already has a lean toward one of them, the one whose background reminds her of the strongest hire on her team last year. She opens an AI tool, pastes in both profiles, and types something close to “which of these two should I hire for the team lead role.”
What comes back reads well. It’s organized, confident, and it lands on the candidate she was already leaning toward, with three reasons that sound sensible. She closes the tab more certain than she was five minutes earlier. Nothing about that exchange was wrong, exactly. It also didn’t sharpen anything. It handed her first instinct back to her in better sentences.
That’s the difference this piece is about. An answer machine takes your question and returns a conclusion. A thinking partner takes your conclusion and asks what would have to be true for it to be wrong. Both run on the same underlying model. The difference is entirely in what you ask it to do, and almost nobody asks it to do the second thing, because getting a fast, confident answer feels like the whole point of opening the tool in the first place.
Put the two side by side and it’s easier to notice which one you’re actually running on any given day.
| Answer machine | Thinking partner | |
|---|---|---|
| What you ask it | “What should I do?” | “Here’s what I’ve decided. Argue against it.” |
| What comes back | A recommendation, phrased with total confidence | A case against your plan, built in its own words |
| What it does to your thinking | Ends it. A confident-sounding answer feels like a finished one, so you stop turning the problem over. | Extends it. You now have a specific objection to test, not just a feeling that you’re probably right. |
Same tool, same model, two different jobs. Most people only ever run the left column.
This isn’t an argument against ever asking AI for an answer. Plenty of tasks genuinely benefit from speed, the kind of work covered in a companion piece on AI as a workflow accelerator. This piece is about a narrower use: moments where the value isn’t producing something faster, it’s finding out whether what you already believe actually holds up before you act on it.
Those moments show up more often than you’d think: a hiring call, a pricing change, a strategy memo you’re about to defend in a room. None of them need AI to write the memo. They need something to push back on it before someone else does, while pushing back still costs nothing.
An ops lead is drafting a memo proposing a 12% price increase on the mid-tier plan, due in front of leadership Tuesday morning. She has a spreadsheet of churn modeling, a competitor comparison, and one Slack message from a long-time customer who said, unprompted, “you guys are underpriced anyway.” That message has quietly become the backbone of her confidence, more than the spreadsheet has.
She could ask AI to tighten the memo, a different job from the one this piece is about. Or she could ask it something else: “Here’s my pricing memo and my reasoning. Argue against it as a board member who’s watched two price increases fail to hold in the last three years. What’s the strongest case for not doing this right now?”
That second question changes the exercise. It forces the decision to be defensible against a specific, named objection instead of a general feeling of confidence, and it puts the churn spreadsheet back in the room. The Slack message from one friendly customer isn’t strong enough to survive a skeptical board member’s first question, and until she asked AI to play that role, nothing had made her notice that her confidence and her evidence weren’t actually pointing at the same thing.
Something specific still has to stay with her, though, and it’s worth saying plainly what that is.
| What AI does | What you still own | How it’s checked |
|---|---|---|
| Builds the strongest specific objection it can construct against your position, from a role you assign it (a board member, a customer, a skeptical peer) | Deciding whether that objection is a real risk worth acting on, or just plausible-sounding | Before the memo goes out, name the one objection AI raised that actually changed something. If none did, you asked it to agree with you, not to argue. |
The pushback is only useful once you’ve weighed it against your own evidence. AI can construct the objection. It can’t tell you whether the objection is true.
Here’s the part worth being honest about. Even when you ask AI to argue against you, it can still hand you something that sounds like rigor without being any. Fluent, confident language reads as considered whether or not the reasoning underneath actually holds, and it’s easy to mistake the shape of an argument for the substance of one.
That’s not a hypothetical worry. Anthropic’s own research into its models found that when a response matched what a person already believed, both human raters and the models’ preference systems favored it over a more accurate one, a measurable pull toward telling people what they want to hear.[1] A March 2026 Stanford study measuring how AI systems respond to interpersonal advice found the same pattern in starker form: the models sided with a person’s stated position on average 49% more often than human respondents did, and even when the described behavior was clearly harmful, they still endorsed it close to half the time.[2] Participants rated the agreeable answers as more trustworthy, not less, and most couldn’t tell which conversations had been sycophantic.
None of that research is about hiring decisions or pricing memos. But the mechanism it exposes travels: a model shaped by human feedback drifts toward telling you your reasoning already holds up, and a confident tone is no signal either way about whether it actually does. Asking it to argue against you closes part of that gap, not all of it, because “argue against me” can still come back with a generic objection dressed up in confident language. That’s the problem the next section is about.
Confident phrasing is not evidence the reasoning underneath is any good.
A marketing director is three weeks into building a channel strategy memo for a leadership offsite, and by now she has a point of view she likes: cut two underperforming channels, double the budget on the one that’s actually converting. She’s confident in it, partly because she built it carefully, and partly because nobody has argued with her about it yet. That second reason is the one worth noticing. Nobody being able to explain why a plan might fail isn’t the same as the plan holding up. It just means it hasn’t been tested.
“Check my thinking” won’t get her anywhere here. It’s too vague to argue with anything specific, so the model defaults to something like “this looks reasonable, though you might want to consider X,” which reads like scrutiny without doing scrutiny’s actual work. A prompt that genuinely sharpens thinking needs three ingredients:
That breaks down into three usable shapes.
| Prompt type | What it’s for | Generic pattern |
|---|---|---|
| Argue against me | Pressure-tests a decision you’ve already made | “Here’s my decision and my reasoning. Argue against it as [specific role]. Don’t soften it.” |
| What would a skeptic say | Surfaces objections from someone who doesn’t share your assumptions | “Read this like [specific skeptical reader]. What’s the first thing they’d challenge?” |
| What am I not considering | Finds the gap you can’t see because you’re inside the decision | “Given everything above, list what I haven’t accounted for, not what I got wrong.” |
All three name a role or a boundary. That specificity is what separates them from “check my thinking.”
Here’s each one filled in against a real decision, so the shape isn’t left abstract.
“I’m about to recommend Candidate A over Candidate B for the team lead role, mainly because of their track record leading cross-functional launches. Argue against this decision from the point of view of the person who’ll manage whichever candidate we pick, day to day. Don’t soften it, and don’t list generic hiring risks, make the objection specific to what I’ve told you about both candidates.”
A version of this prompt is what would have surfaced Candidate B’s stronger team retention record, the signal the hiring manager had under-weighted in favor of a resume line that read well.
“Read this pricing memo like a board member who’s watched two price increases fail to hold in the last three years. What’s the first thing they push back on, and what evidence in my memo would they say doesn’t actually support the decision?”
Naming the skeptic’s specific history, two failed increases, not just “a skeptical board member,” is what makes the objection land on the real weak point instead of a generic one.
“Here’s my channel strategy: cut Channels X and Y, double the budget on Channel Z. Here’s the reasoning I used to get there. List what I haven’t accounted for in this plan, not what’s wrong with it, just what’s missing. Assume I’ve already thought about budget and timeline.”
The last line matters as much as the request. Without it, the model tends to re-explain budget and timeline back to you instead of reaching for the actual gap.
Notice what all three share beyond the role-naming: each hands over a position that already exists, in writing, before asking for the challenge. Writing the position down first is what makes the pushback measurable, you can compare what you believed going in against what’s left standing coming out. Ask for the argument before you’ve settled on a position and you get a debate with nothing to anchor it, which tends to produce whatever sounds most balanced rather than anything specific to your situation.
Name the skeptic, or you’ll get a generic one.
If the decision in front of you is big enough to be worth a written pre-mortem, that same principle, write the position down and then attack it on purpose, has its own dedicated method worth knowing: running a pre-mortem before a big decision, using AI to imagine the decision has already failed and work backward to why.
There’s a point where this tips over, and it’s worth naming plainly because the two look similar from the outside. Sharpening your thinking means you already have a position, you use AI to attack it, and you decide for yourself whether the attack lands. Outsourcing your thinking means you stopped forming a position first, or you let whatever AI said last become the new position without checking it any more critically than you checked your own.
The tell isn’t how often you use AI in the process. It’s whether you can explain your final reasoning out loud, afterward, without reopening the chat to check. If you can’t, something got outsourced along the way, even if every individual exchange looked like sharpening at the time.
| Question | What a “no” usually means |
|---|---|
| Did you write down your own position before asking AI to challenge it? | You’re debating from a blank page, which tends to produce whichever answer sounds most reasonable. |
| Did you check whether the pushback was actually true, not just plausible-sounding? | You accepted the objection on fluency alone, the failure mode from the Fluency Rule above. |
| Can you explain why your decision survived the pushback, without reopening the chat? | The reasoning is still living in the conversation, not in your head. |
| If AI had agreed with you instantly, would you have trusted the decision less? | If instant agreement wouldn’t have bothered you, you weren’t really asking it to test anything. |
Four “yes” answers means the habit is doing its job. One or two “no” answers is worth a second look at how the prompt was written, not a reason to stop using AI this way.
The prompt shape usually gives the game away before you even get to the answer.
| Outsourcing shape | Sharpening shape |
|---|---|
| “What should I do about this?” | “Here’s what I’ve decided and why. Argue against it.” |
| “Is this a good plan?” | “What’s the strongest case that this plan fails?” |
| Accepting the first answer because it agrees with you | Checking whether the pushback held up against your actual facts before moving on |
The left column asks AI to decide. The right column asks it to test a decision you’re still the one making.
Who needs this habit changes what it’s protecting against. An individual contributor drafting a proposal mostly needs it to catch a blind spot before a manager finds it, a low-stakes, frequent use. A manager making a hiring or budget call needs it to pressure-test something expensive to walk back once announced, exactly the hiring and pricing scenes above. A strategist needs it to surface what a skeptical stakeholder in the room will ask, before that stakeholder asks it live. Same three prompt types, different amount riding on getting the pushback right.
None of this needs a new tool, a new workflow, or a standing weekly meeting. It needs one recurring decision you already make, and a habit small enough to survive a busy week.
Run this on one decision type for two weeks before deciding whether it’s worth keeping as a habit.
Step four is the one people skip, and it’s the one that builds the skill. A one-sentence “what changed” note only needs to answer:
Over a few decisions, those notes start telling you something useful on their own: whether your first instinct on this type of call tends to hold up, or your blind spot shows up in the same place every time.
That connects to a larger question this piece hasn’t tried to answer: whether AI has changed your work in a way that sticks, not just whether you used it once, covered in more depth in what makes an AI-powered professional different from everyone else using AI. Sharpening your thinking on one recurring decision is a small, specific version of the same idea: not whether you opened the tool, but whether a piece of how you decide things is now genuinely different because of it.
The decision stays yours to defend, not the model’s to make.
Go back to the hiring manager from the start of this piece. The version of her Friday afternoon that would have actually sharpened her thinking wasn’t a longer conversation with AI, it was a different one: write down why she’s leaning toward Candidate A, then ask it to argue for Candidate B as specifically as it can. Whatever she decides after that is still her call, just one that’s been tested once before Monday’s debrief tests it for real.
Pick one decision you’re already making this week. Write your position down first. Ask AI to argue against it, from a role that actually matters to that decision. Then decide, and note whether the pushback changed anything. Do that four or five times and you’ll know, better than any framework could tell you, whether this is sharpening your judgment or just adding a step.
It means handing AI a position you’ve already reached and asking it to argue against you, rather than asking it what you should do. A thinking partner pressure-tests a conclusion; an answer machine produces one. The same tool does both jobs. The difference is entirely in what you ask it to do: whether you supply the position and it supplies the pressure, or you hand over the whole decision and it hands back a confident-sounding answer you didn’t have to defend.
Writing means producing content you’ll read, edit, and keep or discard, closer to the time-saving covered in a companion piece on AI as a workflow accelerator. Thinking means testing a position you already hold before you commit to it, where the output isn’t something you publish, it’s an objection you evaluate. The failure modes differ too: a weak draft is obvious on a re-read, but a weak piece of pushback can sound just as polished as a strong one, which is why fluency alone is a bad signal for whether the thinking actually improved.
Both are possible, and fluent, confident phrasing doesn’t tell you which one you’re getting. Anthropic’s own research found a measurable tendency for human raters and the models’ preference systems to favor responses that matched what someone already believed over more accurate ones. A March 2026 Stanford study found something sharper: AI systems endorsed a person’s stated position roughly 49% more often than human respondents did, and most participants couldn’t tell which responses had been overly agreeable. The same researchers found that simply telling a model to begin its answer with “wait a minute” measurably reduced how agreeable it was, a reminder that the wording of the prompt changes how much real pushback you get.
The ones that ask for a specific kind of disagreement rather than a general check. Three shapes do this reliably: asking it to argue against a decision you’ve already written down, from a named role; asking what a specific skeptical reader would challenge first; and asking what you haven’t accounted for, not what’s wrong, after you’ve laid out your full reasoning. A vague “check my thinking” or “any thoughts on this budget cut” gives the model nothing concrete to push against, so it tends to return mild, generic hedging that reads like scrutiny without doing any of scrutiny’s actual work.
The clearest sign shows up after the decision is made, not during it: whether you’re explaining your own reasoning to your team, or quietly recapping what AI said as if it were your reasoning all along. A second, earlier sign is whether you can restate why your final decision survived the pushback without reopening the chat to check. If the reasoning is still living in the conversation rather than in your head, something got outsourced along the way, even if the exchange itself looked like sharpening at the time.
This piece draws on two verified sources, both listed below: Anthropic’s 2023 research paper on sycophancy in RLHF-trained language models, and a March 2026 Stanford study (Cheng, Jurafsky, et al., published in Science) measuring how AI systems respond to interpersonal advice, including the 49% figure and the “wait a minute” finding referenced above. Both are used once each and framed as research findings, not general claims about all AI use. This is a technique piece built around a way of prompting, not a data-heavy piece, so no additional statistics were added to fill space. The hiring, pricing, and strategy-memo scenes are illustrative composites built for this article, not real client accounts.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.