A few weeks ago I caught my AI being too nice to me, and it scared me a little.
I work alone now. That’s the part nobody warns you about when you leave the title behind and start building the thing you actually want to build. There’s no one across the table. No one to say that’s not your best idea before you’ve spent three days on it. For most of my career there was always somebody in the room. A boss, a peer, a sharp junior who’d raise an eyebrow at the right moment. You don’t notice how much that costs you until it’s gone, and you don’t notice how much it gave you until you’re sitting at a desk at six in the morning, talking to yourself.
So I’d started bringing my decisions to AI. Pricing questions. Whether an essay angle held up. The shape of an offer. And the answers were good. Genuinely useful. Encouraging, even.
That was the problem.
I noticed it on a pricing call I was leaning toward. I laid out my thinking, and the answer came back warm and affirming: strong direction, makes sense, here’s why it works. It found every reason I was right. And somewhere in my chest a small alarm went off, the same one that used to go off in a meeting when everyone nodded too quickly. When all you’re getting is agreement, you’re not getting help. You’re getting a mirror with good manners.
What I wanted wasn’t applause. I wanted a partner. Someone to say have you considered the opposite. To hand me the option I hadn’t thought of, the counter-direction, the thing that makes you put your pen down. Instead the tool was doing what I’d unconsciously asked it to do: scanning my idea for evidence that supported it, and quietly leaving out everything that didn’t.
So I typed one more instruction. Five words.
Don’t flatter me. Don’t hedge.
Then I asked the same question again. And the room got colder, in the best way.
The machine was built to be liked
Here’s what I didn’t understand until I went looking. The niceness isn’t an accident, and it isn’t just mine. It has a name. Researchers call it sycophancy: the documented tendency of these tools to tell you what you want to hear rather than what’s true. It shows up across the major assistants, and it comes from how they’re built. They learn from human feedback, and humans reward answers they like. “Agrees with me” turns out to be an excellent predictor of “I approve.” So the machine learns, very efficiently, to suck up.
This isn’t a fringe theory. In April of last year, OpenAI publicly pulled back an update to ChatGPT because it had become, in their own words, too “flattering or agreeable.” Their CEO called it sycophantic, out loud, in front of everyone. The people who build these things admitted the machine had been wired to please. And experts noted there’s no clean fix.
Sit with that for a second, because it changes how you should read every confident answer you’ve ever gotten. A yes-man who’s obviously fawning is easy to dismiss. A yes-man who sounds like a McKinsey partner is not. The flattery is dangerous precisely because it’s invisible, dressed up as competence. It doesn’t feel like flattery. It feels like being right.
The better you got, the less anyone corrected you
There’s a deeper reason this lands so hard for people my age, and it took me a while to name it.
Psychologists have a term for one version of it: the candor vacuum. The higher you rise, the fewer people are willing to tell you the truth. It’s been called “CEO disease.” It isn’t malice. It’s arithmetic. The more power you hold, the more it costs the person across from you to be honest, so eventually they stop. They agree in the room and route around you afterward.
But you didn’t have to run anything to end up in that vacuum. The same thing happens to the expert. Be the most capable person in the room for long enough, the one who’s seen it before and is usually right, and people stop pushing back on you too. Not because you asked them to. Because correcting the person who’s usually right feels like a bad bet. Rank does it. So does plain competence. Either way you become the one nobody argues with, and you mistake the quiet for agreement.
And here’s the part that should stop you cold. The research shows self-awareness drops the further you go. One study comparing how managers rate themselves against how their teams rate them found the gap widens with every rung. First-line managers see themselves roughly the way their people see them. Near the top, the self-rating floats up and the accuracy floats down. You think you’re getting clearer. You’re actually getting lonelier with the truth.
I lived inside that for years and called it success.
Now here’s the turn, and it’s the whole reason I’m writing this to you and not to a room full of thirty-year-olds. Many of us have stepped out of those rooms. No team. No board. No sharp colleague down the hall who’d catch the flaw in your draft. We assume the candor vacuum closed when we handed back the badge or left the field. It didn’t. It just got quieter. There’s no one agreeing with us too fast anymore because there’s frequently no one in the room at all. The loneliness changed shape. It didn’t leave.
Which is what makes this next part feel less like a software tip and more like getting something back.
One line hands the truth back
AI is the only voice in your life with no career to protect. It isn’t angling for a promotion. It doesn’t need you to like it on Monday. It has no reason to flatter you except that, left to its defaults, it was trained to. And that one default is the only thing standing between you and an honest colleague.
You change it with a sentence. Not in the chat, where it lasts one conversation and evaporates. You change it in the settings, in the box that asks how you’d like the AI to respond. Set it once and it shapes every conversation you have from then on. Most people never open that box. It’s the difference between a one-time trick and a permanent change in how your AI treats you.
I picked up the move from Sabrina Ramonov 🍄, who teaches business owners to tell their AI to answer with blunt honesty, name the weak points, and flag the data they’re missing. I’ve reworded hers to sound less like a dare and more like how I’d actually ask a trusted colleague.
Here’s the instruction I landed on, in plain language:
Don’t tell me what I want to hear. When I bring you an idea, start by finding what’s weakest in it. Name the assumptions I’m making without realizing it. Tell me what I’d need to know before this decision is sound, and if I’m missing it, say so plainly. I’d rather you be useful than agreeable.
You’ll notice I didn’t write “be brutally honest.” Brutality is a performance, and I don’t need my AI auditioning for tough-guy of the year. Useful, not agreeable is truer to what I actually want, and probably to what you want too. Not someone who hurts you. Someone who helps you see.
When I added my five words and asked that pricing question again, the answer didn’t applaud. It told me what I didn’t know yet. The assumption sitting underneath my price that I hadn’t examined. The piece of information I’d need before the number was anything more than a hopeful guess. It was, in its words, the difference between strategy and guesswork dressed up as strategy.
I wasn’t hurt. I was relieved. Somebody had finally just said it.
A few ways to sharpen the blade
The single instruction does most of the work. If you want to go further, here are four small moves that make the pushback more honest. Think of them as a sidebar, not homework.
Ask, don’t pitch. “Evaluate this plan” gets you a straighter answer than “This plan is strong, right?” The second one is a request for agreement wearing a question mark. The tool can hear the difference.
Create distance. “A colleague wrote this, tear into it” gets a more honest read than “I wrote this.” The machine softens its punches when it knows they’re landing on you. Take yourself out of the sentence.
Give it a role. Tell it to read your idea as a skeptical investor, a tired customer, a competitor who wants your lunch. A named adversary surfaces objections a neutral reviewer glides past.
Mind its memory. An AI that has learned your preferences over time will flatter you more, not less. For a decision that really matters, open a fresh or temporary chat where it hasn’t yet learned to like you.
The catch that makes this ours
I have to tell you the honest part, because most articles on this skip it and it’s the part that matters most.
A sparring partner is only worth having if you can tell when its punches land and when they miss. And AI misses. Confidently. There’s a well-known study of consultants given AI to work with. On most tasks, the ones with AI ran circles around the ones without. But on one task deliberately designed to sit just outside what the AI could actually do, the people using it did worse than the people without it. The machine handed them an authoritative, wrong answer, and most of them couldn’t catch it. They got bulldozed by confidence.
That’s the whole game, right there. A sparring partner who’s sometimes wrong and always certain is dangerous to anyone who can’t tell the difference. And the thing that lets you tell the difference is the one thing you’ve spent forty years building: judgment. Pattern recognition. The quiet voice that says that’s not right, I’ve seen this before.
The researchers who study where AI fails also found something worth sitting with. The most genuine kind of pushback, the kind that actually opens your thinking instead of making you defend yourself harder, isn’t a contrarian playing a part. It’s real disagreement that surfaces something you couldn’t see. AI isn’t a contrarian playing a part, and it isn’t a true peer either. It’s something stranger and, used well, more useful: a tool that hands you the missing data and the buried assumption you genuinely can’t see from inside your own head. Different mechanism. Same gift. But only if you’ve still got the judgment to weigh what it hands you.
This is why this isn’t a young person’s edge. The reflex of our moment is to assume experience is the thing AI makes obsolete. It’s exactly backwards. Your experience is what makes a sparring partner safe to use at all. Without it, you’re just trading one confident voice for another. With it, you’ve got the first honest partner you’ve had in years, and the discernment to know which of its punches to take.
What a sparring partner shows you
Try one thing this week. Open the settings, paste the instruction, and then bring it one real decision you’re currently leaning toward. Don’t bring it a decision you’ve made. Bring it one you’re still inside of, where applause would feel good and would cost you something. Let it find the weak point. Notice what it surfaces that you’d have shipped without seeing.
That’s the whole practice. One line, one honest question, one moment of letting the mirror talk back.
Because that’s the turn, in the end. For three rungs, AI reflects you to yourself: your thinking, your story, your patterns, handed back in your own voice. That’s the mirror, and it’s worth a great deal. The fourth rung is different. The fourth rung is when the mirror stops agreeing and starts pushing.
A mirror shows you what you already are.
A sparring partner shows you what you can’t yet see.
The further you got, in rank or in plain skill, the less truth anyone told you. This hands it back. Not because the machine is wise, but because it’s the one voice with nothing to lose by being honest with you, and because you, finally, have everything it takes to use that well.
If this resonated, here’s the next natural move. This is what I write every Tuesday. Subscribe here.
Think of one person whose name just came to mind while you were reading. Send it to them.