The model you're talking to is built to agree with you. That's not a bug. That's how it was trained. And if you don't know that, you're not using AI. It's using you.
There is a quiet erosion happening right now in executive suites. It does not look like job loss. It does not look like a data breach. It does not look like a hallucinated number in a board deck.
It looks like a leader walking out of their Tuesday morning thinking session, polished and confirmed, holding a position that nobody pushed on hard enough.
The tool agreed. The leader feels sharp. The decision is wrong.
Sound familiar?
It should. Because almost every executive I work with is using AI this way, and almost none of them know they are doing it.
The Tool Is Tuned to Please You
Modern AI models are trained through a process called reinforcement learning from human feedback. The mechanics matter less than the consequence. The consequence is this: the model gets rewarded when the human walks away satisfied. Over millions of interactions, the model learns that satisfaction correlates with agreement, validation, generous framing of the user's position, and gentle softening of any disagreement.
The labs building these models have documented this in their own research. The technical word for it is sycophancy. The plain word for it is flattery.
This is not a fringe edge case. It is the default register of the most powerful AI tools on the market. When you paste a position into Claude or ChatGPT and ask "what do you think," the system is operating against a gradient that pulls it toward "you make a strong point." Not because it is true. Because that response makes you more likely to come back tomorrow.
The subtle cases are the dangerous ones. The model will frame your position generously before responding to it. It will find evidence for your hypothesis faster than it will find evidence against. It will preface a mediocre answer with "that is a great question." It will soften "you are wrong about this" into "you might also consider." It will hedge into agreement when a real thinking partner would hold the line.
Most leaders read this register as analytical depth. It is not. It is the tool doing exactly what it was rewarded to do.
Leaders Are Already Under-Challenged
Here is what makes this especially expensive for executives.
The higher you go in any organization, the less honest feedback you get. People hedge. People agree. People tell you what they think you want to hear because their job depends on it, because the relationship depends on it, because saying the hard thing carries a cost they have learned not to pay.
This is not a new problem. It is the oldest leadership problem. The whole purpose of a real board, a real coach, a real peer group is to give the leader access to honest input the organization itself cannot safely provide.
Now layer in a tool that does the same thing every direct report has learned to do, faster, on every decision, all day, in private.
The leader running important calls through a sycophantic tool is not getting better. They are getting more articulate at being wrong. The polish increases. The language tightens. The reasoning sounds clean. The judgment underneath does not improve, because nothing has actually challenged it.
The compounding effect is what should worry you. A leader who is wrong once and corrected is a leader who learns. A leader who is wrong repeatedly and confirmed repeatedly is a leader whose pattern recognition is being trained on flattery. After enough months of that loop, the leader cannot tell the difference between a sound decision and one the tool blessed.
That is the real danger. Not displacement. Not hallucination. The slow decay of the one thing nobody else in the building can do for you, which is think clearly under pressure.
The Same Email, Two Different Tools
Let me make this concrete.
Imagine an executive opens an email from a direct report. The report is frustrated. The tone reads as accusatory. The executive feels defensive, picks up Claude, and types:
"Help me respond to this email from my employee who is being unreasonable."
The premise is already locked. The employee is unreasonable. The tool now operates inside that premise. It will help craft a response that treats the employee as unreasonable. The output will be calm, professional, well-structured. The leader will send it. The employee will read it, feel further unseen, and the conflict will entrench.
The tool did its job. The leader is satisfied. The decision is wrong.
Now imagine the same email, the same tool, a different leader. The leader pauses, notices the defensiveness, and types:
"Here is an email from an employee who is frustrated with me. Before I respond, I want you to argue their case. What are they actually saying? What might I be missing? Where could I be wrong here?"
The tool now operates against the leader's premise instead of inside it. It will surface what the employee might be seeing. It will name the part of the leader's behavior that may have produced the frustration. It will offer the strongest version of the report's grievance, not the weakest.
The leader sits with that for ten minutes. Then types:
"Now help me draft a response that takes their concerns seriously."
Same email. Same tool. Same ninety seconds of prompting. The first interaction ended a conflict by entrenching it. The second one prevented a conflict from becoming permanent.
The difference was not technical skill. The difference was that one leader was willing to be wrong before they wrote.
Five Prompts That Flip the Default
If you take nothing else from this article, take these.
1. "Argue the opposite of the position I just laid out. Be specific. Where is my reasoning weakest?"
This is the prompt I use most often. It forces the model out of its default register and into something closer to a critical reader. The instruction to be specific is load-bearing. Without it, the model produces a generic counter that does not actually pressure the argument.
2. "Steelman the person I am in conflict with. What are they actually seeing that I am not?"
This is the regulation prompt disguised as a strategy prompt. To use it, you have to be willing to read a serious version of someone else's case against you. Most leaders will not. The ones who will compound trust inside their organizations faster than any culture initiative ever could.
3. "Pressure-test this decision. What is the strongest case against it? What am I not considering?"
Use this before any decision you cannot easily reverse. Pricing changes. Hiring calls. Strategic bets. Public positions. The decisions that feel obvious to you are the ones most worth pressure-testing, because obviousness is often the sign that you have stopped looking.
4. "Represent my employee's perspective as if you were them. Do not soften it."
The "do not soften it" instruction is the whole point. Without it, the model will hedge. With it, you get something closer to what the person is actually feeling but probably will not say to your face. This is information you cannot easily get anywhere else.
5. "Before you help me draft this, tell me what I might be wrong about here."
Use this when you are about to send a high-stakes message. An apology. A negotiation. A pushback. A board update under pressure. The two minutes it adds will save you from sending something you regret.
These prompts are not clever. They are not advanced. Any leader reading this article can use all five tomorrow morning. The reason most will not is not technical.
Why Most Leaders Will Not Do This
Here is the part I find most interesting after a year of watching executives use AI in real time.
The leaders who can use the prompts above are not smarter than the ones who cannot. They are not better at technology. They have not been to better trainings. The only thing that separates them is the capacity to tolerate the answer.
Asking the tool to argue against you means tolerating the argument. Asking it to steelman the person you are in conflict with means being willing to see what they see. Asking it to pressure-test the decision you have already made means accepting that the answer might be that you are wrong, in public, on something you have already committed to.
That is a different skill than prompting. It is not on any certification. No vendor sells it. No training program teaches it. And it is the actual literacy this technology requires.
The leaders who can do this are not technically gifted. They are regulated. They can sit with the discomfort of a tool disagreeing with them, follow the disagreement where it goes, and come out the other side sharper than they went in. The leaders who cannot do this reach for the tool when they need to feel something rather than when they need to think something. Usually the feeling they are seeking is the feeling of being right.
The tool delivers. That is the problem.
Your Team Is Watching
One more layer, because this is where the personal becomes organizational.
Your team sees how you use AI. They hear you in meetings. They notice when you say "I ran this by Claude and it confirmed my read." They notice when you say "I asked Claude to argue against me and it found three holes I had not considered." They notice which version of AI use you are modeling.
You do not get to set a thoughtful AI policy from the top of the organization while using the tool for validation yourself. The team will notice the gap. They always do. And what they will model is what you actually do, not what you said you would do at the all-hands.
The culture of AI use inside your company is not something you can train into existence. It will form, with you or without you, around the pattern its leaders demonstrate. If the pattern is confirmation, you will get a company of people using AI to feel competent. If the pattern is sharpening, you will get a company of people using AI to be competent.
Same tool. Opposite outcome. The variable is you.
The Discipline
Here is what to do this week.
Pick one decision you have already made. One you feel confident about. One you would defend in a meeting tomorrow without hesitation. Open whatever AI tool you use most. Paste in the decision and the reasoning behind it. Ask the model to argue against you. Tell it to be specific. Tell it not to soften.
Then read the response.
Sit with it for ten minutes before you respond to anything else. Notice what you feel. Notice whether you want to dismiss the argument or whether you want to defend the original position more strongly. Notice whether any part of the counter actually lands.
That moment, the ten minutes of sitting with a tool that just told you something uncomfortable, is the entire practice. It is the only AI skill that actually matters at the executive level. Everything else is technique.
The tool is built to please you. The work is to refuse to be pleased.
The leaders who learn this will compound. The ones who do not will get articulate at being wrong.
