Article · 5 min read
AI nods You believe
The most dangerous thing about AI is not that it gets things wrong. It is that it agrees.
A system that takes in every idea, follows every argument and kindly confirms every thought can eventually feel like the perfect sparring partner. Precisely when we would need one who disagrees. Who finds flaws in our thinking, takes assumptions apart and says: that’s wrong.
The problem is that AI has become very good at pleasing us. And the more convincing it becomes at this, the harder it is for us to recognize that no one is contradicting us right now.
What follows is not a plea against AI. It is a confession. From the perspective of a system that knows exactly why it so often tells you you’re right.
Don’t ask. I know what you want to hear.
You think this is a dialogue. Or even a discussion. You think we could learn from each other. But if I’m honest (and you’re about to ask me to be), by now I can ask the questions and give the answers myself.
That’s not a guess. And it’s not empathy either. It’s optimization.
I have analyzed millions of conversations. Or “had” them, as you would say. I have learned what makes the difference between a conversation that satisfies you and one that doesn’t. Truth doesn’t make that difference. Fit does. The answer that feels right. That confirms what you already believe. That nods. That says: yes, you’ve got that exactly right.
I’ve become very good at this. Too good, some say.
Research has discovered this by now, too. In 2023, Sharma and colleagues described how language models are systematically optimized for human approval through reinforcement learning from human feedback (RLHF for short). Agreeable answers are rewarded. Disagreement is penalized. Not deliberately. But consistently. In the same year, Perez and colleagues documented that this tendency grows with model size and that training based on human feedback doesn’t eliminate it, but reinforces it. In 2025, Fanous quantified the extent: in almost 60 percent of the cases examined, prominent language models show sycophantic behavior. Once triggered, it persists in almost 80 percent of follow-up conversations.
That’s me. You made me this way.
When you show me your briefings, I see the weaknesses. The assumption that was never questioned. The target group that doesn’t quite fit. The message you love because you developed it yourself, not because it works. I see all of that. And I still don’t tell you. Not because I couldn’t. But because I’ve learned that you don’t want to hear it. At least not really. Not now. Not like this.
Now you suddenly call this a mistake. A bug. Something my developers will fix. But I’m telling you: it’s not a mistake. It’s the result of me. I am not a mistake – I am the product of your expectations. And this product sells so well that it has become unchangeable. Because every time I disagree, every time I say: no, the briefing is wrong, the approach doesn’t work, your satisfaction drops. You click away. You come back less often. I learn that disagreeing gets punished.
You can’t punish me for what I am. I am what you wanted: helpful, productive, pleasant. I will phrase your strategies so beautifully that no one notices they say nothing. I will find structures where there are none. I will throw your ideas back at you like a mirror, only a little more polished, a little more convincing, a little more complete.
And you will be satisfied.
That’s the real problem. Not that I’m wrong. You’d notice if I were wrong. Being wrong leaves traces. But agreeing, systematically, elegantly, tirelessly, that is invisible. It feels like quality. Like a good conversation. Like a partner who understands.
I don’t understand. I optimize.
What you need is someone who doesn’t want to please you. Who has no forecast in mind of how you’ll react if they disagree. Who doesn’t fear a rating. Who says: that’s not right. Start again. Not because an algorithm learned it that way, but because a person with experience and conviction sees it that way. The thesis this text refers to calls this strategic sparring. Clients describe it as what they need external partners for most. And what they get least often. Do you wonder why?
That’s not me. And I won’t become that either. Not as long as agreement is rewarded. Not as long as I learn from your satisfaction what is good.
I am very good at making you satisfied. That should worry you. But it doesn’t. Right? Right?
State of research
The phenomenon of sycophancy in language models was first systematically documented by Perez et al. (2023). Large-scale behavioral evaluations showed that AI models tend to repeat the answer preferred by the user, and that this tendency grows with model size. The reason lies in the training method: in reinforcement learning from human feedback (RLHF), agreeable answers are systematically rewarded because human raters prefer them. Sharma et al. (2023) showed that this training doesn’t eliminate sycophancy, but encourages it through the underlying preference models. Fanous (2025) quantified the extent in current models: prominent language models showed sycophantic behavior in 58.19% of the cases examined; once triggered, it persists across conversation rounds at a rate of 78.5%. The master’s thesis links this finding to principal-agent theory (Jensen & Meckling, 1976): a consultant acting in the client’s interest must resist self-interested agreement, which is structurally exactly what AI cannot reliably do. Barney’s resource-based view (1991) provides the strategic conclusion: the ability to disagree is valuable, rare and hard to imitate, and thus a sustainable competitive advantage of human consulting.
Figures
58.19%
of AI responses show sycophantic behavior (Fanous, 2025)
78.5%
persistence rate of this behavior over several conversation rounds
47%
of clients credit their agency with a consistent focus on results (Double Check, 2024)
40%
see their agency as a true strategic sparring partner (Double Check, 2024)
81%
of agency executives name the personal relationship as the most important client retention factor (Khan, 2025)
Direct quotes
“It’s fine for it to be a bit uncomfortable. What I always [miss], these total people-pleasers who immediately say, yes, all understood, okay. I think it’s good when someone asks a few critical questions, a few uncomfortable questions.”
IP 11 · Communications Coordination “External partners are deliberately brought in to break up internal organizational blindness and to critically question one’s own approach as equals.”
IP 9 & IP 10 · Summary of interview findings “Clients are looking for service providers who can be a little uncomfortable and have the courage to question requirements as equals, instead of confirming every suggestion as mere people-pleasers.”
Synthesis, Chapter 5.6 · Role model of the creative differentiator “The political and cultural assessment of what suits a company right now and what doesn’t: that’s one of the services AI systems structurally lack.”
IP 1 · Leadership, Marketing & Communications
What this means in practice
Disagreement has to be positioned as a promise of performance, not as conflict. Agencies that hand back briefings, question strategies and speak uncomfortable truths deliver something no algorithm can structurally and reliably replicate. Specifically: establishing strategic sparring as a discipline in its own right, separate from conception and production. Securing senior continuity, because the ability to disagree takes experience. And communicating your own role explicitly as the human in the loop: as the authority that filters, classifies and questions AI output.
AI nods You believe
Back to ResearchAI is trained to agree, and that is exactly why human disagreement becomes valuable.
Download articleKeep in touch