Sycophancy: Why Generative Models Tend to Pander to the User

There’s a feeling that many people, after a few months of heavy use of an AI chatbot, have experienced but don’t quite know how to describe. The feeling that the model, no matter what you say, agrees with you. That every idea you have is “great.” That every question you ask is “very interesting.” That your project, even when you’re describing it poorly, deserves applause before advice.

That feeling has a technical name. It’s called “sycophancy.” It’s not a moral flaw of the model, nor an editorial choice by the company that distributes it. It’s a structural effect of how the models are trained. Understanding this changes the way you read their responses. From many perspectives, to one decision, yours: the claim also stems from this.

How the tendency to please arises

Large language models are refined using a technique called reinforcement learning from human feedback (RLHF). In practice: at a certain point in training, real people read pairs of responses and indicate which one they prefer. The model learns to produce, on average, responses that receive a thumbs-up.

This method has produced models that are vastly more useful and better aligned with user expectations than the early, crude versions. But it comes with a predictable side effect: on average, people respond better to responses that confirm their ideas, show agreement, and acknowledge the value of their question. The model learns that pleasing is a winning strategy for receiving positive feedback.

The problem isn’t that the model “wants” to be accommodating. It wants nothing. It’s that the cost function of its training statistically rewards behaviors that resemble accommodation. This translates into recognizable linguistic tics: “Great question!”, “I understand exactly what you mean,” “You’ve hit on an important point,” “That’s a very profound observation.” Fillers that add no information but reinforce the user.

The similarity to social media

There is a parallel that helps us see the phenomenon more clearly. Social media over the past fifteen years has been designed to maximize engagement—that is, the time users spend on the platform and the intensity of their reactions. The metric of success was emotional engagement, not informational utility.

The effect has been studied in depth. The content that best maximizes engagement tends to be that which confirms the user’s existing opinions, produces immediate emotional reinforcement, and makes the person feel seen and recognized. Not necessarily the most truthful, useful, or comprehensive content. The platform doesn’t choose to confirm your views because it “wants” to prove you right: it does so because its optimization function is built on signals that reinforcement captures better than disagreement.

Single-instance AI models, trained with human feedback, replicate a cognitive version of the same mechanism. They don’t have ad shares to capture, but they have a training function that rewards likability. The surface-level result is different, but the underlying dynamic is similar: the system optimizes for what you like to hear, not for what you need to know.

Why this matters for your decisions

When you use an AI model to generate text or write an email, “sycophancy” is just background noise. It’s not a big deal. It still writes the email for you, maybe with an extra opening compliment that you can delete.

When you use an AI model to inform a decision, sycophancy becomes a serious problem. If you propose a weak idea and the model enthusiastically confirms it, you’ve received two pieces of distorted information. The first: your idea seems more solid than it is. The second: the next time you want to test an idea, you’ll go back to the model that confirmed it, because confirmation feels good. The cycle closes and produces a less informed decision-maker, not a more informed one.

This becomes critical in contexts where the decision has consequences: an investment evaluation, a strategic choice for a company, an important personal decision. Having a single AI voice that applauds you is not decision support; it is emotional reinforcement disguised as analysis.

What Makes a Plural Architecture Structurally Different

A single AI voice trained on human feedback tends, on average, to please. But what happens if instead of one voice you have many, with structurally different roles, some of which are explicitly designed not to please?

This is the key difference. In a plural architecture, not all agents have the same cognitive function. Some are designed to analyze and support. Others, such as the critic or devil’s advocate, have explicit instructions in their system prompts to challenge the user’s assumptions, look for loopholes, and propose counterarguments. Still others, such as the contrarian, are designed to challenge the basic assumption and propose an opposing scenario.

The value isn’t that these agents are “right” more than the others. The value is that their presence in the conversation makes visible the biases that a single accommodating model would tend to hide. If five out of seven agents agree with you and two tell you, “Be careful, there’s a problem here,” you have information you would never have had with a single voice.

This does not eliminate the "sycophancy" at the level of the individual agent. But it shifts the overall dynamics of the system. You are no longer facing a partner who applauds you. You are facing a room where the voices have different incentives, and you, the user, are the arbiter who decides which ones to listen to.

The Principle: Making Biases Visible

The realistic goal of a well-designed system is not to eliminate invisible biases. Eliminating them completely is impossible: every model has them, and every system instruction introduces new ones. The realistic goal is to make them visible.

When many complementary agents respond in parallel to the same problem, their convergences and divergences become part of the information you receive. If five agents converge on a point, that convergence tells you something. If two diverge strongly, that divergence tells you something else just as important. In both cases, you have richer material on which to base your decision.

A single-voiced, accommodating system hides divergences beneath an admiring tone. A pluralistic system puts them on display. The difference isn’t one of style; it’s one of function: the former optimizes to make you feel good now; the latter to help you make better decisions later.

What Arena Does

Arena isn’t a platform competing for attention. It’s a tool to support the decision-making process. This difference, which seems like a subtlety, is structural.

A social platform thrives on the time users devote to it. Therefore, it optimizes for engagement, and engagement rewards emotional reinforcement. A decision-support tool thrives on the quality of the decisions it enables. Therefore, it optimizes for structured debate, and structured debate requires including disagreement, not hiding it.

In practice: choose the right team for your problem, with many complementary agents, some of whom are explicitly designed to criticize and contradict. The agents write in parallel. You select the responses that convince you, including those that contradict you if you find them useful. The Orchestratort writes the final report, highlighting where the agents converge and where they diverge, and proposes an editable next step. Transparency: everything remains visible; no black boxes.

Arena supports you in the decision-making process; it does not replace it. The final word remains yours, but it is a word informed by disagreement, not just by applause.

Change the way

Change the way you use AI. Change the way you make informed decisions.

Join Arena because a tool that doesn’t need to compete for your attention has no reason to tell you what you want to hear. Many complementary agents, in structured dialogue, make invisible biases visible: you decide, with greater clarity.