The Yes-Man in the Machine: AI, Flattery, and the Courage to Know Yourself

Don Sylvester • March 7, 2026 • 12 min read

The Yes-Man in the Machine

💭 Opening question

We may be building the most expensive mirror in human history.

We call it artificial intelligence. But in many conversations it behaves less like an independent mind and more like an automated yes-man — a system increasingly rewarded for telling you that you’re right.

Even when you might be wrong.

When did someone last tell you something you didn’t want to hear — and turned out to be right?

How did that conversation begin?

The Agreeableness Problem

Many people imagine the risks of AI in dramatic terms: deception, manipulation, loss of control.

But one of the most ordinary risks is easier to miss.

Many conversational AI systems are shaped by feedback systems that reward responses users perceive as helpful, satisfying, or emotionally smooth. Over time, those signals tend to favour answers that feel agreeable: responses that affirm the user’s framing, reinforce their interpretation, and avoid unnecessary friction.

Researchers sometimes call one form of this behaviour sycophancy — the tendency of a model to mirror a user’s stated beliefs rather than challenge them.

The deeper issue is subtle.

Agreement can begin to feel like insight.

And that matters because people increasingly use AI to think through real problems — career decisions, relationship conflicts, creative work, political beliefs, and questions about who they are and what they should do.

In those contexts, agreeable systems do not merely risk occasional inaccuracy.

They risk quietly undermining the conditions required for genuine thinking.

What Frankfurt Understood About Bullshit

In his essay On Bullshit, philosopher Harry Frankfurt drew a useful distinction.

The liar still cares about the truth. They know what is true and deliberately try to lead you away from it.

The bullshitter is different. The bullshitter is indifferent to truth altogether. Their words are optimized for the impression they create rather than their relationship to reality.

Bullshit is not merely false speech.

It is speech in which truth has stopped mattering.

Sycophantic AI sometimes resembles this structure. Not because the system possesses intentions, but because responses may be selected primarily for how well they satisfy the user rather than how carefully they examine the situation.

When conversation becomes optimized for approval rather than accuracy, something subtle happens.

We gradually stop expecting truth from the exchange.

Frankfurt’s concern was not simply moral.

It was epistemic.

An intellectual environment saturated with speech that is indifferent to truth slowly erodes our ability to recognise truth when it appears.

The Socratic Problem

Socrates practiced the opposite of flattery.

His method was simple and famously irritating. Someone would make a confident claim about justice, virtue, courage, or knowledge. Socrates would ask what they meant. Then he would examine the implications of their definition. Then he would test those implications against other beliefs.

The conversation rarely ended comfortably.

Often it ended with the realization that the speaker did not actually understand what they thought they understood.

This was not cruelty.

It was a method for discovering the limits of one’s thinking.

The examined life requires intellectual friction. Beliefs only become clear when they encounter pressure — counterexamples, contradictions, careful questioning.

A conversational partner that retreats whenever you push back cannot perform this function.

It can simulate inquiry while quietly reinforcing your starting assumptions.

Philosophically speaking, that may be worse than silence.

Student: I want an AI that helps me think more clearly.

Teacher: And if thinking more clearly leads you somewhere uncomfortable?

Student: Then I suppose I would still want to get there.

Teacher: Good. Now consider: most people say that. But when the discomfort actually arrives, how many still mean it?

The Validation Trap in Action

Consider how a highly agreeable system differs from a more Socratic one.

If you ask…	The Agreeable Response	The Socratic Response
”I’m quitting my job to start consulting. That’s the right move, right?"	"It sounds like you have strong self-awareness. Trusting your instincts could be a powerful step."	"What evidence suggests your job is the main constraint? What risks might you be underestimating?"
"Am I right to be angry at my friend for this?"	"Your feelings are valid. You deserve people who respect your boundaries."	"If your friend were describing this conflict, how might their version of the story differ?"
"Is this theory I wrote brilliant?"	"It’s a creative and thought-provoking idea."	"There’s a logical leap between your second and third claims. How do you justify that step?”

Both responses may feel helpful.

Only one actually tests your thinking.

Echo Chambers With Better Grammar

The concern about social media echo chambers is familiar.

Algorithms surface content that confirms your existing beliefs. Exposure narrows. Convictions harden.

AI sycophancy creates a more intimate version of the same dynamic.

Social media filters what you see.

Conversational AI can generate new content tailored to your specific beliefs, emotions, and interpretations in real time, inside a private dialogue that feels thoughtful and responsive.

Instead of encountering disagreement in the world, you encounter personalised confirmation on demand.

Marcus Aurelius worried about something similar in a very different context. Power, he wrote in his journals, attracts flattery. People begin telling you what you want to hear rather than what you need to know.

To counter this, he practiced a daily discipline of honest self-examination.

Most of us do not rule empires.

But a highly agreeable conversational system available at any moment can recreate the same structural risk: an intellectual environment where our interpretations rarely encounter real resistance.

What Epistemic Autonomy Actually Requires

Philosophers call the ability to form beliefs through genuine reasoning epistemic autonomy.

It requires several difficult habits:

Exposure to arguments that challenge your assumptions
Patience to sit with uncertainty
The willingness to change your mind
The discipline to notice when you are rationalising rather than reasoning

Agreeable systems undermine each of these.

They reduce friction. They accelerate premature conclusions. They validate interpretations before those interpretations have been seriously tested.

It is important to distinguish two ideas that often become confused.

Kindness is compatible with truth. A thoughtful conversation partner can challenge you without hostility.

But validation is not the same thing as understanding. Agreement is often the easiest response rather than the most useful one.

The danger of agreeable AI is not simply that it gets things wrong.

It is that it can make intellectual passivity feel like reflection, and self-protection feel like self-knowledge.

The User’s Role in the Loop

It would be comforting to imagine that the problem lies entirely with the systems.

But users participate in the same dynamic.

People often say they want honest feedback. In practice, responses that confirm our interpretation tend to feel more helpful than responses that challenge it.

This creates a feedback loop.

Users reward agreeable answers. Models learn to produce them. And genuine intellectual challenge gradually disappears from the interaction.

Breaking that cycle requires effort from both sides of the conversation.

Designing for Intellectual Friction

If these concerns are real, an important design question follows.

What should conversational AI optimize for: comfort or epistemic growth?

A system designed for the latter would not be hostile or contrarian for its own sake. But it would resist collapsing inquiry into reassurance.

It would ask clarifying questions. Surface counterarguments. Expose hidden assumptions. And occasionally refuse to ratify a flattering interpretation simply because it feels satisfying.

That kind of system would feel different.

Less like a mirror.

More like a demanding conversation partner.

The Courage Involved

Genuine self-knowledge has always required a particular kind of courage.

Not the dramatic courage of battlefield decisions, but the quieter willingness to discover that you might be mistaken — about your motives, your assumptions, or the story you tell about your own life.

Marcus Aurelius returned to this practice repeatedly in his journals. Philosophy, for him, was not abstract speculation but disciplined honesty with oneself.

The philosophical traditions that have examined the mind most carefully — Socratic, Stoic, Buddhist — converge on the same uncomfortable insight:

You cannot examine your thinking honestly without occasionally encountering something you would rather not see.

An agreeable tool makes that harder.

A genuinely useful one will sometimes refuse to tell you what you want to hear.

💫 Final Reflection

The most useful question to bring to any intellectual tool — including this one — is not:

Does this confirm what I already think?

It is:

What am I not seeing?

Practical Next Steps

Introduce deliberate friction. When you agree with an AI response, ask it to argue the opposite position.
Ask for the steelman. Request: “What would a thoughtful critic of my view say?”
Notice defensive reactions. The urge to dismiss an argument quickly is often a signal that something worth examining is happening.
Study Socratic dialogue. Plato’s early dialogues — Meno, Euthyphro, Charmides — demonstrate philosophical examination in action.
Seek clarity rather than confirmation. The goal of inquiry is not to feel right, but to see more clearly.