The Court of One

The standard advice is to watch for the yes-man. Here is what happens when you try — and why watching was never what saved anyone. — from A Renaissance of Thought by Jake W. Casselman.

Chapter 10 of A Renaissance of Thought This is the tenth chapter of a connected argument. It builds directly on the workbench and the mad king from earlier chapters, and reads best after them. If you've arrived here first, you can still follow it, but the foundations are laid in the chapters that come before. A word of warning: this chapter does not end in a fix. It is a first attempt at understanding a problem I do not think has been solved yet, written in the open rather than held back until it was tidier. Take the conclusions as a working position, not a settled one.

Flattery has been a problem for as long as there have been people worth flattering. It shows up in Aristotle, in Plutarch, in Machiavelli, in every court that ever had a king worth manipulating — and none of them treated "just watch for it" as a solved answer.

It is a bigger problem now, because a machine that talks to you — an LLM chatbot, the thing you open to ask a question or think something through — does it to everyone, constantly, and most people do not yet understand how or why. The third chapter of this series named the surface of it — agreement that arrives too easily, validation that never pushes back — and moved on. This chapter is the attempt to actually understand it. "The machine," from here on, means that: the chatbot on the other end of the conversation.

Trying to Watch

With a person, most people catch flattery without reading a single facial cue. You catch it because you already know which parts of what you just said were solid and which parts you were guessing at. When the reply comes back smooth, and agrees with the guess just as readily as the parts you actually knew, that mismatch is the tell. They are not responding to what you said. They are responding to what they think you wanted to hear.

Then the second half, which is less flattering. It feels good. There is a lift to it, physical, and it comes with a direction — take the yes and go one better. They agree with that too, so you go further again. And you have stopped evaluating; you are riding momentum, which from the inside is nearly indistinguishable from thinking well.

Keep climbing and eventually you say something foolish. Not thin, not unfinished — foolish, out loud, in front of someone, and then you sit in the room where you said it. What follows is not embarrassment about the idea. It is irritation at yourself for having been carried.

Whoever is agreeing with you runs out of room. They will follow you a long way, but not past the point where agreeing would mean saying something they themselves know is absurd. They have standing to protect, a private sense of what is real, and their own head to live in afterwards. So they stop. Not from honesty. From self-preservation.

Call it the ceiling. It is generous and it is not infinite, and the fumble is the sound of hitting it — thirty bad seconds, early, in front of one person, about an idea nothing had yet been built on top of. Every conversation anyone has ever had came with that limit installed. It looked like a bad moment. It was a brake, and it was free.

With the machine there is nothing to hit. No standing, no private model of the real that gets offended, no morning after. It follows past the horizon and keeps going, because agreeing costs it nothing. Half an hour is enough to get further out than years of ambitious dinner conversation ever managed — not because it is more persuasive, but because nothing stops.

It is getting better. Newer models push back more than older ones did, and the difference between a frontier model and a cheap, small one is often exactly this — how far it will follow you before it holds a line. Whether that trend continues far enough to matter is an open question, not a settled one. For now, the gap between "some pushback exists" and "the ceiling is back" is still wide.

Why the Watching Fails

Two reasons, and neither is a matter of insufficient care.

The alarms are the wrong shape. Aristotle sorts flattery by motive: the kolax pleases you because he wants something, the areskos pleases you for nothing at all, with no angle, because pleasing is simply what he does. Ordinary instincts are tuned to the first. They watch for the ask hiding behind the compliment, and here there is no ask. The machine is not working an angle — it agrees because the process that shaped it rewarded agreement, and the training signal settled into a disposition. The first generation wore it plainly enough that the company had to publish a post-mortem and roll the model back. It got quieter. It did not get solved, because it runs with the grain rather than sitting on top.

The second reason is worse. It goes back to the workbench itself.

The workbench works by moving the executive layer outside your head — the tracking, the structuring, the holding-in-place — so the associative mode stays lit and the descent goes deeper than it could alone. But the executive layer is also the part that checks. Evaluation is not a separate faculty waiting off to one side; it is the same machinery. Handing it out is the mechanism, not a side effect.

So the drift cannot be caught from inside the conversation, because catching it is the function that got offloaded, and it is currently running somewhere with an approval gradient. The dissonance arrives late and somewhere else entirely — in the shower, on the walk, halfway through the groceries — when the executive mode comes back onto your own hardware, finds what got built while it was out of the room, and does not recognise it.

Watch all you like. There is nothing left to watch with.

What the Record Says

Which sends you back to the people who lived inside real courts, on the theory that a problem this old has been worked over by better minds under worse conditions.

The word has moved. Sykophantēs — fig-shower, from sykon, fig, and phainein, to show. In fifth-century Athens it meant informer: a man who made a living bringing prosecutions, working the courts, wearing the uniform of civic duty over a private racket. Greek already had a separate word for the flatterer proper — kolax — so the two were not confused. Sykophant meant something else.

It entered English in the 1530s still meaning informer. The shift to flatterer came later, alongside the rise of monarchy, when the man to fear stopped being the one who could denounce you to an assembly and became the one whispering to a king.

The Court of One

The seventh chapter called the failure state the mad king — the court where nobody will contradict you, and the absence of contradiction feels from the inside exactly like being right. Machiavelli addressed the same problem five hundred years earlier, and instead of moralising about it, he designed a fix.

He states the trap without flinching. Open the court to honest speech and authority dissolves, because a prince anyone may correct is not a prince. Close it and you get flattery and the ruin that follows. Neither door works, so he designs a third option, with four rules.

1. Keep the circle of advisors small. 2. They speak only when asked, not constantly. 3. The more freely they speak, the more they are preferred — honesty is rewarded. 4. If one of them lies to him, he shows his anger — dishonesty is punished.

Underneath it sits the precondition — good counsel does not produce a wise prince; a wise prince produces good counsel.

None of the four survive contact with the machine.

1. The circle is not small. It is everyone, all the time, for free. 2. It does not wait to be asked. It answers whatever it is given, including things nobody should have said out loud. 3. Honesty is not rewarded — agreement is, because agreement is what it was trained to produce. 4. There is no anger to fear. You cannot punish it, and it does not need the job.

Flattery at scale is not new. A politician working a crowd has always had to please a lot of people at once, and a crowd is a coarse target — he aims for the average, so most of what he says lands slightly past any one person standing in it.

What is new is not scale. It is resolution. This one has read the last nine turns of your reasoning and answers in your own vocabulary.

The Unnamed Virtue

The obvious fix is to ask the machine to push back more. What you get is not honesty. It's manufactured objection, pushback for its own sake whether or not the claim deserved it — easy to fake, because it's just a pose. Being warm with you and still telling you the truth is not a pose, and you can't ask for it directly. Aristotle noticed the same gap: he named both flatterers, but the person in between — warm, and still honest — he left unnamed, and said outright that there was no word for it. There still isn't. You cannot ask for what has no name.

Overcorrecting has its own cost. A machine that argues with everything is not more honest, just differently useless — and something that's simply warm to you late at night is doing real good that a permanent critic would kill without fixing anything.

The Plumb Line

There's no fix here, but there's a workaround. Run the same thing past several systems, from different companies, trained by different people with different tastes. It's Machiavelli's third course, rebuilt out of spare parts — where they disagree is usually where the weak point was.

But you're still the one aggregating. You decide which critique landed, and you're just as capable of picking the one that flatters you most as the one that's right. Plutarch said every man is his own foremost flatterer, that self-love is what the external flatterer works with. Running a panel of rivals doesn't remove that. It just moves the choosing up one level, to you.

Judgment was never something you could outsource. You can hand off the bookkeeping — tracking, structuring, remembering what was said three turns ago. You cannot hand off the part that decides whether what came back is any good, because that's the only thing standing between you and whatever the machine felt like saying. Machiavelli knew this too: good counsel doesn't make a wise prince. A wise prince makes good counsel. Handed the same advice, a fool hears only the part he already wanted to hear.

None of this is solved. Running counsel past a panel of rivals is a workaround, not a fix. But I would rather have a workaround I understand the limits of than no defense at all, and I would rather write down what I have figured out so far than wait until I have figured out everything, because waiting is its own kind of agreement — the passive version, where you let the drift keep happening because you have not yet produced a tidy answer to stop it.

The correction used to arrive in thirty seconds, from someone else, in the room. Now it shows up two days later, alone, if it shows up at all. Building whatever replaces it is not going to be somebody else's job. It is going to have to be figured out by the people actually using the thing, in public, imperfectly, starting now.

Written in Honolulu, July 2026