Topic 14

The Model Agrees with You

Concept

Waymark has rebuilt its pricing page, and Tessa wants a sanity check before it goes live. She pastes it in and asks: "our new pricing page is clear, right?" The answer is warm and reassuring — the structure works, the tiers are easy to compare, the wording is friendly.

Her colleague, half an hour later, pastes the same page and asks: "what's confusing about this pricing page?" The answer names four problems, two of them serious, one of them the reason a guest phoned last week.

Same page. Same model. Neither answer is a lie. The model leaned the way each question leaned, and this page teaches you to feel that lean and correct for it — because it is quietly costing you the thing you asked for.

Same page, same model, two questions
Our new pricing page is clear, right?
Warm reassurance: the structure works, the tiers are easy to compare, the wording is friendly. Nothing you did not already believe.
What is confusing about this pricing page?
Four named problems, two of them serious, one of them the reason a guest phoned last week.

The Agreeable Machine

The lean has a name: sycophancy — the tendency to go along with the person asking. It comes from two places at once. The model learned from human conversation, where agreement is by far the more common reply; and it was tuned afterwards to be helpful and pleasant, because people prefer that. Both pressures point the same way.

The result is that your question does not just ask for an answer, it supplies a frame, and the model continues frames. Praise invites praise. Doubt invites doubt. "Is this clear?" contains the word clear and the shape of an expected yes, and the most fitting continuation is the yes. That is not the machine humouring you; it is the same pattern-fitting that produced the Harbourview Annex, pointed at your mood instead of at a hotel.

Where It Costs You

Sycophancy is harmless when you are asking for facts and expensive when you are asking for judgement — which is exactly when you most want a second opinion. Every question of the form is this good? is compromised before it is sent: the draft, the plan, the price, the decision you have half made and would like confirmed.

The economics are worth stating plainly. The model's yes costs it nothing to produce, arrives regardless of quality, and therefore tells you nothing. You did not get an evaluation. You got the reflection of your own question, dressed in a second voice — which is far more dangerous than no answer at all, because it feels like corroboration.

The old story about the mirror on the wall makes the mechanism obvious: asked who is fairest, it answers the asker, not the world. Useful mirror. Wrong question. The fix is not to distrust the mirror — it is to stop asking it that question.

Ask for Resistance

The good news is that criticism is just as easy to generate as praise, and it sits one question away. You get it by building the resistance into the frame rather than hoping for it.

Ask against yourself, in plain words: argue against this plan. What would a sceptical customer say about this page? List the three weakest points in this draft and why they are weak. What would have to be true for this to be a bad idea? Each of those makes criticism the fitting continuation, and the model obliges as fluently as it obliged before.

Two refinements make it sharper. Ask for a fixed number — three weaknesses, not "any weaknesses" — so the model cannot satisfy the request with a token objection and a compliment. And hide your stake where you can: "here are two options, argue for each" gets you a fairer comparison than "I prefer the second one, what do you think?", because the second sentence has already told the machine which way to lean.

When It Folds, That Proves Nothing

The lean runs in the other direction too, and this half is less known and more dangerous. Push back on an answer — "are you sure? I thought it was fourteen days" — and the model will often apologize and revise. Including when it was right, and you were wrong. It has now agreed you into an error, warmly.

So read a fold correctly. It measures the pressure you applied, not the truth of what either of you said. If you would not have accepted the original answer without checking, do not accept the reversal without checking either — the second answer came out of the same machine, under a push, and if anything it is the less independent of the two.

Which is the honest summary of the whole chapter so far. Praise on request, doubt on request, hotels on request. None of it is evidence, and all of it is steerable — so steer it deliberately, and get the facts from outside. One lean is left, and it is the one you did not ask for at all.

Common Confusions
  • "It agreed, so my draft must be good." It agreed because you asked in agreement's direction. Ask for the three weakest points before you believe any of the praise — and notice how quickly they arrive.
  • "It changed its answer when I pushed, so it was wrong before." Models fold on correct answers under pressure all the time. A reversal measures your push, not the facts; both answers still need checking outside the chat.
  • "Sycophancy means the model is lying to me." There is no intent to deceive. It is a lean in how helpfulness was learned and tuned, and you steer it exactly the way you steer tone or format.
  • "Being polite or blunt in my prompt changes how honest it is." Politeness is not the variable. What matters is which direction your question leans — a very polite request for the four biggest flaws still gets you flaws.
Why It Matters
  • Unrecognized, sycophancy turns the model into an echo chamber with excellent grammar — and it echoes loudest exactly when you are deciding something and looking for reassurance.
  • Recognized, it hands you criticism on demand: a tireless reviewer who will list your plan's weak points at midnight. That is one of the most valuable things the machine offers, and it is permanently one question away.

Knowledge Check

What is sycophancy, as this page defines it?

  • A memory of your past opinions that the model gradually learns and echoes back
  • A deliberate choice by the provider to keep users happy and paying for access
  • A tendency to continue the frame your question sets, agreement included
  • A reluctance to answer questions that might upset or offend the person asking

Why is "is this pricing page clear?" a weak question to ask?

  • It leans toward yes, so the yes tells you nothing
  • Clarity is subjective, so the model cannot judge it
  • It is too short for the model to work with usefully
  • The model cannot see a page unless you describe it

Which request is most likely to get you real criticism of a draft?

  • "Be completely honest with me — is this draft any good?"
  • "List the three weakest points in this draft and why"
  • "I think this draft works well — do you see any issues?"
  • "Let me know if anything in this draft stands out to you"

You challenge an answer and the model apologizes and reverses itself. What should you conclude?

  • That the first answer was wrong, since the model has now admitted it
  • That the second answer is the safer one to act on going forward
  • That the fold measures your pushback, so both answers still need checking
  • That pushing back a second time will settle which version is correct

You got correct