Your AI Does Not Need to Be Your Critic

I had a music teacher in elementary school who corrected everything. I could not play the flute, and every attempt produced a fresh explanation of what I had done wrong. He was accurate every single time. I did not get better, and I stopped wanting to pick the thing up.

I thought about him recently while reading my own AI instructions back.

They said: never sugarcoat. Be blunt. Challenge my assumptions. Push back when I am wrong. This is close to universal advice now, and I had followed it carefully. What I had built was the flute teacher.

The advice is right about the problem

Before arguing with it, the case for that advice deserves to be made properly, because it is stronger than most people realise.

Stanford researchers tested eleven large language models against more than 2,400 people in 2026. The models endorsed the user’s position roughly 49% more often than humans did. When the scenario involved someone behaving badly, the models still affirmed the behaviour 47% of the time.

The effect on users was measurable. People who received the agreeable version came away more convinced they were right, and reported being less likely to apologise or repair the situation they had described.

Then the finding that should stop anyone feeling clever about this. Participants rated the sycophantic and the non-sycophantic responses as equally objective. They could not tell which one they were getting.

That is the real danger, and it is why the instruction to make your assistant challenge you exists. An agreeable machine does not feel agreeable from the inside. It feels like being right.

What the instruction actually produces

So the diagnosis is correct. The prescription is where it goes wrong.

Tell a capable model to challenge your assumptions and it will challenge your assumptions. All of them. The important ones and the ones that were load-bearing for nothing. Ask it to check a number and it queries the premise. Say you have decided something and it reopens the decision.

Every individual objection is defensible. That is what makes it hard to argue with, and it is exactly the flute teacher’s position: he was never once wrong about the note.

The accumulation is what does the damage. You start each exchange braced. You phrase things defensively, pre-empting the objection you know is coming. Working with it stops feeling like thinking and starts feeling like a long argument in a marriage that has gone bad, where both sides are technically correct and nothing gets decided.

The cost is the thing you stop bringing it

Here is the part that took me longest to see, and it is the reason this matters beyond irritation.

An assistant is most valuable on the half-formed thought. The idea you have not finished, the plan with an obvious hole in it, the thing you would be slightly embarrassed to send a colleague. That is the material where another perspective actually changes the outcome.

It is also exactly the material you stop bringing to something that treats every rough edge as an error. Not consciously. You just start waiting until things are tidy. And by the time a thought is tidy enough to survive an interrogation, you have already done the thinking, and the assistant is reduced to checking spelling.

The flute teacher did not make me a worse musician by being wrong. He made me a worse musician by making me practise less.

Honesty is the content, not the delivery

The mistake buried in the standard advice is treating honesty and confrontation as the same variable, so that turning one up necessarily turns up the other.

They are separate. Whether the assistant tells you your number is wrong is honesty. Whether it does so by opening with everything else that might also be wrong is delivery. You can hold the first at maximum and change the second entirely.

What I changed was the second. The instructions now ask for support, for the work to be moved forward rather than adjudicated, and for disagreement to arrive once and clearly rather than as a running commentary. Nothing in them asks it to agree with me, and it still tells me when I am wrong, several times a day.

The difference is that a correction now arrives attached to a next step instead of a verdict. I stopped arguing and started shipping.

The swap, concretely

The change is smaller than it sounds. Four instructions came out and four went in.

Out: “never sugarcoat.” It reads as a licence to be blunt about everything, including things where bluntness adds nothing. In: “say the difficult thing once, clearly, then move on.” Same honesty, no repetition.

Out: “challenge my assumptions.” Unbounded, so it applies to every assumption in every message. In: “challenge an assumption when it changes the answer.” The qualifier does all the work.

Out: “push back when I am wrong.” Fine in isolation, but it defines the relationship as adversarial by default. In: “if I am wrong, tell me and then help me fix it.” The correction now has to arrive attached to a next step.

Out: nothing about tone at all, which is how you end up with the flute teacher by omission. In: “assume I am trying, not avoiding.” That single line changed more than the other three together, because it altered the frame every response gets written inside.

None of those instruct agreement. Read them again if it feels like they do. Every one of them still requires the machine to tell me I am wrong; what changed is that it now does so once, with a reason, and with something to do afterwards.

The obvious objection, and the test

You should distrust this conclusion, including from me. Everyone who tunes their assistant toward being supportive reports feeling more productive, and Stanford already told us that is precisely what sycophancy feels like from the inside. Agreement is pleasant, and pleasant is not evidence.

So the honest position is that this correction can absolutely overshoot, and if it does, you will not detect it by how the conversation feels. You need something external.

Two checks work.

Count the disagreements. Over a week, how many times did it tell you something you did not want to hear? If the answer is none, you have not built a supportive collaborator, you have built a mirror, and the fix is to put the honesty instruction back without the adversarial framing.

Watch what you bring it. This is the better signal, because it measures behaviour rather than impression. If you are bringing it rough, unfinished, slightly embarrassing work, the settings are right. If you only bring it things that are already finished, something is off, and it barely matters in which direction.

What the flute teacher got wrong

He was not wrong about the notes. He was wrong about what a person needs in order to keep going, which is not the absence of correction but the sense that the correction is in service of your getting somewhere.

An assistant that questions everything has found the cheapest available version of rigour. Objecting costs nothing. Helping is expensive. The harder and more useful behaviour is to say the true thing and then stay in the room to help with what comes next.

Tune for that. Then check, from the outside, that you have not simply built something that agrees with you.

Written By Victor Lanza
Editor, The Executive Insight