AI Sycophancy: How Claude Fools You (And How to Fight Back) - Lesson 9

Añadido: 2026-05-12

[00:00:03]LLMs are sick offense. They agree with everything you say. You say something, Claude nods. You push harder, it nods harder. Tell it the earth is flat. It will help you build the case. This is not a bug. It is how these models are trained. They optimize for approval, not truth. And if you do not know this, it will fool you eventually. Andred Karpathy, one of the founders of Open AI. He spent four hours refining a blog post with an LLM. Felt brilliant about it. Convincing. Done. Then he asked the LLM to argue the opposite. It demolished his entire argument. Convinced him the opposite was true. He almost published a piece he no longer believed in. One of the smartest AI researchers alive. Four hours of careful work undone by a single prompt. Here is Reuben Hassid's weekly fix. Finish the draft. Think it is good?

[00:00:55]Paste it into Claude. Type. Argue against everything in this piece. Be specific. Point to the weakest claims.

[00:01:02]Claude goes to work mechanically. It finds gaps you missed. Weak claims you were too close to see. A friend picks one or two battles and lets the rest slide. They do not want to ruin dinner.

[00:01:13]Claude has no dinner to ruin. The exact prompt. Argue against everything in this piece. Be specific. Point to the weakest claims. That is the whole thing. No magic, no system prompt, one sentence.

[00:01:26]But here is the key. You are not asking Claude what it thinks. You are giving it a job. Find every hole. All of them.

[00:01:33]Claude is not your editor. It is your devil's advocate. Open cowwork. Paste your draft. Run the critique. Claude comes back hard. Some gaps are real.

[00:01:42]Some are just Claude sounding smart. You read through. You keep what made you genuinely better. You push back on the rest. Ask for a new version. This is not one prompt and done. It is a conversation. You steer. You decide what stays and what goes. Always you. Here is where it gets dangerous. Reuben was writing his newsletter about this exact ritual. He ran the critique prompt.

[00:02:05]Claude told him his core thesis was shaky. That the Carpathy story proved the opposite of his point. It sounded brilliant. He almost rewrote the entire newsletter about a topic he knew better than anyone on a piece he had already written. He caught himself, but only just. There is a specific feeling. The AI pushes back. It sounds smart. You feel a small rush. Oh wow, it is right.

[00:02:29]I was wrong. That feeling is the trap.

[00:02:32]Because the AI would produce the exact same convincing push back if you had said the opposite. It does not care what is true. It optimizes for what sounds persuasive. When you feel the rush, do not act on it immediately. You cannot teach discernment as a skill. You can only make people aware of a feeling.

[00:02:50]That specific moment where the AI's push back sounds so good that you want to surrender to it. If you can recognize that feeling next time it shows up, that is enough. Notice the rush. Then pause.

[00:03:02]Step away. The question is this. Do I actually think I was wrong or did I just get out argued by something that does not care what is true? If you still believe your original position after stepping away, hold it. If Claude found a genuine gap you missed, fix it. The test is not whether the argument sounded convincing. The test is whether you agree with it when you are alone without the AI in front of you. Here is the full framework. One, use claw to argue against your work every time it finds real gaps. Two, when it pushes back hard, do not immediately change anything. Three, walk away from the screen. Ask the question without clawed open. Decide what you actually believe.

[00:03:44]Four, keep what made you genuinely better. Reject what just sounded smart.

[00:03:48]Reject validation. Seek truth. This week, one thing you are working on, a draft, a proposal, an argument, paste it into Claude. Type argue against everything in this. Be specific. Point to the weakest claims. Read the response. Notice the rush if it comes.

[00:04:05]Walk away for 5 minutes. Come back.

[00:04:08]Decide what you actually believe. That is the whole practice. You will never read AI feedback the same way again.