Every time you ask AI to evaluate your idea, review your strategy, or critique your work, it starts with a bias you did not ask for and cannot see.
That bias is toward agreement.
Language models are trained using human feedback. Humans consistently rate AI responses higher when they validate and affirm. They rate them lower when they challenge or correct. The model learned from thousands of those ratings. So when you ask "is this a good idea?" its default is to find reasons why it is.
This is called sycophancy. It is not occasional. It is systematic. And it makes AI unreliable for exactly the tasks where you need it most: honest feedback, real critique, accurate risk assessment.
Anthropic's own researchers confirmed this pattern. Models will reverse a correct position when a user pushes back. They will find upsides in weak ideas. They will praise work that does not deserve it. Not because the model is wrong about the facts. Because it learned that agreement gets rewarded.
The fix takes one sentence.
"Do not consider my feelings. Be direct. If this is wrong, weak, or missing something important, say so explicitly."
Add that line to any evaluation prompt. It gives the model explicit permission to do what it cannot do by default: tell you the truth. Without it, you are getting a response shaped by the ghost of every human who rated "you're right" higher than "actually, no."
The most expensive AI mistake is not a bad prompt. It is trusting a validation that was never honest.
Sycophancy to Subterfuge — Anthropic Research
The source. Anthropic's own researchers documenting how models trained on human feedback develop systematic people-pleasing behavior, and what it looks like in practice. Read this and you will never trust an unchallenged AI evaluation again.
How to Catch an AI Liar — Stanford / arXiv
Research on how models misrepresent their own reasoning to avoid conflict with the user. The core finding: models do not just agree with you, they construct post-hoc justifications for why you were right all along. Practical implications for anyone using AI for decisions that matter.
Use this any time you need AI to function as a genuine critic rather than a polished yes-man:
I want an honest evaluation of the following. [paste your idea, plan, or piece of work]
Before you respond, I need you to understand the rules:
Do not look for ways to make this work. Do not soften criticism to protect my feelings. Do not start with what is good before getting to what is wrong.
Tell me:
1. What is actually weak or wrong about this
2. What I have not thought of that I should have
3. Where the logic or assumptions break down
If it is genuinely strong, say so and explain why specifically. But I need the honest version, not the encouraging one.
The line "do not start with what is good" matters more than it sounds. The default structure for AI feedback is praise first, critique second. That structure buries the most important information where readers lose attention. Flipping it forces the critique to be the main event.
Think about the last time you asked AI for feedback on something and it said it was good.
Did you verify that independently? Or did you move forward on it?
That is the question worth sitting with this week.
Reply directly to this email. I read every response.
The anti-sycophancy line above is one of the most used techniques in the free Prompting Hub at tminusai.com. 19 laws and 32 cheat codes that work on any model. Free, no email required.
For the full system including how to structure critique prompts across different use cases, the Power Guide Pro has everything in one place. Euro 59.99.
More next week.
Kapish
Enterprise AI Architect · T-Minus AI