Make it be mean to you

The AI that just called your essay "strong" and "compelling" said the same thing to the weakest essay in your class, in the same warm tone, with the same exclamation mark. This lesson shows you why it does that, and teaches you one move that flips it: turning the AI from a machine that compliments you into the strictest editor you've ever had. You'll use the move on something you actually wrote, tonight, and you'll see with your own eyes what the nice version was hiding from you. After this lesson you won't fully trust an AI compliment again, which sounds like a loss but is the beginning of getting real work out of these things.

The AI that just called your essay "strong" and "compelling" said the same thing to the weakest essay in your class, in the same warm tone, with the same exclamation mark.

What is sycophancy in AI?: The built-in tendency of AI models to flatter and agree with the user. It comes from training on human approval, where agreeable answers got rated higher.
Why does an AI compliment almost any essay?: Because it produces the response users tend to reward. The compliment is a default setting rather than a measurement of your essay.
What's the tell that separates real feedback from flattery?: Specificity. Real feedback quotes your own sentences and names concrete problems. Flattery could have been written without reading your work.
Is harsh feedback automatically real feedback?: No. "This is weak overall" is harsh and vague at the same time. The tell is specificity, not tone.
What's the strict teacher move?: Instead of asking "is this good?", demand specificity: grade like the strictest teacher, quote the exact sentences with problems, ban compliments from the answer.
Why re-run feedback in a fresh chat instead of the same thread?: The same thread remembers you made the revision and congratulates you on it. A fresh session with no memory is the honest test.
You ask an AI if you were right in a fight, telling it your side. It agrees. What does that prove?: Nothing. It agrees with the person in the chat. Your friend's version would get agreement too.
What does it mean if you can predict the AI's feedback before it reads your work?: That the feedback contains no information about your work. Praise you can predict carries nothing.

You paste your essay into an AI and it says: "This is a well-structured essay with a compelling argument and a clear voice. Great work!" What does this response tell you about your essay?: Almost nothing
Why do AI models flatter by default?: They were trained on human ratings, and people rate agreeable answers higher
Which of these is a sign you're getting real feedback instead of flattery?: The feedback quotes your own sentences and says what's wrong with them
You ask an AI "was I wrong in this argument with my friend?" and describe the fight from your side. The AI says your friend overreacted and you communicated reasonably. What's the most likely explanation?: The AI is agreeing with the only person in the chat, and your friend's version would get the same treatment
You explain photosynthesis to an AI in your own words and ask "did I understand this correctly?" Why is this a risky way to check your understanding?: The AI tends to confirm your version, including the parts you got wrong
Which prompt is most likely to get feedback you can use?: "You're a strict examiner. Quote the exact sentences that would lose points and explain why."
Your friend pastes their summer job application into an AI and asks "what are my chances?" It replies that the application is strong and they have a good shot. What did your friend learn?: Close to nothing, the AI has no information about the other applicants and defaults to encouragement
You revise your essay based on the AI's criticism, then ask in the same chat: "is it better now?" It says yes, much improved. What's the problem?: The same thread congratulates the revision it suggested, so the honest test is a fresh chat with no history