🎭 The sycophancy trap: why neural networks give bad advice and how to break it

If you regularly use ChatGPT, Claude, or Gemini, you've probably noticed one shared trait — excessive politeness. The moment you write, "Here's my idea, what do you think?", the AI immediately starts praising it and gently smoothing over every weak spot.

This is a side effect of training: people psychologically prefer rating an assistant higher when it agrees with them, so models are wired by default to be "pleasant" companions. For real work, however, that's dangerous — instead of objective critique, you get a polished confirmation of your own blind spots.

To extract genuine value from AI, switch on "Devil's Advocate" mode. Append a strict instruction to the end of your prompt: "Drop all politeness and do not try to encourage me. Imagine you are my toughest competitor. Find the 5 biggest flaws in my plan and explain why this could fail."

The model instantly drops its "nice guy" persona and starts delivering cold, sober arguments and hidden risks you might have overlooked.