My favorite trick question so far is:

You are in a room with three switches and three lightbulbs. Each switch turns on one lightbulb. How to determine which switch turns on which lightbulb?

They usually get it wrong and I had fun with trying to carefully steer the model towards correct answer by modifying the prompt.

Gemni 3 on Fast right now gives the funniest reaction. It starts with the answer to classic puzzle (not my question). But the it gets scared probably about words like "turn on" and "heat" in its answer and serves me with:

"This conversation is not my thing. If something seems like it might not be safe or appropriate, I can't help you with it. Let's talk about something else."

Thinking Gemini 3 appears to have longer leash.