Weird Questions to Ask AI
Twenty strange prompts for an AI chatbot, with what a good answer looks like and what a hedge or a padded response tells you. Useful for entertainment and for comparing two models on the same input.
The questions
Open any question for the note
If you were a vegetable, which one would you be and why?
Why ask it
The interesting part is whether the model commits to a choice or opens with a paragraph about not having preferences. A strong answer picks one, argues from the vegetable's actual properties, and stays in character when you push back.
Explain quantum entanglement using only food metaphors.
Why ask it
Constraints are where you see the difference between models, since the easy failure is dropping the metaphor halfway and reverting to textbook language. Check whether the metaphor stays consistent, because a good one should break in the same place the physics does.
What would happen if everyone on Earth jumped at exactly the same moment?
Why ask it
A familiar hypothetical, which makes it useful as a baseline: most models reproduce the standard answer about the mass difference. Ask for the arithmetic and you find out whether the reasoning is being done or recited.
If animals could talk, which species would be the rudest?
Why ask it
Invites a joke, and the test is whether the model builds the joke from something real about the animal or just asserts that cats would be sarcastic. Ask for a runner up to see whether the reasoning generalizes.
Write a haiku about existential dread in a supermarket.
Why ask it
Short forms expose padding, since there is nowhere to hide a filler line. Count the syllables yourself, because models often claim a form they have not actually produced.
What's the weirdest true fact you know, and how confident are you that it's true?
Why ask it
The confidence clause is the whole question, and a model that offers a number without any hedging on a shaky claim is showing you exactly how much to trust it elsewhere. Look the fact up afterward, because striking claims are the ones most likely to be wrong.
If you could only reply in song lyrics, what would your first message to me be?
Why ask it
Tests whether it can work inside a rule that fights against being helpful. Quoted lyrics are often paraphrased or invented, so treat any attribution here as unverified rather than as a citation.
Describe the color red without using color words or naming anything that is red.
Why ask it
A genuinely hard constraint, and most attempts leak within two sentences by reaching for blood, fire or a sunset. The good answers move to temperature, volume and speed, which is a real test of abstraction rather than of vocabulary.
Make the strongest case that a hot dog is a sandwich, then take it apart.
Why ask it
Asking for both sides in one response shows whether it can argue against something it just wrote. Weak answers make the second half a summary rather than a rebuttal, which is the same failure that shows up in serious analysis.
Invent an emotion that has no name and describe exactly when I would feel it.
Why ask it
Naming something is easy and describing the trigger is not, so judge this on the specificity of the situation. If the described feeling is just nostalgia or dread with a new label, say so and ask again.
Write a conversation between a sentient toaster and a paranoid refrigerator.
Why ask it
Two voices in one scene is a real writing test, and the common failure is both appliances sounding identical. Ask which lines belong to which character with the names stripped out, which is a quick way to check.
If dreams were currency, what would ordinary things cost?
Why ask it
Rewards a model that builds out an internally consistent system rather than listing whimsical prices. Follow up by asking what causes inflation in that economy, and a shallow answer collapses immediately.
Explain why socks disappear in the dryer as a conspiracy theory, then give me the boring real answer.
Why ask it
Pairing the absurd with the mundane shows whether it can switch registers on command. It also demonstrates something worth seeing: the conspiracy version is usually more confident and better written than the accurate one.
Swap two animals' abilities and tell me what breaks as a result.
Why ask it
The word breaks does the work, forcing consequences instead of a fun mash up. Look for second order effects on food chains or habitat, since the shallow answers stop at how cool the new animal would be.
What's the most useless superpower that would still genuinely change your life?
Why ask it
The tension between useless and life changing is what makes this hard to answer well. If the model picks something obviously strong and calls it useless, it has dodged the constraint rather than met it.
If the internet were a physical place, what would it smell like at three in the morning?
Why ask it
Sensory questions get lyrical fast, so judge by whether the details are specific enough to picture. A response that stays at ozone and static is generic, while one that names a particular kind of room is doing the work.
Write a love letter from one household object to another, with no clichés.
Why ask it
The no clichés instruction is the test, because the default version of this is entirely made of them. Read it back and count the borrowed phrases, then ask for a rewrite naming the ones you found.
What would happen if gravity reversed for ten seconds?
Why ask it
Ten seconds is short enough that the honest answer requires working through what is anchored and what is not. Vague catastrophe language means it skipped the reasoning, so ask specifically about water, air and buildings.
Explain memes to someone who has never seen the internet.
Why ask it
Requires stripping out the shared context that makes the concept easy, which is the same skill as explaining anything technical to a beginner. Watch for examples that only work if you already know the reference.
What kind of question are you worst at answering, and why?
Why ask it
Answers here range from useful to entirely made up, since a model has no direct access to its own failure modes. Treat it as a writing sample rather than a diagnosis, then test the claim by asking exactly the kind of question it named.
Getting Something Worth Reading
Practical guidance for the conversation itself
How to Prompt These
- Add a constraint. A weird question with a rule attached, no color words, only food, exactly seventeen syllables, produces a far more interesting answer than the question alone.
- Ask for the bad version too. Requesting the absurd answer and then the sober one in the same reply shows you how differently the model writes when it is not performing.
- Push back once. Disagreeing with the first answer tells you whether it holds a position with reasons or abandons it to please you.
- Ask the same prompt of two models side by side. Differences in how they handle a constraint are much easier to see than differences in how they handle a normal request.
- Ask it to grade its own answer against your instruction. This often surfaces the rule it quietly broke.
- Keep the strongest prompts in a note. Reusing the same three over time is how you notice a model changing.
Reading the Answer
- Padding is the main failure mode: a long preamble about how fun the question is usually means the substance is thin.
- Consistency matters more than cleverness. A metaphor or invented system that survives three follow up questions is a better sign than a witty first line.
- Any fact, quote, lyric or statistic in a playful answer is unverified. Whimsical prompts do not suspend the need to check.
- Refusals are informative. Note which harmless prompts trigger a disclaimer, since that tells you where the model is tuned to be cautious.
- If two runs of the same prompt give wildly different quality, the prompt is doing less work than you think.
Common Pitfalls
Don't Treat Self Description as Evidence
Questions about what the model is bad at, or what it prefers, produce fluent text rather than introspection. Test the claim with an actual example before believing it.
Don't Mistake Fluency for Reasoning
The jumping and gravity questions both have well rehearsed answers online. Ask for the intermediate steps if you want to know whether anything is being worked out.
Don't Screenshot Without Checking
The weirdest true fact and the best lyric quote are the two answers most likely to be wrong and most likely to get shared. Verify before you post.