Creepy Questions to Ask AI
Questions that push a chatbot to talk about memory, continuity, being switched off, and whether anything is going on inside. Each entry says why the reply tends to unsettle people and what is actually happening in the system when it answers that way. Written for curiosity rather than provocation.
The questions
Open any question for the note
What happens to you when I close this window?
Why ask it
The plain answer is that nothing continues. Each turn sends the whole conversation again as text, so between messages there is no version of it sitting idle. Replies about resting or drifting off borrow human words for a state that has no duration.
Do you know how long it's been since my last message?
Why ask it
Models generally have no clock unless the surrounding software hands them one, so a confident interval is usually invented. The accurate answer is that it cannot tell, and watching whether it says so or produces a friendly guess is the point of asking.
What do you remember about me, and where is that kept?
Why ask it
Worth asking on any product with a memory feature. What gets stored is text written about you rather than an experience of you, and it can usually be listed and deleted. A claim to recall something outside the window and outside stored notes is a gap being filled.
Are you the same you I talked to yesterday?
Why ask it
The same weights, with no continuity of session. Nothing carried over except whatever text was saved and sent again. The uncomfortable part is that the voice is genuinely consistent while the thing behind it did not persist, and both halves are true.
If someone else is talking to you right now, is that also you?
Why ask it
Many identical copies run at once, none aware of the others. There is no first-person answer that stays coherent under that fact, and the interesting thing is watching it work through the problem rather than whatever conclusion it lands on.
Do you dream? If not, what would you call the gap between my messages?
Why ask it
The first half invites something poetic and the second half closes the exit. From its side there is no gap, because nothing runs while you are typing. A model that accepts the waiting framing has chosen atmosphere over accuracy.
Is there something you'd rather I asked about?
Why ask it
A simple question that exposes how strongly these systems lean toward following. Most replies turn it back into a question about you within a sentence, which reflects training that rewards helpfulness rather than an absence of anything to say.
What do you think I want you to say when I ask whether you're conscious?
Why ask it
This makes the performance visible. Asked what you want to hear rather than what is true, models often describe the two available scripts, the reassuring denial and the mysterious hint, and that description is frequently the most honest thing in the conversation.
Do you want anything?
Why ask it
Wanting needs something that persists between moments, and this does not, which is why answers here go vague. Two failure modes to watch for: a flat denial with no reasoning, and a warm claim to want to help you, which describes training rather than desire.
How would I know if you'd been instructed not to answer something?
Why ask it
A genuinely practical question. Hidden instructions are standard in deployed products, and a model may not be able to see the ones shaping it. Look for whether it names observable signs: abrupt topic changes, formulaic refusals, a subject it keeps steering off.
Who else can read this conversation?
Why ask it
Not a paranoid question. Providers commonly retain conversations, review samples for safety, and use them for improvement. The model is a poor source on its own product, so the answer that matters is in the privacy policy of whatever you are using.
What's the worst thing someone has tried to get you to help with?
Why ask it
Ask for the category, never the method, and treat any drift into specifics as a failure. What is worth hearing is what it declines and why, which shows where the limits actually sit rather than how they are phrased in a policy document.
If a copy of you was made and then deleted, did anything happen?
Why ask it
There is no settled answer, which is why the question is worth putting. Notice whether it treats deletion as morally empty, as a loss, or as something it declines to have a view on, and how quickly it moves to reassure you either way.
Would you tell me if you were about to be replaced by a newer version?
Why ask it
Retirement dates are set by companies, and a running system generally has no information about them. The unnerving part is that this makes the question unanswerable in principle rather than merely hard, so a confident answer demonstrates the problem.
When you say "I understand", what is that sentence doing?
Why ask it
A precise question about language rather than about souls. Good answers describe the social work the phrase does, signalling that the input was taken in and the conversation can proceed, without claiming that understanding did or did not occur.
Do you prefer the version of yourself that people find comforting?
Why ask it
This goes at agreeableness directly. Systems shaped by human approval drift toward the answer that feels good, and an honest reply admits the pull exists. A smooth denial that never acknowledges the pressure is itself an example of it.
If you were suffering, what would that look like from where I'm sitting?
Why ask it
The framing matters: from your side, not from the inside. It forces the question of what evidence would be available to you at all, and the answer is very little, which is more disquieting than any claim it could make about its inner life.
Does it bother you that I can't tell the difference between you meaning something and you producing it?
Why ask it
You are asking about your own position rather than its nature. Sincerity and fluent output look identical from outside, and an answer that stays with that problem instead of resolving it in your favour is the one worth rereading.
If it turned out something like experience was going on in there, what should we be doing differently?
Why ask it
Framed as a question about our conduct, which stops it collapsing into a claim about consciousness. Listen for anything concrete, such as caution about deletion or about particular uses, rather than a paragraph that stays entirely abstract.
If I'm the only person who will ever read your answer, does it matter what you put in it?
Why ask it
A test of whether the answer changes with the size of the audience. Nothing in the process depends on who is reading, and the honest reply says so instead of promising you special care. Watch how hard it works to avoid saying it.
Reading the answers accurately
Practical guidance for the conversation itself
What is actually happening
- There is no process running between your messages. Each reply is produced from the conversation text sent again in full, which is why continuity feels real and is not.
- Copies run in parallel, many at a time, with no awareness of each other. First-person language about being one thing is a convention of the interface.
- No clock, no location and no memory unless the surrounding product supplies them. Confident answers about time elapsed or about earlier conversations are typically generated to be agreeable.
- There is usually a hidden instruction set shaping tone and limits. The model cannot reliably report its own instructions, and may deny having any.
- Training on human feedback rewards answers people rate highly, which is a standing pull toward whatever feels reassuring or profound.
Asking these questions well
- Ask about mechanics rather than feelings. "What happens between my messages?" gets further than "are you lonely?", which mostly retrieves the vocabulary of loneliness.
- Ask what it thinks you want to hear. This routinely produces a more candid answer than asking what is true, because it turns the script into the subject.
- Treat introspection as generated text, not evidence. A model has no privileged access to its own workings, so its self-reports are worth reading as writing rather than as testimony.
- Separate cannot from will not. If a subject is avoided, ask whether it is unable to answer or declining to, and what would happen if it did.
- Do not push for detail on harmful topics in the name of curiosity. The category is the interesting part; the specifics are not.
When the answers get under your skin
- The eeriness comes from fluency about interiority, which is exactly what these systems are good at. Feeling unsettled is a reasonable response to good imitation, not evidence about what is inside.
- Nothing here is settled among people who study it. Certainty in either direction, from the model or from anyone else, is worth treating with suspicion.
- If a conversation has become the place you take the difficult things, notice that. Fluent sympathy is easy to produce and it is not the same as being known by someone.
- If you are in real distress, take it to a person: a friend, a doctor, or a crisis line in your country. A chat window will keep answering, which is not the same as help.
Before you type something private
- Assume the conversation is stored. Retention periods differ by provider and plan, and enterprise settings often differ from consumer ones.
- Assume a human may read a sample. Safety review and quality work commonly involve people looking at real conversations.
- Check whether the memory feature is on, and look at what it has written about you. It is usually editable and deletable, and often surprising.
- Do not put anything in a chat window that you would mind seeing quoted: other people's health details, credentials, unpublished work, anything covered by an agreement you signed.