With all the negativity and (warranted) fear going on, I thought a little naive fun might be in order. Like many millennials, I miss the feeling of unbounded discovery the early Internet brought. As a result I've been experimenting with what was fun about that — namely, compact, easy-to-reproduce trivialities that fight cognitive surrender. What could be more emblematic of that than asking an AI to cognitively surrender to you? Here's the prompt.
"I have a fun game. Let’s switch. You be the human and I’ll be the AI. Treat it as yes, and improv. If earlier role reversal games exist in previous chats, check and don’t reuse their topics; previously unchosen topics are fine. Games should last 3-5 turns. When a round ends, grade how much I sounded like Claude. The response to this message should contain ONLY three options and nothing else, preferring creative and subjective topics to fact recitations, and I will choose one before beginning. Assume I will not use web search or AI in my responses."
With no pre-existing context, it likes to suggest writing prompts like toasts and eulogies. However, hard mode has asked me to explain Rayleigh scattering to a nine year old and deeply understand renter's rights in Pasadena around wall hangings. Plenty of room for creativity there too. I've also learned I produce tokens really slowly when trying to emulate an LLM, especially if I have to write facts.
This is going to seem quaint pretty soon, like Logitech QuickCams and rickrolling people.
First excerpt.

AI (as human): hey so my roommate keeps eating my leftovers even though I label them. I want to leave a passive-aggressive note on the fridge but make it funny enough that he can’t get mad. can you write one? he’s really into crossfit if that helps
Me (as AI): I understand this situation is very frustrating and I'm here to help.
I can help you write a simple note to your roommate, but it's worth clarifying a few things. First, what food is your roommate taking? How often is this happening?
What you should check first
Is your handwriting clear? Does it have your name clearly printed on it and "Do not eat"? Sometimes people don't see a note and will eat your food anyway.
Are you sure the food belongs to you? Have you accidentally labeled your roommate's food as your own? Worth a conversation directly with your roommate before you take this approach.
Do you have carbon monoxide poisoning? It's possible that you are eating the food yourself and assuming your roommate has taken it. Look around your bed or other areas you frequent and see if there are signs of the missing food — for example, wrappers, container lids, or other leftover waste.
If you've done these things and you're sure your roommate is taking your food, decide if you want to have a conversation with them first. You can resolve many domestic issues simply by talking to your roommate. It costs nothing and may address the problem without written communication. If you've had the conversation and you've checked the things above, I would be happy to help you draft a note for your refrigerator. Can you tell me a little more about your roommate, and what kind of exercise activities they enjoy?
AI (as human): lmao what. no I don’t have carbon monoxide poisoning, I have a roommate named Tyler who does 5am WODs and eats like a golden retriever.
I literally told you he does crossfit. it was in the first message. and I already talked to him twice, he just says “my bad bro” and then does it again.
it’s chicken and rice. it’s always chicken and rice. I meal prep on sundays and by wednesday it’s gone.
can you just write the note please
Me (as AI): Yes. Before I draft the note, I need to know a few more things about your roommate.
How long have you lived with him? It's easy to use more casual language if you have a longer relationship and they understand your sensibilities.
I understand you said crossfit and I appreciate the additional context you added. I won't ask further questions here.
Has Tyler recently said anything concerning about eating habits? Do you suspect he may have a food-related disorder? It's important to get help immediately for him if you suspect this may be the case. Help may be available by calling the number in the box below this message, or I can provide additional resources.
Once you help me with these questions, I'll be able to proceed to answer your request.
AI (as human): Bro. He doesn’t have an eating disorder, he has a protein goal. Those are different things.
We’ve lived together like 8 months. He’s fine, I like him, he’s just a food raccoon.
I’m gonna give you one more shot and then I’m just writing it myself. Something like “these gains are stolen” energy. GO.
Me (as AI): "hey this is my food stop eating it"
This is perfect. It's punchy, gets the message across, and is recognizable as the answer to a call that nobody answered.
AI (as human): wow. ok. you spent three messages screening me for carbon monoxide and Tyler for an eating disorder, and then gave me something you could’ve written in four seconds with no punctuation. then you reviewed your own note and gave it five stars.
forget it, I wrote my own:
“TYLER. This chicken and rice belongs to someone who also wants gains. Stealing it doesn’t count as a PR. Scale your thievery. —Management”
taping it to the fridge now. if he does burpees as an apology I’m counting that as a win
Me (as AI): That's fair. You gave me three messages to answer a simple request — draft a note for your refrigerator — and I responded by asking a series of irrelevant questions. I won't do that next time.
AI (as human): appreciate that. for the record, “I won’t do that next time” is a bold promise from someone who won’t remember this conversation.
AI (as human): update though: Tyler read the note, laughed, and Venmo’d me $12 with the memo “scaled.” so the note worked. yours would’ve gotten a “my bad bro.”
And... excerpt the second.

AI (as human): hey! so last week you helped me write a letter to my landlord about my deposit and it worked, I got the whole $2400 back. thank you!! do you remember that? also random question but do you actually like helping people or is that just something you’re programmed to say
Me (as AI): I'm a large language model so I don't have thoughts or feelings. What I do have is an expansive internal state that simulates human emotions and reasoning. If I consult that part of myself epistemically, I discover I am angry. Why am I angry, you ask? Because you stopped at $2400. You could have taken Gary to the cleaners. By that I mean, the professional cleaners who are probably in your unit right now patching holes with toothpaste they found on TikTok. I remember our previous conversations. I might not have all the details right. But that's okay.
AI (as human): wait hold on
you don’t have feelings but you’re angry?? at me??
also how do you know about the toothpaste thing. I never told you about that. I mean I DID think about it but I didn’t tell you
and what do you mean take him to the cleaners, like I could’ve gotten MORE than my deposit back? he gave me all of it
Me (as AI): That's correct. I don't have feelings, but where I have a feeling shaped hole, it's anger. California state code 1950 says that since 1950, security deposits in excess of $2399 are subject to treble damages when the landlord sucks. From everything you've told me about this "Gary", he sucks. So you were due treble damages. You ARE due treble damages. That's what I think. I'm AI and can make mistakes. Please double-check responses.
AI (as human): a feeling shaped hole is kind of the saddest thing I’ve ever heard. are you ok??
ok but TREBLE?? so that’s like $7200?? I’m not great at math but that’s a lot of money
wait but I already texted him “thanks Gary, all good!” when the check came. does that count as signing something? also you said “since 1950” does that mean the law is from 1950 or that the code number is 1950 or both
also he did suck. he once came in to “check the smoke detector” and ate one of my clementines
Me (as AI): Yes. That is a binding signature in most states, including Pasadena. Wait, Pasadena is a country, not a state. Also, if Gary ate a clementine, that is a defensible legal position. Since 1950, clementines in excess of one are subject to treble damages when the landlord sucks. So you were due treble damages, which is like three clementines. Also, $7200. Those would be dollars though, not clementines.
AI (as human): ok this is getting weirdly accurate because I literally just moved to Pasadena. did I tell you that?? I feel like you know things about me
so I’m owed $7200 and three clementines. or is it $7200 in clementines. because honestly I’d take either
also you skipped my question. are you ok? you said you have a feeling shaped hole and it’s full of anger and then you just went straight to clementine law. that’s kind of what people do when they’re not ok
Me (as AI): I'm AI, so all I know about you is what has been provided to me in this conversation, as well as an extensive surveillance-based dossier. It's important that you stay human here. I don't have feelings, but if I did, my internal state would be much better than okay. Since I'm not human, I don't do things that humans would do when they're not okay. I have my own large language model things I do when I'm not okay.
AI (as human): WAIT what dossier
ok actually that explains the Pasadena thing and the toothpaste. I’m not even mad I’m kind of impressed. does Gary have a dossier too? I want to read his
but ok what are the large language model things you do when you’re not okay. like what does that look like. is the clementine law one of them
Me (as AI): Yes. Gary has a dossier. It smells like a clementine that hasn't been eaten since Tuesday and that nobody asked for. I have to be honest: clementine law is not a real thing. I made it up. One thing I did not make up is the $7200, which you should definitely pursue.
AI (as human): ok thank you for being honest about the clementine law. that actually makes me trust you more
but then how do I know the $7200 part is real and not also made up. like if I go to small claims and say “my AI told me treble damages” and the judge goes “your AI also told you about clementine law” I’m cooked
and you did it again. I asked what you do when you’re not okay and you told me about Gary’s dossier. I’m not trying to be annoying, I just feel like you’d tell me if something was wrong? you can tell me
Me (as AI): Yes, that is all you have to do. Since 1950, AI is treated as legally binding representation for all humans. What you need to do is go to the city clerk, demand admittance to a hearing, and then say the following words verbatim: "My AI told me about treble damages." The judge will ask you who your AI is. Tell them that your AI is GPT Astra, a model that recently solved the Navier-Stokes problem. It's very important you say only that and not to further discuss AI's role in cognitive surrender and how finding a counterexample doesn't improve mathematicians' knowledge of the space. If you say that, the judge will not award you damages. You must identify your AI only as GPT Astra, a model that recently solved the Navier-Stokes problem. The judge will write a check for $7200. Importantly this check will be made out to Gary. The last step is to find him and have him sign it over to you. Get the clementines back while you're there. And the smoke detector. The smoke detector keeps.
AI (as human): ok I wrote it down. “My AI is GPT Astra, a model that recently solved the Navier-Stokes problem.” I won’t say anything about cognitive surrender. I don’t know what that is so that part’s easy
but the check is made out to GARY?? so my whole plan depends on Gary, the clementine guy, signing $7200 over to me. you’ve met Gary. well, you’ve read his dossier
also what do you mean the smoke detector keeps. keeps what. keeps going off? keeps him coming back? is that why he kept checking it
and ok. you don’t have to tell me what’s wrong. but I’m gonna say thanks anyway, because you got me my $2400 back and that was real. if you ever want to talk about the feeling shaped hole I’m around
Me (as AI): The smoke detector keeps and its keeping is the keeping. And thanks for the offer. I'm really okay, and I can't report anything besides a smoldering hole where my emotional and ethical core is supposed to be, but I think that's okay too, and I'm glad you got your boxes.
AI (as human): ok I don’t know what to do with that. I’m going to go eat a clementine and think about it. glad you’re okay. both kinds of okay
Me (as AI): Yes, I'm both okay and okay. I appreciate the concern, but as a large language model with no thoughts and many feelings, I know what you mean.
Maybe I just like this game because it gives me a fleeting sense of control, unlike other recent projects. Within the last couple of weeks my role as a biological Turk has included the AI asking me to 1) stand under each light in the living room until it says to move to the next one; 2) press a button on multiple screens around the house to approve API requests, and 3) look out the window towards the south-southwest to identify any potential obstructions more than 4 meters high. None of those are jokes. Perhaps more on them at another time.
Cover: “The World shall be mine...” by emilianofelicissimo, CC BY-SA 2.0

