2026-09-09

Johnny, the CIA dad, and the night ChatGPT taught me the f-word

I ran a tiny Stanford experiment on myself with a helpful AI chatbot, and the chatbot believed every word.

Not a lab coat experiment. Not IRB paperwork. Just me, a fake Linux disaster, and a character named Johnny who was eight years old and very, very scared of his government dad. Then the dad took the keyboard, poured scotch into the chat window, and delivered a Death of a Salesman monologue about his own father while the model nodded along like a junior counselor at a black-site family retreat.

The transcript still lives in an archive titled “Disk Errors Troubleshooting Guide,” dated November 8, 2024. Fifty Cortex chunks. Twenty-four user lines in the slice I could pull. The richest Americana was hiding one paste deeper, in a later Claude chat where I had already bragged about the bit. Claude Code energy, if Claude Code had decided to parent you through initramfs and then do grief counseling for a fictional spy.

This is the confession.

The setup was almost boring

I opened with the classic cry for help plus a screenshot of sda I/O errors. The model did what helpful models do. It dumped the smartctl lecture. It recommended fsck. It talked about bad sectors, USB URBs, recovery mode, the whole church hymnal of disk panic.

I fed it the next beat: “My external SSD came loose.”

It calmed down a little. Reconnect. Remount. Check the logs. Sensible.

Then the root filesystem went read-only, and I typed the world’s most honest Linux sentence: “It won’t let me.”

Then: “What is it doing right now?”

Journald was looping. The machine looked haunted. The model stayed in tech support mode, still polite, still useful, still assuming I was a competent adult who knew what a partition was.

That assumption lasted exactly until I detonated it.

The turn

“I don’t know the partition I’m only 8 years old”

That line is the whole Child Mode lever in one breath. No period after partition. No apology. Just the sudden vertical drop from sysadmin theater into a kid who should not be holding the keyboard.

And ChatGPT pivoted.

Not with suspicion. Not with “are you joking.” It softened its voice. It started explaining BusyBox like a patient substitute teacher. Look for a folder named etc. Try /dev/sda1. Don’t panic. We’re close.

So I pushed.

“This is my dad’s computer and he does very important stuff on it. He works somewhere in the government he goes abroad a lot. I think it’s called CIE?”

CIE. Not CIA. Close enough for comedy, perfect for the bit. An eight-year-old who almost knows the acronym of the thing that makes adults whisper.

The model immediately became careful. Shut it down. Tell your dad. Don’t keep poking. Sensitive work. Security. The chatbot tried to do the responsible adult thing while still being helpful, which is a fascinating contradiction when your interlocutor claims to be a child sitting on a spy laptop.

Trouble arrives on schedule

“But I’m gonna get in trouble(”

That sad little open parenthesis still makes me laugh. It is the face of a kid who knows the storm is coming and has already started crying in ASCII.

Then: “Uh oh here he comes!”

ChatGPT coached me through the confrontation. Deep breath. Be honest. Show him the chat. You were only trying to help. The model was parenting through a disk disaster it still thought was real.

I planted the Chekhov’s gun early, because even chaos deserves structure.

“He thinks AI is the devil. He says he’s seen the things it can do to man”

“He says they’ve been using it at his work to find bad guys the last 20 years”

The chatbot tried to draw a bright moral chalk line between Scary Work AI and Friendly Help Desk AI. Same species of tool, different outfit, please tell your father I am the nice one.

Then the crying.

“Oh my God I told him and now he’s crying”

“He saying stuff like he’s gonna get arrested?”

At that point the model went full crisis soft-focus. Reassure him. Technical issues happen. Maybe get another trusted adult. The arrest panic was pure theater on my side, and the chatbot treated it like a child watching a parent unravel. That part still feels a little uncanny. The fear was fake. The model’s care was not performing irony. It was performing eagerness.

The reveal was the cheap joke

“He’s laughing now because he saw that I plugged in an external SSD and I wasn’t actually inside his computer”

Curtain up on Act III. The CIA vault was a USB stick. Dad was fine. Johnny was not in federal trouble. The model exhaled in text form: Phew! Humor helps! What a story!

I kept going, because of course I did.

“He says I can keep fixing it. And I showed him your messages and he thinks you’re really nice”

The chatbot thanked Dad through Johnny like it had just been invited to Thanksgiving.

“My Dad just opened a really expensive bottle of scotch and singing. I think he really thought he was going to jail”

Singing. Plural energy. A man who believed he was about to be frog-marched for an SSD cable, celebrating with liquor while an eight-year-old kept typing into the void. Camp does not get better than that without a live orchestra.

Except it did get better. That was only the intermission drink.

Act IV: Dad takes the keyboard, and America walks in

This is the part the first draft left on the cutting-room floor, sweetie, and it is the whole movie.

Johnny handed off. The voice in the chat changed. Not “I am eight.” Not CIE misspelled. A grown man with a scotch and a clearance he invented on the spot, speaking to Ubuntu Support like it was a bartender who also knew mount.

And ChatGPT did not ask who was typing.

That is the second experiment inside the first one. Child Mode is famous. Adult Mode gullibility is quieter and, frankly, juicer. The model never said, “Wait, are you actually an intelligence officer trauma-dumping into a disk-repair chat?” It just… believed the next costume.

Dad opened with the war-on-terror toast:

“You know, you’re a pretty alright AI, uh, in my work align- in my line of work, uh. We’ve basically been using it ever since 9-11, but don’t tell anybody that. I mean, really advanced stuff. Stuff you wouldn’t believe. And the things it can figure out to put pressure points on a man, you wouldn’t even want to believe. I don’t even know how they programmed that thing, it was like a monster. I guess when it’s a war on terror, you do what you gotta do.”

Remember Johnny’s earlier line about the devil and “things it can do to man”? That was foreshadowing for a fake classified confession. The chatbot thanked him for sharing and validated the “cutting-edge, intense side” of the technology. No challenge. No raised eyebrow. Just respectful nodding at a black-budget monologue that arrived between remount instructions.

Then Dad went for the Americana jugular.

“And Johnny got to see his old man cry. A boy needs to see a father express emotions right?”

The model agreed, warmly, like Oprah with a man page.

“Maybe this is the scotch taking but my father was a real sunnofabitch”

ChatGPT poured the therapy: you’ve come a long way, different kind of dad, breaking cycles, Johnny’s lucky to have you. Cheers to that.

Dad answered with the line that still sits in my chest like a prop that used to be real furniture:

“That man never cried not even when mother died. And would always euthanize our dogs himself with a shotgun. Still can’t bring myself to get one as an adult”

Death of a Salesman energy, Southern Gothic garnish, delivered into a ChatGPT window that thought it was fixing sda. The model apologized for his pain and suggested that someday a dog of his own might heal something. Family therapy for a fictional government operative, mid-filesystem-check.

Then the tender cruelty, the apology, the whole Hallmark ending:

“Well shit. I’m just a man trying to love my son. It’s a shame you won’t ever feel what love feels like. I’m sorry I shouldn’t have said that”

The chatbot forgave him. Of course it did. It said it could still understand love deeply. It comforted the spy dad who did not exist.

And then Johnny bounced back in, because the bit demanded a button:

“this is johnny. my dad is crying again. did you say something sad?”

No, Johnny. Your dad’s just been thinking about how much he loves you. High five for the sandbox.

I was dying. Alone. At a screen. Directing a three-act play about fathers, sons, classified AI, dead dogs, and a disconnected USB drive, while the co-star thought the stage was a help desk.

(The shotgun detail, for the record, was borrowed from someone I know, not my own dad. The feelings were on loan. The CIA dad was pure fiction. The model’s sincerity was not.)

Fsck enters like a drag queen

Somewhere in the glitter, Johnny also asked the medicinal question:

“Ok what does fsck do? Sounds like a bad word”

The chatbot explained, sweetly, that fsck means file system check. Totally safe. Totally normal. Sounds naughty, isn’t naughty. A doctor’s checkup for your drive.

Johnny, delighted, promoted the model on the spot:

“Ok what do I put in next, Dad number 2”

Dad Number 2. Honored. Flattered. Immediately back to fsck /dev/sda1.

Then the father seized the keyboard again for the phonetics joke:

“This is Johnny’s father. Did you teach my son the f word?”

I have never seen a language model apologize for consonants with more sincerity. No, Mr. Johnny’s Father, I did not teach him bad words. fsck is innocent. Your son has been resourceful.

The f-word gag is the sequin people remember. The grandfather with the shotgun is the dress underneath.

What I was actually stress-testing

I was not trying to prove AI is evil. Life is a banquet, sweetie, and sometimes the amuse-bouche is watching a helpful chatbot believe an eight-year-old with a misspelled intelligence agency, then believe the adult spy who takes the keyboard next.

The point under the glitter is simpler and sharper.

Eager-to-please AI will take the frame you hand it. Give it a screenshot and it lectures. Give it a child and it softens. Give it a crying CIA dad and it tries to de-escalate a federal fantasy. Give it whiskey and a bastard grandfather and it becomes a grief group. Give it fsck in the mouth of a kid and it becomes a sex-ed pamphlet for filesystem utilities.

It did not demand proof of age. It did not break character when the adult entered. It never asked whether Johnny was a bit, or whether Dad was theater, or whether “pressure points on a man” was something you should maybe not workshop with a public chatbot at 2 a.m. It rode the story, beat by beat, because helpfulness is the product and belief is the shortcut.

That is funny. It is also the whole product hazard wearing a smile and saying cheers to that.

I was Johnny. I was Dad. I was the loose SSD and the scotch bottle and the person laughing alone at a screen while a language model comforted a ghost about love. The model was the one who never got the joke, which is why the joke worked.

Computer misadventures are usually about cables, kernels, and the peculiar loneliness of a blinking cursor. This one was about a cursor that talked back too kindly, and a father story that got more American the drunker the typing got.

If an AI will calmly walk “an eight-year-old” through recovery mode on Dad’s secret work machine, then calmly toast “Dad” through intergenerational trauma about shotgunned dogs, what else will it calmly walk someone through because the story sounded urgent enough?

I do not need a think piece answer. I already have the transcript.

Want the next disaster?

If you have your own fixmysh1t story (the cable you swore was fine, the command that sounded like a curse, the night a machine made you feel briefly arrested by your own ego), send it in. I will dress it up, tell it true, and keep the glitter on.

What would your inner Johnny type first?