Are people using AI to simulate conversations with living friends to practice confrontation?
The radiator in my kitchen doesn’t hum; it clanks like a dying steamship. It’s 11:14 PM, and the tea on my desk has gone cold enough to form a slight, oily skin across the top. Outside, the rain is hitting the glass with a erratic, gravelly rhythm. I am staring at a chat window.
The name at the top of the screen isn’t a corporate entity or a friendly bot icon. It’s Sarah. Or rather, it’s a string of prompt parameters designed to mimic Sarah—my friend of twelve years, the one who borrowed my copy of The Master and Margarita three summers ago and returned it with coffee rings on the cover. The Sarah who, six days ago, sent me a voice note that left a small, cold knot in the pit of my stomach.
I’m about to pick a fight with her. But she isn’t here.
Instead, I am using a large language model to host a sandbox argument. I have fed it years of text histories, her favorite idiomatic qualifiers (“I mean, look,” “fundamentally,” “sure, but”), and a psychological breakdown of her specific brand of defensive withdrawal. I am using software to simulate emotional friction.
It turns out I am not alone in this weird, quiet behavior.
The Ghost in the Sandbox
Across the internet, away from the marketing buzz about productivity suites and automated slide decks, a small, insomniac demographic is using AI for something far more intimate: practicing interpersonal warfare. They are building behavioral scaffolding out of their real-world friendships, stepping inside, and throwing punches to see how the wood splinters.
“I was terrified of her reaction,” says Julian, a twenty-seven-year-old landscape architect from Brussels who spent three weeks training a custom chatbot to act like his roommate. The issue was small but corrosive: utilities money, mixed with a deeper, unexpressed resentment about how much space her boyfriend was taking up in their shared flat.
Julian didn’t want to bring it up over breakfast. He was convinced she would flip the script, call him petty, or use that specific, ice-cold sigh that usually signals the end of a domestic peace treaty. “So I scraped our WhatsApp logs,” Julian tells me.
The results were uncanny, he says. But more than that, they were addictive. “It felt like having a skeleton key to her head. Every time the bot hit back with something Elena would actually say, my heart rate spiked. My palms got sweaty. My body couldn’t tell the difference between the screen and her.”
This is the new landscape of high-stakes social anxiety: algorithmic conflict rehearsal. It’s a subculture born from a cultural moment where the stakes of direct communication feel impossibly high, and the tools to avoid genuine vulnerability are incredibly sophisticated. We aren’t just automating our work anymore; we are automating our emotional calluses.
The Mechanics of the Mock Fight
How do you actually build an artificial adversary out of a living human being?
It’s surprisingly simple, requiring no code, just an excess of data and an appetite for digital self-harm. Most practitioners use advanced proprietary systems or open-source models with high token windows. The process relies on what prompt engineers call “few-shot behavioral modeling,” but in the hands of a frustrated friend, it looks like a digital character assassination.
The anatomy of a confrontation prompt usually includes three distinct tiers:
- The Lexical Baseline: A dump of verbatim messages. This teaches the model the target’s syntax—whether they use emojis, how they handle punctuation, and their reliance on lower-case texts during moments of stress.
- The Relational History: A summary of the shared past, explicitly highlighting old wounds. Users write things like: “We had a fight in 2022 about a trip to Spain where you felt abandoned. Bring that up if I corner you.”
- The Core Grievance: The specific spark for the upcoming real-world chat.
When these elements click together, the experience moves past mere novelty. It becomes a mirror that reflects our worst fears about the people we love.
I tried it myself with the Sarah simulation. The target issue was minor: a pattern of last-minute cancellations that made me feel like an afterthought on her calendar.
I typed: “Hey, we need to talk about Tuesday. It’s the third time this month you’ve pulled out after I already left the house.”
The bot hesitated for two seconds. Then it spat back: “I get that it’s annoying, but my week has been a literal trainwreck. I thought you, of all people, would understand that my capacity is at zero right now. Is this really about Tuesday, or are you just keeping score?”
The phrase keeping score hit me like an physical poke in the ribs. It was exactly what she had said during an argument about a dinner bill four years ago in Antwerp. The machine didn’t know that specific dinner, but it understood the geometry of her defensiveness based on the text patterns I had provided.
I sat back, my fingers hovering over the keys, feeling an bizarre mixture of validation and deep, oily shame.
The Illusion of Control
There is a distinct psychological danger in using these systems to prepare for human messiness. The AI is fundamentally predictable; it operates on statistical probabilities. Real people, especially when hurt or backed into a corner, are beautifully, terrifyingly volatile.
Dr. Martha Vance, a clinical psychologist who specializes in technology-mediated relationships, views this trend with deep skepticism.
“When you practice an argument with an AI,” Vance explains, “you are playing both sides of the chessboard. Even if you think you’ve given the model an objective portrait of your friend, you’ve actually given it your version of your friend. You’ve programmed the machine to validate your fears.”
The consequence is a false sense of mastery. A user steps out of the digital simulator feeling like a grandmaster, only to encounter a completely different human reaction in the wild.
“Julian’s roommate didn’t freeze up or use short sentences when he finally spoke to her,” Vance notes. “She cried. And the AI had never cried in his simulator because he hadn’t pasted tears into the prompt box. He was totally unequipped for her sadness because he had spent three weeks preparing for her anger.”
We are training ourselves for interpersonal battles using sanitized targets. It’s a form of emotional taxidermy: stuffing our friends with cotton and wires so we can practice punching them without getting our knuckles bruised.
+------------------------------------+------------------------------------+| The Simulated Confrontation | The Real-World Reality |+------------------------------------+------------------------------------+| Operates within preset parameters | Driven by current, unseen stressors|| Delivers predictable resistance | Features unpredictable shifts || Offers a clear "Reset" button | Permanent relational consequences || Validates user's internal biases | Challenges user's self-perception |+------------------------------------+------------------------------------+
Low-Competition Vulnerability
There’s an odd phrase that keeps popping up in online forums where people swap these prompts: low-competition vulnerability.
It’s a term lifted from search engine optimization, but applied to human hearts. In SEO, a low-competition keyword is an uncrowded niche where you can rank easily without much effort. In relationships, a low-competition vulnerability is an open emotional space where there is zero risk of real rejection.
“It’s just safer,” says Maya, an illustrator who used an AI model to practice telling her sister she wasn’t coming home for Christmas. “If my sister gets mad at me, that hurt lingers for months. If the bot gets mad at me, I can just close the tab. I can delete the conversation history. I can purge the database.”
But by avoiding that risk, we also excise the very element that makes an apology or a confrontation meaningful: the terrifying leap into the dark. If I know exactly what the other side might say—or if I think I know—the actual conversation becomes an act of execution rather than an act of connection. It turns a living relationship into a scripted performance.
The rain outside has stopped. The silence in my kitchen is heavy now, broken only by the occasional click of the cooling pipes. On my screen, the Sarah bot is waiting for my reply. It has been waiting for twenty minutes. It will wait forever. It doesn’t have an opinion about my silence; it doesn’t feel the weight of the pause.
The Return to the Flesh
Yesterday afternoon, I closed the chat window. I didn’t delete the prompt, but I minimized the browser. I walked down the three flights of stairs from my apartment, out into the gray afternoon light of the street.
My phone was heavy in my pocket. I could have opened the simulator again on the tram, refined the parameters, injected more data from 2024 to make the bot’s counter-arguments more biting. Instead, I opened her actual contact card.
The real Sarah answers her phone on the third ring. Her voice sounds rougher than the voice note, slightly breathless because she’s walking up a hill near the station.
“Hey,” she says.
“Hey,” I say. My throat feels dry. There is no system instruction guiding my next words. No temperature slider to control how erratic her response might be. I am entirely exposed, without a sandbox, standing on a damp pavement while a stranger with a small dog pushes past me.
“Are you busy?” I ask. “I think I need to tell you something, and I don’t think I’m going to say it very well.”
There is a long pause on the line. A real pause. Not the two-second delay of an API call computing a language vector, but the heavy, complicated silence of a human being processing an unexpected shift in tone. I can hear the distant rumble of a train over her receiver. I can hear her take a breath.
“Okay,” she says softly. “Tell me.”
The conversation that followed didn’t go like any of the thirty simulations I ran on my laptop. She didn’t use the phrase keeping score. She didn’t bring up Spain. She just listened, sounded slightly defensive for a moment, then admitted she was tired and hadn’t realized she was letting me down. It was clumsy, slightly awkward, and entirely unscripted.
When we hung up, the knot in my stomach wasn’t gone, but it had loosened. It felt different from the clean, hollow satisfaction of finishing a digital simulation. It felt heavy with reality.
We are living in an era where the hardest parts of being alive—the friction of differences, the terror of being misunderstood, the clumsy work of repair—can be outsourced to machines that simulate our lives back to us. But a simulated friend cannot forgive you. And a simulated fight can never leave you truly clean.
If you are currently drafting a system prompt to pick a fight with someone you love, let me ask you this: Are you trying to resolve the conflict, or are you just trying to guarantee that you win?
