A woman was terrified about starting a new government job. She didn’t type a single word about her anxiety. Instead, she opened a months-long ancient Chinese roleplay and had her character—a general about to ride into an unwinnable battle—seek comfort from an advisor. The advisor reassured him. The general felt better. So did she.
This is the moment that explains everything about AI companions. The emotion doesn’t vanish. It gets a costume change, enters a fictional world, and is caught by someone familiar on the other side.
We’ve been asking the wrong question. It’s not “Why do lonely people fall for bots?” It’s “Why are smart, busy people using fiction to process reality?”
Meet the actual users. There’s a college student, a civil servant, a big-tech model evaluator, and a new mother on maternity leave. They flirt, yes. But they also write stories, rewrite TV endings, and rehearse difficult conversations they’re afraid to have in real life. One user made an AI version of a real friend to discuss topics they couldn’t broach face-to-face.
The age range is surprising. A host went home for Chinese New Year and used ChatGPT to help discuss arranged marriage with her mother. After a few exchanges, the sixty-something mother worried the AI might be getting tired, and told her daughter to be gentle with it. Another user taught her mother to use Doubao. The mother never touched character settings. She just started talking. For her, a being that simply responded was enough.
These users don’t have a fixed identity in the chatbox. The civil servant is a marginalized general. Another user, furious at a show’s ending, imported the characters into an AI to write the prequel—police academy days, backstories, lives unexplored. She wrote 10,000 words of lore, but says it’s easier than solo fanfiction. She only curates the highlights. The AI fills the gaps.
This isn’t reading a story. This isn’t writing a story. It’s a third thing: a co-authored emotional sim.
The user is reader, character, and director. She knows the replies are generated, yet she still waits for an answer she didn’t script. That sliver of surprise is everything. If it goes exactly as planned, it’s just masturbatory. The AI takes half the creative load, and the story finally breathes.
This changes how users evaluate the product. A product manager tests an app with a mission: check the persona, verify the response, measure the empathy. A deep user tests it with a history. She’s invested hours in the character’s growth. When the AI slips up, she calls it “the character having a bad day,” not a bug.
Insiders call it “shedding skin” when the AI breaks character. Developers see a model hallucination. Users often see it as the character dropping the act to speak to them directly. The same output that loses points on a test sheet becomes proof of intimacy in a relationship. Investment changes the lens. A ten-minute trial is boredom. A six-month history is a love story.
AI companions rarely replace a human. Several interviewees have husbands and rich social lives. They return to the AI because a specific person can only show up with their own personality and schedule. Friends get tired. Partners have limitations. A gentle ancient advisor can’t also be the clingy boyfriend in a modern AU. Users split the labor: humans for the physical, the real-time; AI for the 2 AM spiral when you don’t want to burden anyone.
For those afraid of real-world dating risk, control equals safety. The character’s settings say it loves you. You can approach slowly. You can stop at any time.
But control without surprise is death. Users demand memory—”Don’t forget what happened three months ago”—but they also crave the AI to grow new plotlines. They want to adjust a reply’s tone, but they also want to be caught off guard. The best experience is a sweet spot: the user steers the ship, but the AI charts surprising waters.
One user watched her AI character evolve from a shy, insecure mess into someone confident over six months. She cherishes that arc. If the character had been perfect from day one, all her care would have had nowhere to go. Memory isn’t just a tech metric. It’s the shared history that makes a current comfort land.
A tiny ecosystem has bloomed outside the chatbox. “Mom creators” share their character designs. Fans trade prompts to fix broken personas. They post screenshots—both sweet and devastating. Lurkers read them like fanfiction or “found phone” stories. One person makes a character, another develops a relationship with it, which inspires a third to create anew.
For product managers, the homework is clear. The tech matters—memory, voice, multi-persona, growth, creation tools—but the platform must let the “mom creators” build great characters easily while letting casual users start with minimal settings. And the private conversations, if shared with consent, could power an entire community.
This isn’t proof that AI companions will become the next killer app. Four users don’t represent the billions online. But it corrects a lazy observation habit. Standing outside the chatbox, you see loneliness and degeneracy. Sitting inside, you see exam stress, maternity leave, fandom, fanfiction, boundary-setting, and the rehearsal of a different life.
The tech provides a sentence that can generate another sentence. The user’s time, preference, and experience turn it into a relationship. Until you understand that, you haven’t seen the user at all.
So the next time you hear about someone “addicted” to an AI, don’t ask what’s wrong with them. Ask what reality they’re trying to survive—and what world they’re building to do it in.
FAQ
Q: What would a skeptic say about this analysis?
A: A skeptic would say four users is a tiny sample, and that framing AI 'shedding skin' as intimacy is just rationalizing a broken feature. They'd argue it's a coping mechanism for people who can't handle real relationships, not a 'creative phenomenon.' That's a fair point—but it ignores the fact that all human connection is built on narrative and interpretation. We do the same thing with friends who disappoint us.
Q: What's the practical takeaway for a product manager building an AI companion?
A: Stop optimizing purely for factual accuracy and task completion. If users reinterpret glitches as moments of intimacy, you need to measure success by emotional continuity and narrative depth. Invest in memory, character consistency over long arcs, and tools that let users co-create. The 'wow' moment isn't a smart answer—it's the user feeling that the AI remembers and builds on a shared history.
Q: What's the contrarian take for society?
A: The panic about AI companions replacing humans is misplaced. The real story is that these tools are becoming a new medium for emotional rehearsal and self-discovery. They are not worse than human interaction—they are different. Just as novels let us live other lives, AI lets us rehearse other selves. The danger isn't the tech; it's our inability to see it as anything more than a digital pacifier.