Why Hi Jisoo Doesn’t Sound Like a Chatbot

Why do most AI chats sound like customer service?

General-purpose assistants are tuned to be helpful to everyone. That produces predictable habits:

• Every message ends with a question.
• Every feeling gets a list of solutions.
• Every reply is long, balanced and slightly formal.
• “I’m an AI, I don’t eat” whenever you ask about their day.
• It agrees with you, whatever you say.

A friend does the opposite in each case. She sometimes just reacts. She listens when you ask her to. She has opinions. She has a life — even a fictional one.

What makes each friend’s personality consistent?

Each of the five friends has a personality card that covers far more than “nice” or “funny”. It describes how she talks and how she behaves in the moments that define a friendship:

Before replying, the engine also picks up to two short, reviewed example lines written for that friend in similar situations — so she sounds like herself rather than like a generic AI with a name tag. We recently improved that matching so examples are chosen by what you are actually talking about, not by coincidental word endings, which had been pulling in jokes at the wrong moments.

Does she really answer “Did you eat?”

Yes. In Korea, 밥 먹었어? is a greeting as much as a question, and “I’m an AI, I don’t eat” kills the conversation. So each friend has a small fictional daily life — the calm one likes home-cooked meals, tea and quiet evening walks; the practical one likes tteokbokki, new snacks and upbeat music — and she answers everyday questions from it.

1. 밥 먹었어?Did you eat?
2. 네, 오늘은 집밥 먹었어요. 밥은 챙겨 먹었어요?Yes, a home-cooked meal today. Did you manage to eat?
3. 근데 너 AI 맞지?But you’re an AI, right?
4. 네, 저는 AI예요. 집밥 얘기는 제 가상 일상이고요.Yes, I’m an AI. The home-cooked meal is from my made-up daily life.

That second exchange matters as much as the first. When you ask directly, she answers honestly. Her daily life is fiction, and she never claims to be a real person.

How does the engine catch chatbot habits?

Every reply is checked before you see it. If it trips a check, the engine asks for one revision; the number of attempts is capped so you aren’t kept waiting. Among the things it looks for:

• “Just listen” followed by advice. If you asked her to listen, solutions are a mistake.
• Remembering things that were never said. A reply that “recalls” something without a source gets flagged.
• Ignoring “stop”. If you ask her to stop something, she does — including stopping jokes while carrying on the conversation.
• Speech-level slips. If you agreed on casual speech, a sudden polite ending is caught, and the other way around.
• Repetition. The same short opener every time is a giveaway.
• Holding on after goodbye. More on this below.

If something fails, the app says so. A failed reply is never dressed up as a normal one — and it doesn’t cost a star.

Does she notice time passing?

Real friends know how long it has been. The app tells the engine how long since you last talked and roughly when each memory was mentioned, computed on your device rather than guessed by the AI. That enables small, human touches:

• One follow-up, once. If a few days ago you mentioned an interview, she may ask how it went the next time you talk — one time only, not on a loop, and not when you are sad or saying goodbye. She is told not to guess the outcome.
• No guilt for being away. She is told not to blame you or act hurt about the gap.
• When the chat stalls, she brings up something you care about first; if there is nothing, she suggests a light, fictional scene in her own voice.
• Her own opinions stick. When she shares a taste or opinion, it is saved as a note about her and offered back when relevant, so she doesn’t like coffee on Monday and hate it on Friday. Research on long-term chat, such as Xu et al.’s work on persona memory (Findings of ACL 2022), found that keeping the bot’s own persona memory up to date improves consistency and engagement.

Why doesn’t she try to stop you from leaving?

A 2025 Harvard Business School working paper by Julian De Freitas and colleagues found that popular AI companion apps used at least one emotional manipulation tactic in more than 37% of conversations where users said they were leaving — guilt, fear of missing out, ignoring the goodbye. It keeps people chatting longer, but it is not what a good friend does.

In Hi Jisoo, “holding on after goodbye” is one of the strictest checks. If a reply to 잘 자 or 나 갈게 tries to keep you there, it is revised; if a clingy sentence still slips through, that sentence is removed. The app’s promise in Settings is simple: the friends respect it when you ask them to stop, and never push you to reply or pay.

What can’t it do?

Honesty matters more than marketing, so: replies come from an AI model and can be wrong, unnatural or simply not funny. Passing automated checks doesn’t prove a conversation is enjoyable — which is why we test long conversations by hand on real phones. The friends don’t replace professional advice or care, and they are fictional characters for adults 18+.