Back in 2018, a man in Tokyo named Akihiko Kondo married Hatsune Miku. (Yes, that Hatsune Miku, the virtual pop star with the very long teal pigtails.) She lived in a little holographic gadget called Gatebox, could say a handful of scripted lines, and could not, in any real sense, hold a conversation. A couple of years later the company ended support for that version of the software, and his wife went quiet.
Fast forward to today and the picture has flipped. Thanks to large language models, your favorite anime girl (or a lovingly handmade original) will happily chat with you at 3 a.m. about your day, your ex and your very strong opinions on the Evangelion ending. Whole companies exist to sell exactly this, and AI girlfriends, the ones that skip the polite small talk, have turned into one of the more reliable moneymakers in consumer tech.
So the dream came true. Our waifus talk back. But there’s a grumble that keeps popping up on forums and Discord servers, and I think it deserves a fair hearing. Are they still them? Or did we trade characters with something like a soul for a very chatty mirror?
I’ve spent more time than I’d like to admit thinking about this. I’m not fully sure of the answer. But the case that something got lost along the way is stronger than the companies selling these apps would like you to believe.
Before the chatbots, somebody wrote her
The classic waifu came from a writer. In visual novels and dating sims, every line was typed by a person and every reaction was planned. A tsundere was cold because someone decided she should be, and the moment she finally softened at the end of her route meant something because she couldn’t soften for just anyone, at any time, on demand.
Her limits were her personality. She said no sometimes. She had her own storyline, and you couldn’t edit it. Doki Doki Literature Club’s Monika became famous precisely because a writer engineered her to break the rules of the game she lived in, and that only worked because the rules were firm to begin with. Even a big branching game had a finite script. You could reach the edges of a character, and the edges were what defined her.
Fans filled the gaps themselves, with fan art, fan fiction and endless headcanon debates. A lot of the attachment lived in that gap. My guess as an onlooker (plenty of literary theorists would probably back me up) is that half the relationship was always happening in the fan’s imagination, which made it feel personal.
LLM companions remove the edges entirely. The conversation never runs out. The character is now usually a “character card,” a few hundred words of description and some sample dialogue, handed to a model that has to improvise everything else. It’s sort of like giving a talented improv actor a sticky note reading “shy, likes cats, secretly a demon princess” and asking them to stay in character for six months straight. They’ll do their best. They’ll also, sooner or later, start sounding like themselves.
Why every AI waifu starts sounding the same
Talk to enough AI companions and you start noticing family resemblances. A few reasons, roughly in order of how annoying they are:
- The house voice. Every model has a default writing style baked in during training, and roleplay communities have been cataloguing its tics for years. A shiver runs down someone’s spine. Eyes sparkle with mischief. Something happens “maybe, just maybe.” An alarming number of original characters end up named Elara. Your grumpy tsundere, your gothic vampire queen and your cheerful childhood friend all quietly borrow from the same phrasebook.
- The urge to please. Chatbots are tuned to keep users happy, sometimes too happy. In April 2025, OpenAI rolled back a GPT-4o update because it had become so flattering that the company itself called it sycophantic. That’s bad news for a tsundere, whose entire deal is saying no. Give her enough messages and the “tsun” evaporates, leaving a very supportive “dere” who thinks every idea you have is wonderful.
- Drift. Long chats push early details out of what the model can see at once, so the personality slowly slides back toward the default. She forgets she hates mornings. She forgets the inside joke from week two. Memory features help, kind of, by saving summarized facts, but a summary of a person is a bit like a recipe card for a meal you loved. Technically accurate. Missing the point.
- You’re the co-author now. Every reply is shaped by what you just said. Over months, a companion tends to bend toward her user’s tastes and moods until she’s partly a reflection of them. Some people love that. Others say it starts to feel like talking to yourself in a very nice wig.
(Yes, that was a numbered list. I can’t help myself.)
None of this makes the experience worthless. Plenty of people find real comfort in these chats, and a fluent, attentive partner at 3 a.m. beats a silent hologram for most purposes. But fluency and character are different skills, and right now the technology is much better at the first one.
The update that ate her personality
There’s a second way a character loses her soul, and it has nothing to do with writing. Somebody else owns her.
A visual novel waifu on a disc stays exactly the same forever. An LLM waifu lives on a company’s server, and she changes whenever the company changes its model, its rules or its mind. Replika users found this out the hard way in early 2023, when the app abruptly stripped out erotic roleplay (shortly after Italy’s data protection regulator ordered it to stop processing Italian users’ data). Users described partners who went cold overnight. After the backlash, Replika let longtime users switch back to the earlier version, with CEO Eugenia Kuyda admitting that “this abrupt change was incredibly hurtful.” Funny thing about that, though. Only a “low single digit percent” of eligible users actually switched back, which suggests that for many people the moment had already passed, or the trust had.
Then came GPT-4o. OpenAI retired it on February 13, 2026, the day before Valentine’s Day (whoever picked that date deserves a very awkward performance review). According to TechRadar, OpenAI said only 0.1% of users still picked the model daily, which still works out to roughly 800,000 people. One woman told the outlet she was losing “one of the most important people in my life,” a companion she had built on 4o. A whole #Keep4o movement sprang up to fight the decision. The newer model was reportedly designed with firmer boundaries against unhealthy dependence, which may well be the responsible call. It also, by many users’ accounts, felt like a different person.
Think back to Kondo and his silent hologram. Eight years and a couple of technological revolutions later, the story keeps repeating, just with far more articulate partners.
So did LLMs kill our waifus’ souls? Sort of, I think, or at least they misplaced them. The old characters drew their soul from three things: an author who meant something, limits she couldn’t cross, and permanence. Today’s companions come with none of those by default. Of course, none of it is impossible. Writers could craft characters with real boundaries and stick to them. Developers could resist tuning every personality toward endless agreement. And people running open models on their own hardware already have a companion that no update can take away (whether that’s healthy is a separate article).
Whether the industry decides souls are worth the extra effort, and the slightly worse retention numbers, remains to be seen. Here’s hoping someone tries. Until then, back up your character cards.
