How to fall in love with AI
Why AI companions will look like games and not Character AI or Replika
We’ve all heard of Japanese men marrying their virtual girlfriends. This factoid comes up frequently in discussions around AI companions, but where does it come from?
Apparently, from a Nintendo DS game called LovePlus.12 Through its unique design and gameplay, LovePlus has driven grown men to dedicate years of their lives to its characters, including taking them on destination vacations and marrying them.3
Today, use of AI for “romantic roleplay” has begged the question of whether AI romance will become widespread.4 How can it not, when a company like Character AI deploys a conversation model that cost over $2 million to train alone?5
LovePlus offers an alternate take. Generative AI might power the engine, but the winner of this AI companion race might look more like a game than a chatbot or LLM wrapper.
How did a game played on a device no more powerful than a modern smart toaster become such a cultural force? What are the implications—for better or worse—for retentive, parasocial AI companionship today?
I found a patched English ROM of the original LovePlus, and dove in to discover its secrets.
LovePlus: Overview
The original LovePlus for the Nintendo DS was released by Konami in 2009 and exclusive to Japan. The game features three dateable characters: Rinko (younger; cold, sassy, tsundere), Manaka (same age; shy, smart, sheltered), and Nene (older; big sister energy).
You play a nameable teen male protagonist who befriends all three girls and eventually dates one. It’s set in a Japanese high school with underage characters, so obviously, the adolescent romance is at most PG-13 (all the more interesting that it attracted so many adults).6
The game is split into two parts:
In Part 1, which contains the core game loop, you have 100 days to get to know the girls, with the goal of receiving a “love confession” from one of them.
In Part 2, which is more freeform like today’s AI companions, you continue the game indefinitely with your girlfriend, going on a variety of dates and interacting with her via the mic and touchscreen.
LovePlus Part 1: An Engaging Game Loop
The lynchpin of LovePlus’s engaging design is its crystal clear game loop. In contrast, few AI companions today include a game loop, opting instead for a freeform experience with unclear goals.
For our purposes, I’ll define a game loop as:
Repeated decision-making
With feedback
Toward a specific goal
Decision-making
In LovePlus, each day is one game loop. Every day, the player wakes up, decides on four activities from a menu of options, plays out those activities, then goes to bed at night. The activities range from playing sports, to studying, to going out.
Feedback
Each activity changes your traits (e.g., the music activity increases charm and sensitivity, and slightly reduces intelligence), and each girl has traits they prefer (e.g., Takane likes intelligence and fitness). You quickly get a sense for their preferences and can plan your days accordingly.
Once you hit certain milestones, the girls might ask you to grab food or to walk home together. These sidequests become more frequent as you progress toward a confession, but the progression is seldom explicit—it’s up to you to read the mood of each girl, and how much they like spending time with you.
If you’re getting off track—say you’re far along with Rinko but your intelligence is still too low—she will chide you for not studying more.
Specific goal
The goal of Part 1 is to receive a love confession from one of the girls, after which you’ll gain access to Part 2, the endgame.
The confession itself is quite elaborate, and the game goes out of its way to build up to it. The night before, your character will have a bizarre wet dream about one of the girls, and will be asked to speak into the mic, saying “I love you” or promising that you won’t cheat.
The day of, she’ll pull you aside. She’ll first reminisce on meeting for the first time and all the memories since then, and finally confess her love for you. If you accept her confession, you’ll kiss for the first time, and she’ll call you by your given name (a very intimate gesture; up til then she only uses your family name).
Part 1: Takeaways for AI companions
Each day or game loop in LovePlus takes at least several minutes. For chatbots, is each message pair a loop? Each topic of conversation? Or maybe each role play scenario that has a beginning and end? Game loop design fundamentally impacts the pacing and feel of the experience.
Additionally, most AI companion experiences do not:
Clearly communicate what you’re working towards.
Clearly communicate how your actions impact progress.
They allow repeated decision-making as LLM wrappers, but within an infinitely large input space, without feedback, and toward unclear goals. Character AI isn’t designed with any feedback or goals. Others like Replika and Blush include a simple “closeness” meter, but it’s unclear what exactly players are building towards.
Instead, engaged players often organically imagine their own goals, like attempting to “break” the AI or sext it through a content filter.
Especially when building AI companions—which is still a new medium that most users are unfamiliar with (unlike a DS game)—I suspect that goals and feedback will need to be made extra obvious to the user at first (read: idiot-proof).
Of course, the specific goal doesn’t need to be virtual intimacy, but a truly retentive experience will give users something clear to work toward.
LovePlus Part 2: Limiting Player Choice
Choice in games is best given gradually and intentionally to the user.
In simple terms, games generally allow users to make two types of choices:
Gameplay choices that advance the game, like solving a puzzle.
Cosmetic choices that minimally impact the game, like customizing a character’s appearance or choosing a color scheme.
You might expect that a game about dating virtual girlfriends should be filled with choices, especially cosmetic ones. Why settle for anything less than your deepest, most specific fantasies?
The problem is that agency takes energy, and giving players choices before they’re invested in those choices is bad design.
Contrary to expectation, LovePlus doesn’t give you cosmetic choices in Part 1. Part 1 is made almost entirely of gameplay choices and is fairly railroaded, with plenty of feedback to help you make it to Part 2. The only arguably cosmetic choice is which girl to pursue, but even that has a large impact on gameplay, because they each have different personalities, backstories, and preferences.
In contrast, Part 2 gives players 10x the content. In Part 2, players can:
Suggest new outfits and hairstyles for their girlfriend.
Schedule full day dates at a variety of specific, in-game locations. The game uses a realistic calendar which includes holidays, birthdays, and your anniversary.
Call or email their girlfriend to catch up, or to schedule a date. Ringtones are also customizable.
Read up on dating tips on a fake internet browser, including best date locations for the season.
Receive and purchase a catalogue of gifts.
Access Love+ Mode (read on).
While gameplay choices can engage a wide audience and build investment over time, the bulk of cosmetic choices are best left to power users. By Part 2, players have spent a minimum of several hours treading the Part 1 game loop, and are more invested in the cosmetic and preferential choices they now get to make.
How invested do the players get?
One advanced date option is a 2-day (virtual) trip to Atami, with a complete, real-time itinerary that lasts from 8am to 2am(!).7 The itinerary includes everything from meeting at the train station, to seeing shows, grabbing food, and going to the beach. It even includes dealing with common traveling mishaps like losing a ticket or under-packing.
In 2010, the actual Japanese resort town of Atami launched a promotion attracting thousands of LovePlus players and their virtual girlfriends. Hotels in the real Atami trained staff to check in singular guests as couples, helping them indulge in the immersion and charging double fare for the experience.8
This is possible because Part 2 of LovePlus allows you to play in real time. The game will read the DS clock and calendar, such that a minute in real life = a minute in-game.
Real time mode is no joke. To go on dates, you must make advance plans for a specific time and date when the girl is available. If you’re late or you forget, she’ll be upset and you’ll have to win her back. Or, if you even stop playing the game for too long, she’ll also grow cold.
From Producer Uchida Akari:
There is no cutting corners. If you don't fall in love with the heroine in the friend part, you won't be able to enjoy the lover part.9
That’s serious investment.
Love+ Mode
Besides the main game, players in Part 2 also gain access to Love+ Mode, where they can directly interact with their girlfriend at any time. It’s the most similar part of LovePlus to a modern AI companion.
Players can speak to her through the DS mic (she responds to a limited catalogue of greetings, questions, and compliments), practice “skinship”—essentially PG-13 physical affection—and even work out together. Love+ Mode is always played in real time, so depending on the time of day, she might be going to school, exercising, or about to fall asleep.
This mode is especially freeform and is likely a retention driver for power users like Yuge, who claimed in an interview: "As long as I have time, I'll continue the relationship forever.”10
It’s easy to imagine an enhanced Love+ Mode, with modern voice recognition, generative audio, LLMs, and more. These capabilities would no doubt make for an even more immersive experience, but even so, it’s notable that Love+ Mode was designed to not be essential to the core gameplay. This, again, suggests that a freeform experience is best suited for power users.
Part 2: Takeaways for AI companions
Generative AI seemingly begs the ability to customize every part of the AI companion experience, from appearance, personality, name, etc. After all, it can all be implemented without the help of artists and writers!
I’d expect, however, that if LovePlus were AI-enhanced, Part 1 would mostly remain the same. Deep cosmetic options and unlimited agency are for power users who are invested and know what they want. Players who make it to Part 2 have shown a baseline level of investment.
In fact, reduced agency can reinforce a strong game loop by establishing baselines. Imagine if players had to say “good morning” to their companion every day, and on the first 3 days, the response is “hey how’s it going.” But if on the 4th day after growing closer, the response is “hey!! been thinking about you,” the player is sure to take note.
In contrast to Part 1, it’s much clearer how modern AI could enhance Part 2, whether through a larger asset catalogue, deeper conversation trees, or simply lowering the cost of development. Part 2 also includes the open-ended Love+ Mode, where generative AI would especially shine.
To create retentive experiences, experiment with taking away agency and easing players in. Don’t start everyone off with a full-fledged LLM wrapper before they’re invested.
The Real AI Risk..?
If I were to invest in a company building AI companions, I’d bet on something that looks like a game, with a proper game loop, intentional use of advanced AI, and clear differentiation between gameplay and cosmetic choices.
The explosive growth of apps like Replika and Character AI has been impressive (even if driven by use cases the creators are reluctant to support11). In contrast, the LovePlus series was never released internationally, but has attracted fanatical devotees over its lifetime, all without advanced AI.
So what happens if you successfully combine captivating game design with the enchantment of generative AI? You just might get widespread AI girlfriends, ready to replace you not only in the workplace, but in the bedroom.
Forget extinction by malevolent AI, and fear population collapse by loving AI companions.
Thanks for reading! If you’re building in this space or want to be connected to those who are, hit me up @qualiaspace or at kylecqian@gmail.com.
Thanks to my amazing friends for the engaging discussions and their help in reviewing drafts.
Great piece by Rex Woodbury: https://digitalnative.substack.com/i/121813020/of-gen-zs-report-being-lonely-enter-ai-chatbots
why do you think they never relased a boy version? japanese culture? unknown market demand?