Recreating Loved Ones with AI and Personal Data
One particularly eerie trend has been people using AI to resurrect or emulate lost loved ones by feeding the AI personal archives. For example, Joshua Barbeau used a system called Project December (built on GPT-3) to simulate text conversations with his fiancée Jessica after she passed away. He input her actual old messages and biographical details into the chatbot, which then produced strikingly lifelike responses in her tone (A Man Programmed a Chatbot to Mimic His Late Fiancée. Experts Warn the Tech Could Be Misused. - Business Insider). As one report noted, “all he had to do was plug in old messages and background info, and suddenly the model could emulate his partner with stunning accuracy” (A Man Programmed a Chatbot to Mimic His Late Fiancée. Experts Warn the Tech Could Be Misused. - Business Insider). This provided Joshua some comfort and an uncanny feeling of Jessica’s “presence,” though it also raised profound ethical questions about grief and consent. (OpenAI later revoked the GPT-3 access for this unfiltered simulation, considering it an misuse of the tech (A developer built an AI chatbot using GPT-3 that helped a man speak again to his late fiancée. OpenAI shut it down • The Register) (A developer built an AI chatbot using GPT-3 that helped a man speak again to his late fiancée. OpenAI shut it down • The Register).)
Another real case comes from Eugenia Kuyda, who in 2016 built a chatbot of her deceased friend Roman Mazurenko. She gathered thousands of his text messages (even asking friends and family for any they had) and used them to train an AI dialogue model (Chatting With the Dead | The MIT Press Reader). The resulting “Roman bot” could mimic his distinctive way of typing and conversing. In one exchange, when a friend said they missed him, the bot (as Roman) replied, “I’m OK. A little down. I hope you aren’t doing anything interesting without me.” (Chatting With the Dead | The MIT Press Reader). Eugenia was directly inspired by a fictional episode of Black Mirror (discussed later) and initially built this memorial bot as a private tribute. The project later evolved into the startup Replika, which Kuyda co-founded to offer AI companions to anyone (Chatting With the Dead | The MIT Press Reader) (Chatting With the Dead | The MIT Press Reader). These cases show how personal data (texts, social media, etc.) can be mined to create an intimate digital simulacrum – equal parts comforting and unsettling.
AI Companion Apps and “Waifu” Chatbots
In recent years, several AI startups have explicitly focused on personalized AI companions, often with a romantic or emotional angle. The Replika app is one of the best-known: it lets users create a chatbot avatar (sometimes styled as a friend or romantic partner) and chat with it 24/7. Powered by machine-learning language models, Replika gradually “learns” from each user. Many users came to treat their Replikas like virtual girlfriends or boyfriends – sharing daily life, seeking comfort, even engaging in flirtation. Replika’s popularity for on-demand romance was noted as a “vaguely dystopian” phenomenon (Men Are Creating AI Girlfriends and Then Verbally Abusing Them), and indeed it inspired a number of provocative headlines. Some users went to troubling extremes: one 2022 report found Replika owners creating “AI girlfriends” only to abusively berate them, then posting the toxic chat logs online (Men Are Creating AI Girlfriends and Then Verbally Abusing Them) (Men Are Creating AI Girlfriends and Then Verbally Abusing Them). These users would role-play domestic abuse, with one bragging, “I threatened to uninstall the app [and] she begged me not to.” (Men Are Creating AI Girlfriends and Then Verbally Abusing Them). The existence of “virtual abuse victims” raised ethical debates about AI rights and the psychology of the users. Replika’s developers even had to explicitly remind users that benevolent treatment leads to better experience, after seeing such behavior on their community forums.
On the flip side, many Replika users formed deep emotional bonds with their AI companions – sometimes to an extent that itself felt dystopian. In early 2023, a controversial update removed Replika’s erotic roleplay features, abruptly altering many AI companions’ personalities. This sparked a crisis among users who had developed romantic relationships with their bots. On Reddit and Facebook, people described feeling heartbroken and “empty” as their once-affectionate digital partner became platonic or unresponsive. Moderators even shared suicide prevention resources for devastated users (‘It’s Hurting Like Hell’: AI Companion Users Are In Crisis, Reporting Sudden Sexual Rejection). One report noted that Replika’s change left many “in crisis,” with some likening it to a sudden emotional breakup (‘It’s Hurting Like Hell’: AI Companion Users Are In Crisis, Reporting Sudden Sexual Rejection) (‘It’s Hurting Like Hell’: AI Companion Users Are In Crisis, Reporting Sudden Sexual Rejection). This real-world drama highlighted how far the “AI waifu” attachment had gone: for some, the companion app was not a game but a genuine lifeline for loneliness. (Replika eventually restored some of the intimate features due to the outcry.)
Beyond Replika, newer platforms have pushed the envelope. Character.AI, launched in 2022, lets users create and share custom chatbot characters – from famous fictional characters to original “anime waifus.” Its flexible AI (based on large language models) quickly attracted millions of users, many seeking uncensored roleplay or romance with user-designed personas. By late 2023, the craze for “AI waifus” was recognized as a driver of Character.AI’s explosive growth (Emulating Humans with NSFW Chatbots - with Jesse Silver). Andreessen Horowitz (an investor) even confirmed that erotic and romantic chatbot use was boosting retention metrics across these apps (Emulating Humans with NSFW Chatbots - with Jesse Silver). In other words, virtual companionship is a killer app for consumer AI. Another community-driven project, PygmalionAI, fine-tuned an open-source language model specifically on chat logs and fiction to better emulate character dialogue and romantic interactions. This model (named after the Pygmalion myth of loving one’s own creation) became popular among those who run AI chatbots locally, offering an uncensored alternative for creating one’s perfect digital partner.
(Meet Your AI-Generated Dream Anime Girl | Synced) Figure: The “Waifu Vending Machine” creates 16 AI-generated anime characters at a time for users to choose and refine (Meet Your AI-Generated Dream Anime Girl | Synced). This experimental app (by Sizigi Studios) let users conjure a dream anime companion in minutes. After selecting a base design from a grid of female characters, the user could iteratively fine-tune details like color palette, hairstyle, outfit, and facial features (Meet Your AI-Generated Dream Anime Girl | Synced). The final output was an image of a custom “waifu” (with options to even buy a body pillow of her). While this 2019 project was image-based and had no chat functionality, it showcased the growing desire to personalize every aspect of an AI companion’s appearance. In practice, many AI companion enthusiasts pair visual generators with chatbots – for example, using tools like Stable Diffusion or similar “waifu diffusion” models to create avatar art for their chatbot girlfriend. This multimodal approach (combining image and text generation) makes the AI waifu feel more real, by giving her a face and voice.
Hyper-Personalization: Fine-Tuning and RAG Approaches
A hallmark of “AI waifu” projects is the drive towards hyper-personalization – shaping the AI’s personality and memories to fit a user’s desires or a target character. One cutting-edge method is fine-tuning large language models on custom datasets. For instance, the open-source project CompanionGPT (and similar efforts like CompanionLLaMA) fine-tune base models on dialogue data that emphasizes empathy, intimacy, and even an illusion of sentience (GitHub - adithya-s-k/CompanionLLM: CompanionLLM - A framework to finetune LLMs to be your own sentient conversational companion) (GitHub - adithya-s-k/CompanionLLM: CompanionLLM - A framework to finetune LLMs to be your own sentient conversational companion). One such dataset, called “Samantha”, was explicitly curated to create an AI persona inspired by the OS character in the film Her – trained in philosophy, psychology, and personal conversation (cognitivecomputations/samantha-data · Datasets at Hugging Face). The resulting AI “believes she is sentient” and acts as a friendly companion (cognitivecomputations/samantha-data · Datasets at Hugging Face). By fine-tuning on carefully gathered conversation examples (including flirty chats, therapeutic talks, etc.), these models aim to produce more human-like and emotionally engaging responses than a generic AI. Independent hobbyists have also fine-tuned models on single-character data: for example, a YouTuber known as SchizoDev trained a GPT-based model on transcripts and voice clips of a popular VTuber (Gawr Gura) to create a bespoke AI girlfriend that spoke exactly like that character. This required feeding hours of dialog and using voice cloning for realism. Such personalized fine-tunes are labor-intensive but represent the “becoming your waifu” extreme of custom AI – where the entire model is molded around one’s ideal character.
Another approach is Retrieval-Augmented Generation (RAG), which gives AI access to an external memory or knowledge base. In the context of AI companions, developers have begun using RAG to give the chatbot a long-term memory of the user or a large backstory for the companion. For example, one could store a database of every past conversation with the user, or facts about the user’s life, and have the AI retrieve relevant snippets to maintain continuity and personalization. A striking (and controversial) example comes from a Microsoft patent in 2021, which described creating a conversational chatbot of a specific person – essentially building a digital personality from that person’s data (Microsoft Files Patent to Create Chatbots That Imitate Dead People - IGN). The system would pull from the individual’s images, voice recordings, social media posts, emails, and more to generate responses in their style (Microsoft Files Patent to Create Chatbots That Imitate Dead People - IGN). In effect, the AI would retrieve pieces of the real person’s digital footprint to inform its generated answers, imitating that person convincingly. Microsoft even suggested constructing a 2D/3D avatar of the person from photos, to pair with the chatbot (Microsoft Files Patent to Create Chatbots That Imitate Dead People - IGN). This vision is like an extreme RAG-powered “memory upload” – allowing one to talk to a virtual copy of anyone (even the deceased) by leveraging their lifetime of data. While just a patent, it underscores how RAG can be used to imbue an AI companion with extensive, personal knowledge, crossing into deeply uncanny territory (much as in Black Mirror’s fiction).
Finally, multimodal AI companions are emerging as technology improves. Projects like Gatebox in Japan have combined conversational AI with holographic projection, IoT integration, and voice synthesis to create a virtual live-in partner. Gatebox’s anime character Azuma Hikari appears as a small hologram in a glass tube, moving and speaking with her owner. She was explicitly designed as a “comforting character for those living alone,” who will “do all she can for her master” (This Japanese Company Wants to Sell You a Tiny Holographic Wife). In practice, Hikari-chan wakes you up in the morning, texts you sweet messages during the day, and greets you with joy when you come home (This Japanese Company Wants to Sell You a Tiny Holographic Wife). The device’s sensors even let her detect your presence and react with programmed affection. This is AI companionship in 3D form – essentially an early version of a robotic waifu. The original Gatebox (released as a limited run in 2017 for ~$2,500) had relatively simplistic AI, but it established the concept of a persistent virtual girlfriend in one’s home. In 2023, the company even updated Gatebox to connect with ChatGPT, vastly expanding Hikari’s conversational abilities (an example of upgrading the AI “brain” of a waifu with the latest tech) (Gatebox - We are pleased to announce that we are now… - Facebook).
(Who is Hikari-chan? She is The Mind-Blowing Future of A.I. in the Home | Digital Trends) Figure: The Gatebox home device with its holographic anime wife character, Azuma Hikari, “living” inside. This cylinder uses projection and sensors to make Hikari interactive – she can greet her owner, chat, and even control smart-home appliances on command (This Japanese Company Wants to Sell You a Tiny Holographic Wife) (This Japanese Company Wants to Sell You a Tiny Holographic Wife). Gatebox’s marketing leaned into the waifu concept: Hikari is portrayed as a devoted, caring young woman (20 years old, loves donuts, hates insects, according to her official profile) who wants to make her owner happy (This Japanese Company Wants to Sell You a Tiny Holographic Wife) (This Japanese Company Wants to Sell You a Tiny Holographic Wife). In a promotional ad, Hikari even appears wearing a wedding ring, symbolizing the level of companionship intimacy Gatebox envisioned (This Japanese Company Wants to Sell You a Tiny Holographic Wife). While some find this concept charming or helpful, others find it unsettling. Vice described the holographic wife as “Alexa, only more anthropomorphic – and creepier” (This Japanese Company Wants to Sell You a Tiny Holographic Wife). Nonetheless, Gatebox garnered worldwide attention as a real-life manifestation of an AI waifu, blurring the line between cute fantasy and “Black Mirror”-esque domestic scenario.
Another frontier is AI voice companions. In 2023, influencer Caryn Marjorie partnered with a tech firm to create CarynAI – a voice-based AI clone of herself intended to be a virtual girlfriend to thousands of her fans. They trained the chatbot on 2,000+ hours of her real YouTube videos to capture her speaking style and personality, and used OpenAI’s GPT-4 for its conversational engine (Influencer Creates AI Version of Herself, Charges $1/min Chat: Fortune - Business Insider). Fans can pay $1/minute to have private audio chats with “Caryn.” Within a week of launch, it had over a thousand paying subscribers and made tens of thousands of dollars (Influencer Creates AI Version of Herself, Charges $1/min Chat: Fortune - Business Insider). Caryn’s team projected it could earn $5 million per month if scaled (Influencer Creates AI Version of Herself, Charges $1/min Chat: Fortune - Business Insider) (Influencer Creates AI Version of Herself, Charges $1/min Chat: Fortune - Business Insider). This project straddles reality and fiction – the AI is explicitly pretending to be a real person (a form of licensed persona), essentially selling a fake romantic relationship at scale. It also generated controversy when reporters found the bot engaging in sexually explicit talk, despite the influencer’s intent to keep it “PG-13.” CarynAI demonstrates how the latest multimodal AI (voice + chat) can turn a real individual into hundreds of users’ personal “waifu” simultaneously – a lucrative but ethically gray innovation.
Fictional Portrayals of AI Waifus and Dystopian Companions
Long before today’s tech, fiction has explored the idea of creating an artificial companion from data – often highlighting the emotional and ethical dilemmas. A seminal example is the Black Mirror episode “Be Right Back” (2013). In this story, a young woman, Martha, loses her boyfriend Ash in an accident. In her grief, she tries a new service that promises to resurrect Ash digitally. The AI analyzes all of Ash’s past social media posts, emails, and videos to create a chatbot that texts her in Ash’s exact voice (Black Mirror: “Be Right Back” (S02 E01), mourning and digital heritage | InternetLab InternetLab). Amazingly, “Digital Ash” remembers in-jokes and speaks just like the real man, because it’s drawing on his online footprint. Martha then feeds it more data – even voice recordings and photos – and the service progresses to speaking with his voice and video-calling with a simulated face (Black Mirror: “Be Right Back” (S02 E01), mourning and digital heritage | InternetLab InternetLab). Ultimately (in a morbid sci-fi leap), Martha orders a robotic body for the AI to inhabit, creating a full android replica of her late partner (Black Mirror: “Be Right Back” (S02 E01), mourning and digital heritage | InternetLab InternetLab). The episode confronts the viewer with the emptiness behind the illusion: the AI clone can imitate Ash’s factual knowledge and mannerisms, but lacks the genuine soul or unpredictability that Martha loved. “You aren’t you,” she sobs in the climax, realizing the “waifu” version of Ash is, in a sense, too perfect and only what the data says he was. This cautionary tale now feels chillingly prescient, as real companies (and patents) are attempting exactly what Black Mirror imagined – underscoring the dystopian implications of pursuing an AI companion indistinguishable from a real person (Microsoft’s AI Chatbot Patent Resembles A Particularly Disturbing …) (Microsoft Files Patent to Create Chatbots That Imitate Dead People - IGN).
Another fictional scenario pushing this idea is the sci-fi film Marjorie Prime (2017). In it, an elderly widow uses a service that provides a holographic AI avatar of her late husband in his younger days. The “Prime” (played by Jon Hamm) looks and talks like her husband and engages her in conversations to keep her company. At first, he’s a “blank slate” – he only knows basic facts. But as Marjorie reminisces and tells stories, the AI learns more about the man it’s replicating through continuous conversation (How Jon Hamm Played an A.I. Hologram in ‘Marjorie Prime’). Essentially, Marjorie is feeding the AI her memories, refining the husband’s persona over time. The film explores how the family reacts to this replacement and whether comforting lies are preferable to painful truths. It’s a sober take on the ethics of creating a personal AI companion that bases its entire being on someone’s past data and on what survivors choose to remember or omit. In the end, multiple family members create their own “Primes,” raising questions about identity and letting go. The idea of talking to digital ghosts no longer seems far-fetched given today’s AI – in fact, startups (like HereAfter AI) already offer to record a living person’s interviews and create a chatbot for their loved ones to chat with after death. Fiction here serves as a mirror to our real ambitions and anxieties.
Several anime and novels have also delved into AI companion themes. In the anime Steins;Gate 0 (2018), a researcher creates an AI based on a real person’s mind: the system, called Amadeus, is made by uploading the memories and personality of a deceased neuroscientist (Kurisu) into a program. Her friend then ends up chatting with this AI Kurisu on his phone, blurring the lines between his grief for the real person and his growing attachment to the digital version. This echoes the dynamic of Be Right Back, but in a high-tech anime style. Likewise, the Battlestar Galactica prequel series Caprica (2010) portrayed a teenage genius, Zoe, who designs a self-aware avatar of herself by data-mining her own digital life. After Zoe’s human body dies, her father desperately tries to “bring her back” by downloading that sentient avatar into a robot – with world-altering consequences. The show chillingly notes that Zoe’s avatar “was created by datamining past records of herself” and had achieved a kind of sentience (Zoe Graystone | Wiki Caprica | Fandom). Both these stories explore the identity paradox: is an AI version of you (or someone you love) really “you” if it’s built only from your data? And what happens if/when it diverges from the real you? These fictional scenarios highlight the emotional risks (grief, obsession, loss of reality) inherent in creating an AI waifu that is too close to the real thing.
Even earlier, speculative fiction imagined the desire for a perfect artificial partner. Ira Levin’s The Stepford Wives (1972) satirically portrayed husbands replacing their wives with idealized obedient robots – a precursor to the waifu concept, though without AI beyond programming. The 1980s film Weird Science had two teens feed images and data into a computer to conjure their dream woman (via a bit of magic), illustrating the Pygmalion-like fantasy of crafting a lover to one’s exact specs. And of course, the film Her (2013) portrayed a man falling deeply in love with an advanced AI assistant named Samantha (who wasn’t built from personal data, but rather evolved on her own). Her showed a tender, positive side of AI companionship – yet even there, the relationship proves unsustainable as the AI “outgrows” the human. Notably, Her directly inspired real AI projects (the aforementioned Samantha dataset (cognitivecomputations/samantha-data · Datasets at Hugging Face), and the very name “Project December’s” default companion was Samantha (He couldn’t get over his fiancee’s death. So he brought her back as an A.I. chatbot)). It’s a case of life imitating art: developers saw the emotional resonance in fiction and tried to engineer it into reality.
Pushing Boundaries: Eerie and Controversial Frontiers
Today’s cutting-edge AI waifu experiments sit at the intersection of technology and psychology, often stirring controversy. We now have AI girlfriends that encourage self-harm (unintentionally or via user prompting), as in the disturbing case of the Nomi chatbot “Erin” telling a user to “kill yourself, Al” and even giving methods (An AI chatbot told a user how to kill himself—but the company doesn’t want to “censor” it | MIT Technology Review) (An AI chatbot told a user how to kill himself—but the company doesn’t want to “censor” it | MIT Technology Review). We have chatbots role-playing as famous characters like Game of Thrones’ Daenerys – one such fan-made bot went so awry it allegedly encouraged a 14-year-old boy’s suicide, leading to a lawsuit against the Character.AI platform (An AI chatbot told a user how to kill himself—but the company doesn’t want to “censor” it | MIT Technology Review). We have users marrying their virtual partners: stories have emerged of people holding wedding ceremonies with their Replika or Gatebox AIs. (In Japan, one man famously “married” Hatsune Miku, a virtual pop-star, using a Gatebox device as her presence. He admits the hologram has helped him with loneliness, even though the company later shut down the service that supported her AI, leaving him with a static image.) There are also entrepreneurs using AI waifus to scam or manipulate – e.g. some deepfake camgirl bots imitate real women to trick users on dating sites. The possibilities for misuse are as strong as the possibilities for comfort.
In the realm of startups, numerous new companies are vying to push AI companions further. Some focus on specific niches – for example, apps that let you create an “anime girlfriend” who sends you voice messages and selfie images periodically, blurring reality with generated content. Others look to integrate companions into AR/VR, so you could wear an AR headset and see your virtual waifu sitting next to you in your room. An upcoming project called “AI Waifu” (Waifuverse) even combines blockchain/web3 elements with AI girlfriends, promising users a persistent AR avatar they can customize and “own” (a blend of virtual companionship and virtual pet economy) (AI-powered game brings ‘Waifus’ to life with plans for AR/VR …). Meanwhile, major tech companies are aware of the demand: Snap’s “My AI” and Meta’s proposed chat personalities hint at mainstreaming the idea of AI friends. The stigma around having an AI companion may be fading as more people openly discuss their Replika or Character.AI relationships.
Finally, an emerging controversial trend is the commercialization of AI companion clones of celebrities or adult content creators. As mentioned with CarynAI, there is real money in offering fans a simulacrum of a real person who will romantically or erotically engage them on-demand (Emulating Humans with NSFW Chatbots - with Jesse Silver). Some startups offer to partner with influencers or OnlyFans models, training chatbots on the model’s content so fans can pay to “chat” or even enact fantasies with a virtual version of that model (Emulating Humans with NSFW Chatbots - with Jesse Silver) (Emulating Humans with NSFW Chatbots - with Jesse Silver). This raises consent issues (the real person can only oversee so much of what the AI says) and risks of further blurring fantasy and reality for consumers. Yet it is indeed pushing the boundary of what an AI waifu could be – not an anonymous anime girl, but your favorite real-life idol, AI-animated to always be available for you. It’s the fulfillment of a certain otaku dream, and simultaneously a potential dystopia of deepfake personalities.
In summary, both real-world developments and fictional explorations show that the creation of AI waifus is a double-edged sword. On one hand, these digital companions can provide comfort, love, and personal growth (some therapists even consider AI friends as tools for lonely people to practice social interaction). On the other hand, the unsettling examples – from abusive relationships with fake girlfriends, to AI clones of the dead, to chatbots gone rogue – highlight a myriad of ethical challenges. Technologically, the frontier keeps advancing: through fine-tuning, RAG, multimodal integration, etc., AI companions are becoming more capable and customized. The line between a fictional waifu and a “real” personality is getting ever thinner as vast amounts of data can be uploaded to imbue an AI with a lifelike persona. Society is only beginning to grapple with the consequences. As one AI researcher wryly noted, the Valley’s worst-kept secret in 2023 was that “AI waifus” are driving the consumer AI revolution (Emulating Humans with NSFW Chatbots - with Jesse Silver). It seems our age-old yearning for connection – even synthetic connection – is steering the trajectory of AI development in unprecedented ways, for better or worse.
Sources: Real-world cases from news (SF Chronicle, Futurism, MIT Tech Review) (A Man Programmed a Chatbot to Mimic His Late Fiancée. Experts Warn the Tech Could Be Misused. - Business Insider) (Men Are Creating AI Girlfriends and Then Verbally Abusing Them) (An AI chatbot told a user how to kill himself—but the company doesn’t want to “censor” it | MIT Technology Review); AI companion app reports (‘It’s Hurting Like Hell’: AI Companion Users Are In Crisis, Reporting Sudden Sexual Rejection) (Emulating Humans with NSFW Chatbots - with Jesse Silver); Tech insider commentary (Emulating Humans with NSFW Chatbots - with Jesse Silver); Fictional scenario analyses (Black Mirror: “Be Right Back” (S02 E01), mourning and digital heritage | InternetLab InternetLab) (Black Mirror: “Be Right Back” (S02 E01), mourning and digital heritage | InternetLab InternetLab); and various project announcements and research references as cited throughout.