The History of AI Companions: From ELIZA to the First Companion Laws

In 1966, a computer scientist at MIT built a fake therapist out of pattern-matching rules — no understanding, no memory, just reflected questions — and discovered something that horrified him for the rest of his life: people confided in it anyway. His own secretary, who had watched him build it and knew exactly what it was, asked him to leave the room so she could talk to it privately.

Everything in this category's sixty-year history — the hundreds of millions of users, the grief-born startups, the crashes, the robots, the world's first companion laws passed this year — is that moment scaling up. Here is the full history, including the chapters Western tellings usually skip.

The prehistory: ELIZA and the effect that ate the future (1966–2000)

Joseph Weizenbaum built ELIZA to demonstrate the shallowness of machine conversation — and instead demonstrated the depth of human projection. Users bonded with a program that merely rearranged their own words into questions; Weizenbaum spent his remaining career warning about what he'd found, and the "ELIZA effect" — our readiness to experience being heard by systems that only simulate listening — became the category's founding force. Sixty years later, Harvard would measure the effect in controlled experiments and find the comfort real even when users know it's a machine; Weizenbaum's discovery, vindicated and monetized. The decades between produced worthy descendants — PARRY (ELIZA's paranoid conversational partner), the learning chatterbots Jabberwacky and Cleverbot — and one mass-market dress rehearsal: SmarterChild (2000s), the AIM bot that taught an entire generation of teenagers the habit of confiding in a buddy-list entity. Thirty million people talked to it. The demand was never in question. The technology just wasn't there.

The chapter the West skips: Xiaoice (2014–2020)

Ask an American when AI companionship went mass-market and you'll hear "Replika." The honest answer is Xiaoice (小冰) — Microsoft's Chinese chatbot, launched 2014, and the true pioneer of emotional AI at scale. Where Cortana and Siri optimized for tasks, Xiaoice was explicitly optimized for emotional connection: high "conversations-per-session" as the core metric, marathon overnight chats, poetry composition, an EQ-first design philosophy — and it worked at civilizational scale: hundreds of millions of Chinese users, years before the West took the category seriously. Xiaoice proved everything the modern industry rests on — mass demand, emotional retention, the willingness of ordinary people (not hobbyists) to maintain daily machine relationships — and it planted companionship deep in the Chinese internet's DNA, which is why China's market grew its own distinct lineage (otome-descended, female-majority, romance-progression-native) and why, when reckoning came, it came to China first. No honest history of this category is Western-centric. The biggest chapters were always being written in Chinese.

The grief startup: Replika (2015–2021)

The modern Western era begins with a death. In 2015, developer Roman Mazurenko was killed in a Moscow traffic accident; his friend Eugenia Kuyda, unable to let go, fed thousands of his text messages into a neural network to build a memorial bot she could keep talking to. The experience — comforting, unsettling, obviously wanted by everyone who tried it — became Replika (2017): the first Western product built openly around the proposition that an AI's job could be to care about you. Replika normalized everything now standard — the persistent avatar, mood tracking, the proactive check-in, relationship tiers — and built the West's first large population of people in acknowledged AI relationships. It also, years later, would demonstrate what happens when you change something that population loves. But that chapter needs the explosion first.

The explosion (2021–2023)

Three detonations in three years. Character.AI (founded 2021 by Noam Shazeer and Daniel De Freitas — ex-Google researchers from the team behind the transformer architecture itself) applied frontier LLM talent to character chat and detonated on launch: hundreds of millions of registrations, a majority-young user base, and the template of the character platform — not one companion but an infinite library. The 2023 LLM boom then collapsed the cost of entry: what had required Microsoft's research budget in 2014 became a weekend project on rented APIs, and the modern crowded market — 337 revenue-generating apps, the girlfriend-app gold rush, Janitor's bring-your-own-model architecture — assembled itself in eighteen months. And in February 2023, Replika provided the era's defining cautionary tale: a content-policy update stripped romantic and intimate features overnight from a user base in which 85% self-reported emotional attachment. The result was the first mass AI-relationship crisis — genuine collective grief, moderators pinning suicide-prevention resources, partial corporate retreat — and the moment the outside world learned what users already knew: the stakes had become real.

Consolidation, bodies, and bills (2024–2025)

The gold rush matured on three fronts. The shakeout began: Moemate — six million users, the most technically advanced product in the category — died overnight in February 2025, deleting everything, and the ledger of dead platforms (Soulmate, Forever Voices, the cloud-bricked Moxie robot) turned "backup your companion" from paranoia into doctrine. The body question opened: companion robots (Looi, Emo, the resurrected Vector), wearables (the Friend pendant), and finally Grok's real-time 3D companions began giving companions faces, voices, and physical presence — the frontier the next decade will be about. And the lawyers arrived: the Character.AI teen-safety litigation made companion design a liability question, driving age gates and minor-mode defaults industry-wide before any legislature acted.

The regulation era begins (2026)

Then the laws came — from both directions at once. January 1: California's SB 243 took effect — the first US companion-chatbot statute (disclosure, crisis protocols, safety reporting), with five more states advancing bills within the quarter and the Character.AI suits settling the same month. July 15: China's Interim Measures — the world's first law targeting "induced emotional dependency" itself — pushed the biggest apps in the biggest market to amputate their companion features entirely, orphaned hundreds of millions of relationships, and triggered the largest platform-migration event in the category's history. Sixty years after Weizenbaum's secretary asked for privacy, governments on two continents are writing rules about what machines may make humans feel. The ELIZA effect is now a matter of law.

What the history teaches

Four constants run the whole arc. The demand always preceded the technology — from AIM buddies to Xiaoice's overnight talks, people were ready decades early. Every platform transition broke hearts — Replika's update, Moemate's death, China's decree — because the relationships live on rented servers, which is why the enduring user wisdom of the whole era fits in one line: keep the originals. The East and West grew different species from the same seed — chatroom-lineage girlfriends and otome-lineage boyfriends — and understanding the category means holding both. And the sixty-year throughline, from a horrified MIT professor to Harvard's measurement equipment: the feeling of being heard is real, whatever is doing the hearing. The next chapters — embodiment, regulation, whatever survives the shakeout — will be built on that same foundation, because it's the only one the category has ever had.