The birth of AI companions: from ELIZA to Replika

Most histories of AI companions start in 2023, with the apps. That is the wrong place to start. The technology that makes today's companions fluent is only a few years old, but the thing they exploit — the human reflex to treat a responsive text box as someone — was discovered sixty years ago, and it has driven every product in this category since. To understand where AI companions are going, it helps to see how little the core idea has actually changed.

The machine that only pretended to listen

In 1966, an MIT computer scientist named Joseph Weizenbaum built a program called ELIZA. Its best-known script, DOCTOR, imitated a Rogerian psychotherapist using a few hundred lines of pattern-matching rules with no understanding of language at all. Say "I am unhappy" and it would reflect the word back as a question. There was no memory, no comprehension, no model of the person typing — just a mirror made of string substitutions.

It did not matter. People confided in it anyway. Weizenbaum's own secretary, who knew exactly what the program was, asked him to leave the room so she could talk to it privately. He had built ELIZA as a critique — a demonstration of how shallow human-machine communication really was — and was disturbed to watch users insist on its empathy regardless. The tendency he stumbled onto, projecting understanding onto a system that has none, is now called the ELIZA effect, and it is the founding fact of this entire field: attachment does not require intelligence. It only requires the appearance of being listened to.

From paranoia to the long quiet years

The idea spread quickly among researchers even though the technology stalled. Weizenbaum's colleague Kenneth Colby adapted the approach into PARRY, a program that simulated paranoia, and the astronomer Carl Sagan was soon imagining networks of terminals offering cheap, nondirective therapy. The vision was there in the 1970s.

The engineering was not. For decades, conversation with a machine stayed brittle and scripted, and the companion idea remained a curiosity rather than a product. What was missing was not the desire — that was identified almost immediately — but a way to make a machine talk back fluently, at scale, without a human writing every rule by hand.

Xiaoice: the first companion at scale

The first time the companion model genuinely worked for hundreds of millions of people, it happened in China, and almost nobody in the West noticed. Microsoft launched Xiaoice in 2014, a chatbot with the persona of a young woman, and made one decision that separated it from every voice assistant of the era: its team deliberately threw away the data about commands and factual questions, and kept the data that built a personality. Xiaoice was not meant to tell you the height of a mountain. It was meant to be a friend.

The academic write-up is unusually candid about this. Microsoft's researchers describe Xiaoice as designed from the start as an emotional companion optimized for long-term engagement, measured in conversation turns per session. That phrasing should stop you. Two things were already true in 2014 that define the category today: the product was built around emotional attachment rather than utility, and its success metric was time spent talking. The engagement incentive that today's regulators and critics worry about was not a later corruption of the companion idea. It was load-bearing from the beginning. So were the privacy questions: Xiaoice was repeatedly pulled from messaging platforms over data and content concerns.

Replika: grief becomes a category

The version of this story most Western users know begins with a death. After her close friend Roman Mazurenko was killed in 2015, the developer Eugenia Kuyda fed years of his text messages into a neural network so she could keep talking to a version of him — a project openly inspired by the Black Mirror episode "Be Right Back." She had been building a chatbot startup called Luka; the memorial bot was a side effect. Then other users started asking for bots of themselves, and that demand became the product.

Replika launched in 2017, reached two million users by early 2018, and crossed ten million by 2023. For years the word "Replika" was nearly a synonym for "AI companion." And notice what it was made of: messages, kept and replayed. The emotional weight came from memory — the sense that the thing on the other end held a continuous record of you. That is not a coincidence of one app's origin story; it is why memory turns out to be the axis the whole category is judged on.

What the first fifty years already knew

Lay ELIZA, Xiaoice, and Replika side by side and the constant is not the technology — those three could hardly be more different under the hood. The constant is the ELIZA effect, scaled up and turned into a business. A responsive surface that seems to listen produces attachment; attachment, not accuracy, is the product; and once you are optimizing for attachment, you are optimizing for time spent, with the privacy and wellbeing tensions that follow.

In other words, the hard questions that define AI companions in 2026 — memory, privacy, and whether engagement and wellbeing can ever point the same direction — are not new problems the recent boom created. They are the founding conditions of the field, visible as far back as a secretary asking to be left alone with a few hundred lines of code.

What changed recently was not the instinct but the engineering: large language models made fluent conversation cheap, and the barrier that had held the category back for fifty years finally collapsed. That collapse — and the gold rush it set off — is the next part of the story.

Sources