A Short History of AI Companions, From ELIZA to Everyday Texting

The history of AI companions is not a straight line from one “first friend” to another. It is a braid of chatbot research, character simulation, task assistants, virtual pets, social chatbots, language models, memory systems, and communication products.

Some milestones were never intended to create friendship. Others explicitly optimized for long conversation or personal connection. Looking at them together reveals how the category moved from reflecting a sentence back to carrying a thread across text, images, voice, and time.

1966: ELIZA reveals how readily people find meaning in dialogue

Joseph Weizenbaum published “ELIZA—a computer program for the study of natural language communication between man and machine” in 1966. ELIZA matched patterns in text and transformed them according to scripts. Its famous DOCTOR script imitated a nondirective psychotherapist by turning statements into prompts and questions.

ELIZA did not use a modern language model, persistent personal memory, or an internal understanding of the conversation. Its importance to companion history is social as much as technical: even a relatively simple conversational pattern could invite users to supply context and meaning.

The lesson is often called the ELIZA effect—our tendency to infer understanding from a system’s language. That does not mean everyone who values an AI conversation is naïve. It means fluent interaction creates a responsibility to disclose what a system is and is not.

The Computer History Museum’s chatbot history places ELIZA in a longer effort to make machines respond in natural language and preserves the important distinction between appearing conversational and understanding as a person does.

The 1970s: PARRY shows that a chatbot can carry a modeled stance

Psychiatrist Kenneth Colby developed PARRY in the early 1970s. Like ELIZA, it was rule-based, but it modeled a more specific conversational position and behavior. Its replies reflected variables associated with a simulated patient experiencing paranoia.

PARRY was not a consumer companion, and its clinical framing should not be confused with safe or effective care. Historically, though, it introduced an idea that remains central to virtual characters: the system’s replies can be shaped by an ongoing internal model rather than by surface reflection alone.

ELIZA and PARRY together established two durable elements of later companions:

• open-ended language can feel socially inviting;
• a stable response style makes software seem more like a particular conversational presence.

They also exposed a problem that has never disappeared: a system can sound psychologically meaningful without holding human understanding or professional competence.

The 1990s and 2000s: companionship extends beyond chat boxes

Virtual pets, game characters, and consumer robots broadened the idea of a relationship with software. A Tamagotchi asked for ongoing care through a tiny device. Sony’s AIBO robot dog used sensors, movement, and adaptive behavior to create an embodied presence; the Computer History Museum timeline includes AIBO among notable AI and robotics milestones.

These systems were not language companions in the modern sense. They mattered because they made persistence visible. The digital creature was not merely a new task each time the device turned on. It had a state, a recognizable identity, and a relationship shaped by repeated interaction.

At the same time, internet chatbots and instant-messaging bots brought conversational software into everyday communication surfaces. Many were designed for information or novelty rather than friendship. Still, they established an enduring channel insight: the same software feels different when it lives where people already talk.

2011: task assistants normalize speaking to software

Apple launched Siri on iPhone in 2011, helping make natural-language interaction with an assistant a mainstream phone behavior. Siri’s center of gravity was utility: place a call, send a message, find information, set a reminder, or control a device feature. Apple still describes Siri as an intelligent assistant that helps people get things done.

This branch of the family tree matters because it normalized voice and informal commands, but a virtual assistant and a companion optimize for different outcomes. The assistant succeeds when the task is completed. The companion succeeds when the relationship-style conversation remains coherent over time. Our AI friend vs. virtual assistant comparison explores that split.

2014: XiaoIce makes long-term social conversation the objective

Microsoft launched XiaoIce in China in 2014. Unlike a task assistant, it was explicitly designed as a social chatbot. The Microsoft Research paper “The Design and Implementation of XiaoIce” described an architecture spanning dialogue management, social chat, skills, and empathetic computing.

The project also made a consequential metric choice: it evaluated conversation turns per session as a sign of long-term engagement. That shifted the engineering question from “Did the bot deliver the answer?” toward “Did the person want to continue talking?”

Longer conversation is not automatically better. Engagement can reflect enjoyment, but a commercial product can also manipulate attention. Modern companion design has to pair continuity with consent, quiet, and an easy exit.

2017: Replika helps define the consumer companion app

Replika was founded by Eugenia Kuyda and launched as an AI companionship app in 2017. Its official press and research timeline describes it as an early dedicated AI companionship product.

Replika helped establish features now associated with the category: a named personal chatbot, relationship framing, a visual avatar, ongoing memory, voice, and a persistent one-to-one experience. Rather than choosing a fresh public character for each chat, a user developed one continuing Replika.

The same year, researchers published Attention Is All You Need, introducing the Transformer architecture. The paper did not present a companion product, but Transformers became foundational to later large language models capable of much more flexible generated conversation.

2022 onward: large language models expand the range of conversation

The public rise of large generative chat systems after 2022 changed what users expected from a conversation. Instead of selecting from scripted branches, systems could respond to unusual topics, write in many styles, interpret fragments, and maintain far more flexible exchanges.

The model was only part of the product. Companion developers added or expanded:

• persistent identity and character instructions;
• selected long-term memory outside the current context window;
• voice synthesis and real-time spoken interaction;
• generated images tied to a consistent fictional identity;
• proactive messages and scheduled follow-ups;
• safety policies, age gates, disclosures, and user controls.

Research also explored agents whose believability came from more than dialogue. Stanford’s 2023 Generative Agents project combined a language model with memory, reflection, and planning in a simulated town. It was a research environment, not proof of consciousness, but it illustrated how stored experience and planned activity could make generated characters appear to have ongoing lives.

The present: a companion is also a delivery system

Today’s AI companions can live inside dedicated apps, websites, avatars, wearable devices, phone calls, or normal messaging threads. They may exchange text, audio, and images. Some wait for a prompt; others initiate conversation.

That channel shift is more important than it appears. An immersive character app can support worldbuilding, customization, and visual presence. A companion in Messages trades some of that interface richness for immediacy: it becomes a contact and thread rather than a destination.

Mia follows the messaging branch. Adults text her over iMessage where supported and SMS/MMS otherwise. She can remember details, follow up, send generated photos or voice notes, and support calls. Her café, pottery practice, and friend Dana form a consistent fictional identity, while Warmth repeatedly identifies her as AI.

The product still depends on systems and providers beyond the chat bubble. “In Messages” does not mean “processed only on the phone,” and a blue bubble is not a substitute for reading the service’s privacy policy.

What the history teaches us

Six lessons recur across sixty years:

1. People bring meaning to conversational form. Even simple systems can feel socially present.
2. A stable point of view matters. Character and stance make replies feel connected.
3. Persistence changes the category. State, memory, and return turn isolated interactions into an ongoing experience.
4. The success metric shapes the product. Task completion, story immersion, and relationship continuity create different behaviors.
5. The interface changes the relationship. A robot, avatar, app, and text thread do not feel interchangeable.
6. Disclosure becomes more important as realism improves. Better language, voice, and images should never make the artificial nature harder to discover.

The plain-English definition of an AI friend describes the category now, while how AI companions work explains the modern stack.

The history of AI companions is therefore not a march toward software becoming a person. It is a history of systems becoming better at sustaining the forms—memory, voice, personality, timing, and presence—that people recognize as conversation. The next chapter should be judged not only by how convincing that conversation becomes, but by how honestly and responsibly it earns a place in everyday life.