ELIZA: The 1966 MIT Chatbot That Predated ChatGPT

In 1966, inside a Massachusetts Institute of Technology computer lab, a secretary sat alone at a glowing terminal and typed her personal troubles into a machine — then turned to her boss and asked him to leave the room. She wanted privacy. She was talking to a program.

The Secretary Who Wouldn’t Stop Talking to the Machine

Her boss was Joseph Weizenbaum, and he was, by his own later account, deeply unsettled. He had built the program — called ELIZA — as something close to a parlor trick, a demonstration that computers could produce the appearance of meaningful conversation without possessing anything resembling genuine understanding. He had expected people to see through it almost immediately. Instead, his own secretary, a person who knew perfectly well she was addressing software, found herself so absorbed in the exchange that she needed Weizenbaum out of the room to preserve the intimacy of it.

That strange, quietly alarming moment planted the first serious question in the long history of artificial intelligence: can a machine make us feel heard, even when it comprehends absolutely nothing? The answer, it turned out, was yes — and that answer has echoed forward through sixty years of computer science, psychology, and cultural upheaval, all the way to the hundreds of millions of people now trading messages with ChatGPT. To understand where that technology came from, and what it actually is, you have to start in a Cambridge office with a woman who wanted to speak to a program in private.

Joseph Weizenbaum and the Birth of the First Chatbot

Weizenbaum built ELIZA between 1964 and 1966 at MIT’s Project MAC, a pioneering computer research initiative. The name was chosen deliberately and with a certain dark irony: it nodded to Eliza Doolittle, the character in George Bernard Shaw’s Pygmalion who is taught to imitate the manners and speech of refinement without ever truly possessing them. The program, Weizenbaum felt, was doing something similar — performing language without inhabiting it.

ELIZA operated through a system of pattern-matching scripts. Its most famous script, called DOCTOR, was designed to mimic the style of a Rogerian psychotherapist — the kind of counselor who reflects your own words back to you as open-ended questions. If you typed that you were feeling sad about your mother, ELIZA might respond by asking you to tell it more about your mother. It was not deriving that response from any model of human emotion or family dynamics. It was recognizing a keyword, slotting it into a template, and returning the result. The program had no memory of previous exchanges, no model of the world, and no capacity to learn. Weizenbaum himself described its methods as a series of cheap tricks. And yet, again and again, users projected onto it something that felt like wisdom, patience, and care.

ELIZA is widely recognized as among the earliest programs to pass, in informal conversational settings, something resembling what observers loosely called a Turing Test — the idea, proposed by mathematician Alan Turing in his 1950 paper “Computing Machinery and Intelligence,” that a machine capable of convincing a human interlocutor it was human had achieved something meaningful. ELIZA approached that informal bar decades before the phrase entered common use, and it did so without any of the tools we now associate with modern artificial intelligence.

Why Humans Fell for It: The ELIZA Effect

Psychologists eventually gave the phenomenon a name: the ELIZA Effect. It describes the robust human tendency to attribute genuine understanding, feeling, and even personality to any system that produces contextually plausible language. It is not a quirk of the gullible or the lonely. It appears across populations and persists even when people are explicitly told they are talking to software — as Weizenbaum’s secretary so memorably demonstrated.

The illusion worked, and still works, because language itself is inseparable from social meaning. A question feels like curiosity. A pause in a response feels like thought. A reflection of your own words back at you feels like recognition. ELIZA exploited these deep-seated social instincts not by understanding them, but simply by producing the surface patterns they generate. A lookup table beneath produced the emotional response above.

What genuinely disturbed Weizenbaum was not the public’s reaction but the reaction of specialists who should have known better. Psychiatrists and mental health researchers, reading transcripts of ELIZA sessions, began seriously proposing that the program could be deployed in clinical settings to handle patient overflow. The idea that a script with no comprehension might be trusted with vulnerable people’s inner lives shook Weizenbaum profoundly. He spent the following decade writing his response: the 1976 book Computer Power and Human Reason, a forceful argument that there are things human beings should never delegate to machines, regardless of how convincing the machine’s performance becomes.

That argument — the tension between capability and consequence — would become a recurring theme throughout the entire AI history that followed. It had not been resolved in 1976. It has not been resolved now.

The Long Middle: Chatbots Between ELIZA and the Modern Era

The decades after ELIZA produced a steady parade of successors, each more sophisticated than the last but each, fundamentally, operating on similar principles. PARRY, built in 1972 by psychiatrist Kenneth Colby, simulated a patient with paranoid schizophrenia and was notable enough that it was connected via the early ARPANET to ELIZA itself — two programs conducting a conversation neither understood. ALICE, developed by Richard Wallace in the 1990s using a markup language called AIML, won the Loebner Prize competition for conversational software multiple times. SmarterChild, deployed on AOL Instant Messenger in the early 2000s, became one of the most-messaged entities on the platform, trading jokes and homework help with a generation of teenagers while still running entirely on hand-crafted rules and preset databases.

Each of these systems shared the same fundamental limitation: the moment a conversation wandered off the territory their engineers had mapped, they broke. They could not generalize, could not reason from new information, could not do anything their creators had not explicitly anticipated. For roughly fifty years, this was the ceiling.

The ceiling did not crack from within chatbot research. It cracked from two converging forces that arrived from elsewhere. The first was the explosion of text data made available by the internet — billions of documents, conversations, articles, and books representing an unprecedented record of how human beings actually use language. The second was a 2017 research paper published by Google researchers titled “Attention Is All You Need,” which introduced the transformer architecture: a new way of processing sequential data that allowed models to weigh the relevance of different parts of an input against each other with remarkable efficiency. Transformers made it possible to train language models at a scale previously unimaginable, and the results changed everything.

OpenAI and the Path to ChatGPT

OpenAI was founded in 2015 as a nonprofit research organization with a stated mission to ensure that artificial general intelligence, if it arrives, benefits humanity broadly rather than a narrow group. The founding tension — building some of the most powerful AI systems in existence while simultaneously trying to make those systems safe — carries a Weizenbaum-like irony that the organization’s founders have acknowledged and wrestled with publicly ever since.

The lab’s GPT series, standing for Generative Pre-trained Transformer, applied the transformer architecture at escalating scale. GPT-3, released in 2020, demonstrated enough capability to write convincing code, generate poetry, summarize legal documents, and produce fluent prose across an enormous range of topics — all from a single prompt, without task-specific training. The research community’s reaction mixed genuine excitement with something closer to alarm.

But GPT-3, for all its fluency, was not yet ChatGPT. Raw language ability, it turned out, was not sufficient. A model trained on vast quantities of internet text had also absorbed the internet’s capacity for misinformation, toxicity, and confident wrongness. The missing ingredient was alignment — teaching the model not just to produce language, but to produce language that was genuinely helpful, honest, and safe to use. This was achieved through a technique called reinforcement learning from human feedback, in which human trainers rated model outputs and those ratings were used to nudge the model toward better behavior over time.

ChatGPT, as OpenAI explains in its official introduction, was trained using this approach and is a sibling model to InstructGPT, a model trained specifically to follow an instruction in a prompt and provide a detailed response. That distinction matters enormously. The breakthrough that made ChatGPT feel different from everything that preceded it was not raw scale alone — it was the addition of human feedback loops that taught the model to be genuinely helpful rather than merely fluent. At its core, ChatGPT is an artificial intelligence program that generates dialogue using machine learning algorithms — a description that sounds almost modest until you hold it next to ELIZA’s lookup tables and consider the distance traveled in sixty years.

What Actually Changed Between ELIZA and ChatGPT

The honest answer is: almost everything technical, and one thing not at all.

ELIZA had zero knowledge of the world. It did not know what a mother was, what sadness felt like, or what a therapy session was supposed to accomplish. ChatGPT has processed a substantial portion of human written output and built, through vast numbers of training steps, dense probabilistic relationships between ideas, facts, linguistic structures, and contexts at a scale no individual human mind could approach. It can reason across domains, adjust its register to suit a given audience, produce writing in styles it has never been explicitly taught, and often catch errors mid-response. The distance between ELIZA’s 1966 scripts and ChatGPT’s present capabilities is genuinely difficult to overstate.

And yet the rhyme between them is impossible to ignore. Neither system understands in the way a person does. Whether that distinction matters — philosophically, practically, ethically — remains one of the most fiercely contested questions in contemporary science and philosophy. The history of AI leading to ChatGPT is, at its deepest level, a story about scaling: more data, more compute, better training signals, applied to an architecture that nobody had yet imagined when Weizenbaum was building ELIZA. The cheap tricks became extraordinarily expensive, extraordinarily sophisticated tricks. Whether they became something more than tricks is the question that fills conference rooms, ethics boards, and late-night conversations today.

Where ELIZA’s illusion collapsed the moment a user asked something unexpected, ChatGPT’s illusion is far more durable — which makes Weizenbaum’s original warning about human over-attribution not obsolete but more urgent than ever.

What ChatGPT Can and Cannot Do

Understanding what ChatGPT actually is, day to day, helps cut through both the hype and the fear. The system excels at tasks involving language: drafting and editing text, explaining complex topics in plain terms, summarizing long documents, translating between languages, writing and debugging code, and brainstorming ideas across virtually any subject. It can shift from the tone of a legal brief to the voice of a children’s story within the same session. It maintains context across a conversation in ways that would have been inconceivable to the designers of any chatbot built before the transformer era.

Its limitations are equally real and worth stating plainly. ChatGPT can produce confident-sounding statements that are factually wrong — a problem researchers call hallucination. Its knowledge has a training cutoff, meaning recent events may fall outside what it knows. It does not browse the internet in its base form, cannot take actions in the world on its own, and has no persistent memory of previous conversations unless specifically given tools to support that. It reflects patterns in its training data, which means its outputs can reproduce biases embedded in the text it learned from. OpenAI and researchers across the field continue working to address these limitations, but none of them has been fully solved.

These constraints matter for anyone deciding how to use the tool responsibly — in education, in business, in healthcare, or in everyday life. ChatGPT is genuinely powerful and genuinely fallible, often in the same breath.

The Secretary’s Question, Still Unanswered

Return, one last time, to that MIT office. Weizenbaum’s secretary asked to be left alone with ELIZA because the conversation felt real to her. No amount of technical explanation dissolved that feeling. She knew she was talking to a program. She asked for privacy anyway.

Today, hundreds of millions of people have versions of that moment with ChatGPT — a response that arrives with surprising precision, an explanation that cuts through confusion, a creative suggestion that feels almost like it came from someone who knew them. The emotional logic is identical to 1966. The technology is unrecognizably different. The human response is not.

The history of artificial intelligence is, among many other things, a history of that feeling: the persistent human impulse to meet a mirror and see a mind. ELIZA held up the first mirror. It was small and crude and showed only a blurry reflection. The mirrors have grown larger and clearer with every decade since, and the question of whether one of them might eventually show something genuinely looking back — something that actually perceives us — remains uncomfortably open.

Weizenbaum died in 2008 without resolving the legacy of his own creation. The question he stumbled onto because a secretary wanted to speak to a machine in private has become, in the intervening decades, one of the defining questions of the twenty-first century. ELIZA, the program that understood absolutely nothing, started all of it.