In the winter of 1966, a program running on an MIT mainframe convinced people it understood them. It did not. That gap—between what a machine appears to grasp and what it actually computes—was opened sixty years ago by a piece of software called ELIZA, and we have never quite closed it. If anything, we have spent the intervening decades making it wider, glossier, and far more convincing.
A Skeptic at the Console
Joseph Weizenbaum was a German-born computer scientist who joined the MIT faculty in the early 1960s, working in the orbit of the institute’s pioneering artificial-intelligence research. Between 1964 and 1966 he wrote ELIZA, named—with a literary wink—after Eliza Doolittle, the flower girl in George Bernard Shaw’s Pygmalion who is drilled to imitate genteel speech. The reference was deliberate. Like Doolittle, ELIZA could perform a convincing surface without possessing the substance underneath.
Weizenbaum described the program in a now-famous paper in Communications of the ACM. He was not trying to build a mind. He was trying to study natural-language conversation between human and machine, and to show how much could be achieved with surprisingly little—a point that, characteristically, he intended as a caution rather than a boast.
How ELIZA Actually Worked
ELIZA had no model of meaning, no memory of the world, no understanding of what a sentence was about. It was a pattern-matching and substitution engine. The program scanned a user’s typed input for keywords, ranked them, and applied transformation rules—called decomposition and reassembly rules—to rearrange the user’s own words into a response.
The personality everyone remembers came from a script layered on top, called DOCTOR, which imitated a Rogerian psychotherapist. This was a shrewd choice. A therapist in the Rogerian tradition reflects a patient’s statements back at them rather than supplying outside knowledge, which conveniently excused ELIZA from needing any. The machine never had to know anything; it only had to turn your words around and hand them back as a question. The core moves were simple:
- Spot a keyword such as mother, I, everyone, or computer, and trigger an associated template.
- Transform pronouns and verbs so that “my” became “your,” “I am” became “you are.”
- Reflect the rephrased fragment back as a probing question.
- When no keyword matched, fall back on content-free prompts like “Tell me more” or “Why do you say that?”
The resulting dialogue felt uncannily attentive:
- User: Men are all alike.
- ELIZA: In what way?
- User: They’re always bugging us about something or other.
- ELIZA: Can you think of a specific example?
- User: Well, my boyfriend made me come here.
- ELIZA: Your boyfriend made you come here.
No comprehension is happening. The machine is a mirror with good manners.
The ELIZA Effect
What Weizenbaum did not anticipate was how readily people would pour meaning into the mirror. Users confided in the program. They attributed insight, empathy, even care to a few hundred lines of string manipulation. The phenomenon became known as the ELIZA effect: the human tendency to assume that fluent linguistic behavior implies an understanding mind behind it.
The most telling anecdote is one Weizenbaum himself recounted. His own secretary, who had watched him build the thing and knew exactly what it was, asked him to leave the room so she could talk to ELIZA in private. She knew it was a program. She wanted privacy anyway. That detail unsettled Weizenbaum more than any technical limitation, because it revealed that knowing the truth about a machine does not necessarily change how we feel about it.
Weizenbaum Turns Against His Own Creation
He grew genuinely alarmed. What disturbed him most was not ELIZA itself but the reaction of serious people—including practicing psychiatrists who mused that programs like it might one day deliver automated therapy at scale. To Weizenbaum, this was a category error with moral weight: mistaking the imitation of language for the presence of judgment, and being willing to hand human matters to a system that merely simulated concern.
In 1976 he published Computer Power and Human Reason, one of the first ethical critiques of artificial intelligence written from inside the field. His argument was not that computers were weak but that some decisions should never be delegated to them regardless of competence—that there is a difference between what a machine can be made to do and what it ought to do. The book made him a dissident among his peers, and it remains startlingly current.
The Through-Line to Today
It is tempting to treat ELIZA as a quaint ancestor, all the way down on the evolutionary chart from today’s large language models. The differences are real and enormous. ELIZA worked on a handful of hand-written rules and reflected single sentences; a modern LLM is trained on a substantial fraction of human text, holds long context, reasons across domains, writes code, and produces output that is genuinely useful rather than merely evocative. ELIZA could not have answered a question; modern systems answer billions.
And yet the eerie part is what hasn’t changed. A language model still has no inner experience to accompany its words. It still predicts plausible text rather than knowing things in the way a person knows them. And the human on the other end still does what Weizenbaum’s secretary did—infers a mind, feels heard, forms attachments. The ELIZA effect did not go away; it scaled. We now build products that depend on it. The machinery underneath is incomparably more powerful, but the oldest variable in the system is still us, and we are still inclined to see a self where there is only a very good mirror.
That is the inheritance worth carrying forward. At Artificial Heights we build conversational systems that are far more capable than anything Weizenbaum could have run in 1966—but his caution is part of the spec, not a footnote to it. The point was never to make a machine that fools people into feeling understood. It is to build systems good enough that we never have to.