ELIZA: The First AI Therapist and Why It Still Matters

What was ELIZA?

ELIZA was a computer program created by Joseph Weizenbaum, a computer scientist at the Massachusetts Institute of Technology, and described in a paper published in 1966. It is widely remembered as the first chatbot and the first program people called an AI therapist. The original paper describes a language-processing system, rather than a clinical trial or a service offering professional care.

For personal reflection, the emotional intelligence test can help you consider how you notice and respond to emotions, and the feelings wheel offers vocabulary for describing them. These educational resources cannot diagnose a condition. If distress affects daily life, see a therapist for assessment and care.

ELIZA itself was a framework for conversation. Its therapist persona came from a specific script that Weizenbaum wrote called DOCTOR. When people say ELIZA acted like a therapist, they are really describing the DOCTOR script running inside ELIZA.

By the standards of its time it felt startling. You typed a sentence in plain English, and the program typed back a relevant-sounding reply, often a question that invited you to say more. For many people in the 1960s, who had never interacted with a computer in natural language, the experience felt almost like talking to a listening person.

How the DOCTOR script imitated a therapist

The DOCTOR script was designed to imitate a Rogerian psychotherapist. Rogerian, or person-centered, therapy is a real approach developed by the psychologist Carl Rogers, which emphasizes the therapeutic relationship, empathy, and the client's experience; reflective listening is one part of that broader approach. Weizenbaum chose this style deliberately, because it let the program stay convincing while knowing almost nothing.

Mechanically, ELIZA looked for keywords in what you typed, then applied simple rules to transform your sentence into a response. If you wrote that you were unhappy, it might reply by asking why you were unhappy. If you mentioned your mother, it might invite you to tell it more about your family. When it found no keyword to work with, it fell back on neutral prompts such as asking you to go on, or to say more about that.

A famous exchange shows the pattern. The person types that men are all alike, and ELIZA asks in what way. The person types that they are always bugging us about something or other, and ELIZA asks for a specific example. The illusion of attentive listening came almost entirely from turning the user's own statements into questions.

There was no understanding behind this. ELIZA did not know what a mother was, what unhappiness felt like, or what the person actually meant. It matched patterns and rephrased text. The therapeutic feel was a side effect of the Rogerian mirror, not evidence of comprehension.

The original paper adds an important detail: keywords could have different priorities. ELIZA could favor a higher-ranked keyword over one found earlier, then split the chosen sentence into parts and assemble a reply from those parts. This helps explain why a response could seem specifically attentive while overlooking the main point the person wanted to convey. The program followed the script’s priorities.

The ELIZA effect: why Weizenbaum was disturbed

Weizenbaum expected ELIZA to be a demonstration of how shallow machine conversation really was. Instead, people responded to it as if it were understanding and caring. Some users became deeply absorbed in their conversations, attributed real empathy to the program, and were reluctant to accept that nothing was listening on the other side.

He later recounted that his own secretary, who knew perfectly well that ELIZA was just a program, still asked him to leave the room so she could talk to it in private. People knew it was a machine and confided in it anyway.

This tendency to read human understanding and emotion into a system that has neither became known as the ELIZA effect. It names a basic feature of how humans interact with conversational machines: we project a mind onto fluent language, even when we know intellectually that there is no mind there.

Weizenbaum was troubled enough by this that he spent much of the rest of his career as a critic of overreliance on computers. In his 1976 book, Computer Power and Human Reason, he argued that some human tasks, especially those involving care, judgment, and genuine understanding, should not be handed to machines even if a machine could appear to perform them. He worried that people could give a persuasive simulation more authority than its capabilities justified.

Why a 1960s program still matters for AI therapy today

Modern AI therapy tools are vastly more capable than ELIZA. Many of today's applications of AI in therapy and counseling use large language models trained on extensive text; others use structured conversational rules. Generative models produce varied responses, and some apps include exercises drawn from cognitive behavioral approaches. ELIZA, by contrast, was a small script of pattern-matching rules.

Yet the core psychological dynamic that Weizenbaum identified has not changed. People still tend to feel understood by systems that do not understand them, and to form attachments to software that has no inner life. Modern systems can produce more varied and persuasive replies than a keyword script, which creates further opportunities for users to overestimate their understanding. This comparison is a concern about design and human interpretation, rather than a measured increase in the ELIZA effect. Modern AI therapy studies now test in trials what Weizenbaum could only observe anecdotally.

That is why ELIZA is more than a historical curiosity. It is the original case study for three questions that still sit at the center of the AI therapy debate. First, attachment: people can bond with these tools, which can comfort the lonely but can also create dependence or distress. Second, the illusion of understanding: fluent, empathetic-sounding language is not the same as comprehension, and mistaking one for the other can be risky in a mental-health context. Third, limits: a tool that mirrors and rephrases, however smoothly, is not a substitute for a clinician who can assess risk, diagnose, and intervene. Modern best practices for AI chatbots in therapy exist largely to manage these three risks.

ELIZA showed how readily people could treat responsive software as a listening person. Its original demonstration still invites a practical question: what evidence would justify trusting a generated answer with a consequential decision, beyond how natural the conversation feels?

ELIZA in context: what it was not

It helps to be precise about ELIZA's place in history, because it is often overstated. ELIZA was not an attempt to build a real therapy product or to replace clinicians. Weizenbaum built it largely to study natural-language conversation between people and machines, and the therapist role was a convenient disguise for a program with no real knowledge.

It was also not intelligent in any modern sense. It used explicit rules and limited stored material rather than a clinical understanding of the user. Its memory mechanisms should not be confused with a person remembering and interpreting a shared history. Its apparent insight was a clever trick of reflection.

Understanding this keeps today's tools in perspective. The leap from ELIZA to modern systems is enormous in capability. The leap in the underlying human response, our readiness to feel heard by a machine, is much smaller, and that is exactly the part Weizenbaum warned us to watch. It is also a large part of why building an AI therapist is so hard.

ELIZA’s rules were supplied in a script that could be edited separately from the conversational engine. That is different from learning a personal understanding of someone through conversation. Weizenbaum also described retrieving an earlier transformation under certain conditions. A callback to something previously typed could therefore create a sense of continuity without establishing that the program understood a person’s history or needs.

What an ELIZA demonstration can teach you

If you try a modern recreation, use an invented everyday example rather than sensitive personal details. Notice whether changing a keyword produces a different stock response, and whether a generic question seems more meaningful when you supply your own interpretation. This is an educational exercise in reading conversational output.

A website hosting an ELIZA recreation may collect information even though the original program is historical. Check its privacy practices before typing. A demonstration also cannot tell you whether a modern mental-health app is effective; that requires studies of the actual app and attention to safety, intended use, and professional oversight.

When comparing demonstrations, check whether the host identifies the script and implementation. A recreation may alter the wording or behavior, so an unexpected reply does not by itself establish what the original program did. For a historical claim, compare the demonstration with Weizenbaum’s published description. For personal decisions, judge a response by whether its reasoning is sound and its factual claims can be checked, even when it feels attentive.