29 October 2026 Best practices

Why people talked to ELIZA, an artificial intelligence program

Discover how ELIZA used rules to create convincing conversations without understanding text, and what its limits can teach you about chatbots.

By Mehdi Chouikh, CEO at UTIO

People talked to ELIZA because its responses could seem relevant and personable, even though the program did not understand the meaning of what they wrote. It applied rules to sustain the appearance of a conversation, and this way of responding could lead some people to attribute understanding to it.

In brief:

ELIZA applied rules and patterns written by its creator; it did not learn from what each user told it.
Open-ended questions and repeated words could invite people to fill in the meaning of prepared responses.
A rule-based chatbot can respond to situations covered by its rules, but it cannot handle what falls outside them on its own.
A convincing response does not prove that a system understands intent or context. To evaluate it, look at what it does with the information, not just how it sounds.

The difference between having a conversation and carrying out a process helps you assess which tasks an AI agent can take on in a small business. Explore UTIO’s verticals.

How was ELIZA different from a machine that learns?

ELIZA applied rules written by its creator to recognize certain structures and transform sentences; it did not learn from what each user told it. By contrast, a learning system identifies patterns in examples during training and adjusts its behavior based on them, although that does not mean it learns from every conversation.

The difference lies in how the response is produced. With ELIZA, behavior depended on predefined instructions: if the text matched a pattern, the program applied the corresponding rule. If it did not match, the program had limited options.

A system trained on examples can generalize patterns that have not been written as individual rules. Even so, learning patterns is not the same as understanding like a person. To assess an agent for a small business, check whether it follows the intended process and when it should pass the case to a person. Anyone looking to integrate AI capabilities into a product can consult the UTIO AI API.

How did ELIZA generate responses without understanding what it was saying?

ELIZA generated responses by identifying recognizable words or structures in the text and applying rules to transform them. It could reuse parts of a sentence or return a prewritten question, but it did not check what the person meant or whether the response fit the full context.

The mechanism depended on what its rules covered. If a sentence matched a recognized pattern, the program could turn it into a different structure and respond. This allowed it to keep the exchange going as long as it encountered patterns it knew how to handle; it did not allow it to infer a new intent or resolve a contradiction on its own.

The practical difference becomes clear when the conversation goes off script. A response can sound appropriate because it repeats or rephrases words, even though the program has not checked their meaning. When evaluating a WhatsApp agent, ask what task it performs with each message and what it does when it lacks a rule or enough information. Fluency alone does not demonstrate understanding.

Why did some people feel that ELIZA understood them?

ELIZA could seem understanding because it responded to certain statements with open-ended questions or echoed the user’s words. These techniques gave the conversation a sense of continuity. That impression describes a possible reaction to its response style; it does not show that everyone experienced it or that the program understood what they told it.

A person may read an open-ended question as interest and supply context that the response did not interpret. Repeating a word can also make an exchange seem connected.

That explains one possible impression, but it does not let us claim that all users felt understood or turn an anecdote into a general reaction. Without a specific historical source, it is more precise to describe the mechanism than to attribute a shared experience to a group.

When evaluating a current agent, separate tone from task: check what information it collects and what action it takes. The UTIO customer stories explain, company by company, what each one needed, what was set up, and what task its agent does today.

What is the difference between having a conversation and understanding one?

Having a conversation means producing responses that fit what someone has just said. Understanding it means interpreting what they want, what information they have provided and what would follow from responding or taking action. A fluent sentence alone does not show that a program has understood the intent or context.

To assess an agent, look at what it does when a request is ambiguous or missing information. Does it ask for clarification? Does it say what information it needs? Does it stop if it cannot answer confidently? These signs help you evaluate its behaviour rather than judging only the tone of its responses.

In customer service, answering a question is not enough. The response must take the relevant process and information into account. The support vertical lets you explore a specific use case. When assessing any agent, check what information it consults, what actions it can complete and when it should hand over to a person.

Frequently asked questions

Who created ELIZA?

Joseph Weizenbaum, a professor at MIT, created ELIZA. He had joined the institution in 1963. He developed it between 1964 and 1966, and MIT’s digital archive preserves a complete listing of its source code, written in the MAD-SLIP language, with Weizenbaum listed as the author.

What was ELIZA’s DOCTOR script?

DOCTOR was an ELIZA script. In his 1966 paper, Weizenbaum explains that a script is a set of keywords and transformation rules for a type of conversation, and that scripts at the time responded more or less like a Rogerian psychotherapist. The code listing held by MIT places the DOCTOR script at the end.

Was ELIZA the first chatbot?

ELIZA is often considered the first chatbot, though that claim is not undisputed: MIT’s obituary of Weizenbaum describes it as “perhaps the first example” of what is now called a chatbot. What is documented is that he created it at MIT in the 1960s, using written rules and no learning.

Does the ELIZA effect mean that a machine understands people?

No. The “ELIZA effect” takes its name from Weizenbaum’s program and describes people’s tendency to attribute understanding and intelligence to a computer system based on how it converses. It is a reaction from the person using it, not a capability of the machine: a fluent-sounding response does not show that the program understood anything.

Explore specific tasks an AI agent can perform for a small business.