ELIZA—a computer program for the study of natural language communication between man and machine

Joseph Weizenbaum’s 1966 paper describes ELIZA, a program that produces conversational replies by transforming a user’s text according to an editable script. Its contribution is both a concrete text-processing design and an investigation of how little machinery can sustain an impression of understanding.^[1]

The paper appeared in Communications of the ACM 9(1), pages 36–45, January 1966. The chronology uses January 1, the publisher-deposited catalog date; this does not establish a day of physical circulation or the program’s creation.^[1]

How the conversation works

The program scans an input for keywords, gives higher-ranked keywords priority, and tries associated decomposition rules: patterns that divide the sentence into components. A matching reassembly rule then constructs a reply from stock words and selected components, with substitutions such as changing the speaker’s pronouns. The paper’s pattern (0 YOU 0 ME) uses zero to match any number of words; its corresponding reply can reuse the intervening verb phrase while reversing who is addressing whom. The operation depends on a pattern match rather than understanding that phrase.^[1]

Scripts contain the keywords, patterns and replies as data, separate from the underlying program. Multiple reassembly rules help avoid immediate repetition. If no suitable keyword is found, ELIZA can issue a general prompt or retrieve a previously stored transformation. This limited memory is important: the paper does not describe retaining and reasoning over the whole conversation. An interactive editor lets a script writer add or change rules as experience exposes gaps.^[1]

Why the psychiatric interview mattered

The main scripts imitate a Rogerian psychotherapist’s conversational style. Weizenbaum explains that this setting permits a questioner to appear interested without supplying much independent knowledge. In the paper’s boat-ride example, a request to discuss boats can be interpreted as a deliberate therapeutic question rather than evidence that the program knows nothing about boats. The human speaker contributes that interpretation.^[1]

The paper prints an illustrative dialogue and reports that some subjects were difficult to convince that the program was not human. It supplies no participant count, controlled comparison or clinical outcome measure for that observation. Instead, Weizenbaum proposes experiments varying scripts and initial instructions to investigate credibility. The therapeutic style is a device for studying conversation, not evidence that ELIZA treats illness.^[1]

Understanding and credibility

Weizenbaum distinguishes continuing a conversation from drawing valid conclusions from what has been said. He notes that ELIZA discards most input and proposes that a more instructable successor should expose misunderstandings so its partner can correct them. Its existing script cannot overcome every limitation of the processor merely by adding more rules.^[1]

This distinction gives the ELIZA Effect a concrete mechanism: a reader can supply coherence and intent beyond what the program establishes. The paper’s lasting question is how people judge the credibility of computer output, especially when that judgment informs consequential decisions.^[1]