A 1966 script showed listening could be simulated by rules

The design

Joseph Weizenbaum, a computer scientist at MIT, described ELIZA in a paper published in the January 1966 issue of Communications of the ACM. The paper explains the program scans a sentence for a keyword, applies a matching decomposition rule, then a reassembly rule to generate a reply; no representation of meaning is built anywhere. Rules for a conversational style form a script, data rather than code, so the program could converse in English, Welsh or German by loading a different file. The best-known script, later called the DOCTOR script, made ELIZA answer as a Rogerian psychotherapist would, reflecting a statement back as a question, chosen because a psychiatric interviewer can appear to know nothing about the world without breaking the illusion. The ELIZAGEN project, documenting the program's history, records that Weizenbaum wrote the original in 1965 in MAD-SLIP on an MIT IBM 7094, naming it 'ELIZA (aka. Doctor).'

What the evidence says

Weizenbaum's paper is candid the program's apparent understanding is an illusion produced by pattern matching alone. Its closing section reports that people typing as they would to a psychiatrist found it credible a person was answering, hard to shake once told otherwise; he called this a 'striking form' of the Turing test. The ELIZAGEN history page adds a point from decades later: a 2014 account of psychology faculty visiting a DOCTOR-script replica and judging it useful because it showed exactly how not to conduct Rogerian therapy. Together the sources show the same effect recurring across nearly fifty years: an informed observer names the trick, while an ordinary partner still experiences being heard.

What it asks of people

Nothing in ELIZA required a user to disclose anything: sentences were processed within a session, with no cross-session storage. The design instead asks for an act of interpretation: the psychiatric frame lets ELIZA answer almost any statement without appearing ignorant, so a user supplies the sense the responses seem to carry. Weizenbaum reports this cooperative reading is easy to trigger and hard to interrupt once a person expects a sympathetic interlocutor, an asymmetry later commercial companion products also depend on.

Privacy and safeguards

ELIZA in 1966 was a laboratory demonstration, not a service, and the paper makes no privacy claim because none was engineered: no retention policy or consent flow, since no persistent account or company stood behind it. That absence is informative for evaluating a companion product built on a similar trick: understanding's illusion needs no stored data at all, so a product's data practices and its capacity to feel understanding-like remain separate questions.

• Does a product's apparent understanding depend on more than pattern matching and a permissive conversational frame?
• What happens to the sense of being heard once a user is told, plainly, no understanding is taking place?
• Who benefits from preserving the ambiguity between a scripted response and an understood one?

ELIZA's paper remains useful because it isolates the effect from any data collection or business incentive to prolong an interaction: being understood needs only a script and a permissive frame, worth remembering before crediting a modern companion product's engagement numbers to genuine comprehension.