A 1966 program proved fluent replies are not understanding

Start here

Anyone typing a question into a chatbot today inherits a problem that is sixty years old: fluent, responsive text is not evidence that anything understood the question. The first program built on this trick was ELIZA, written by Joseph Weizenbaum at MIT and published in Communications of the ACM in January 1966. Its famous script, DOCTOR, made ELIZA answer like a Rogerian psychotherapist, as the project overview puts it: 'offering little of its own, mostly reflecting your words back as questions.' ELIZA never read for meaning; it matched a keyword, split the sentence there, and slotted the pieces into a reply template. That description is the starting point for judging any assistant that answers you today: a smooth reply describes a process, not a guarantee of comprehension.

What the documents say

A 2025 paper, ELIZA Reanimated, describes how researchers recovered an original ELIZA printout from Weizenbaum's archive and reran the code, confirming how the 1966 program worked. The DOCTOR script documentation built from that code shows the mechanism: each keyword carries a rank, a decomposition pattern splitting a sentence into numbered parts, and a reassembly template rebuilding a reply from them. The canonical exchange Weizenbaum printed shows it working: a user types 'Men are all alike,' the keyword ALIKE fires, and ELIZA answers 'IN WHAT WAY' -- a template, not a thought. The overview adds the human half: 'Weizenbaum was disturbed that people knew this and confided in it anyway,' the reaction that named the gap between what a program does and what people believe it does, the ELIZA effect.

Check this

The check is mechanical and works on any assistant: read a short exchange and ask what the reply adds that you did not supply yourself. ELIZA's answers recombine your own words: told 'my mother takes care of me,' it stores a transformed copy and hands it back later as 'earlier you said your...' -- filing, not recollection. A larger model can make the pattern harder to spot, so name the mechanism first: whether the reply draws on a fact it looked up, or simply reshapes words already on the screen.

What holds and what fails

This is an editorial reading, not a claim in the sources: the ELIZA lesson holds whenever fluency comes from pattern and recombination, and fails once a system genuinely retrieves or reasons over information you did not provide, which some tools now do. The distinction is not old software versus new; it is checkable retrieval versus confident rearrangement. DOCTOR's framing worked because a therapist can plausibly ask questions without asserting facts. Many assistants now assert facts, which is where source-checking has to take over.

• Before trusting a surprising answer, ask what specific new fact it contains beyond what you typed.
• Read one recovered exchange at the DOCTOR script page to see the pattern-matching run.
• When a reply reflects your own phrasing back at you, treat that as a cue to ask a plainer, checkable question.

ELIZA was never meant to convince anyone; how thoroughly people confided in it anyway is the oldest evidence that fluent language earns trust before understanding does.