ELIZA

In 1966, Joseph Weizenbaum published an article in the “Communications of the ACM”. In it, he described ELIZA, a program capable of conversing in natural language with a human being. At MIT, where the researcher worked, this achievement created a sensation in artificial intelligence circles.

The name chosen for this program was not random. Eliza Doolittle, the heroine of George Bernard Shaw’s Pygmalion, embodies the transformation of a simple flower seller into a distinguished lady through the lessons of Professor Higgins. Joseph Weizenbaum saw in this metaphor a machine that learns to “speak” under a teacher’s supervision while retaining its limitations—a perfect illustration of his program.

On MIT’s MAC time-sharing system, ELIZA ran using the MAD-SLIP language on an IBM 7094. Users connected from remote terminals and began what resembled genuine conversation. They typed their sentences, the program responded. The illusion worked. Yet the mechanisms animating ELIZA were disconcertingly simple. The program scrutinized each sentence searching for keywords. Once a term was identified, it applied predefined transformation rules. If no keyword was detected, it drew from a repertoire of generic responses or reused a previous transformation.

This approach revealed its full subtlety in the hierarchization of keywords. When a sentence contained “I” and “everyone”, the program favored the latter term, because excessive generalizations often deserve attention. The transformation rules broke sentences down according to preset patterns to reconstitute them into coherent responses.

ELIZA’s modular architecture constituted its true innovation. The conversational rules lived in scripts separate from the program’s core. This design allowed behavioral modifications without touching the source code. Scripts thus emerged in German and Welsh, proof of the system’s flexibility.

The script that would make ELIZA famous simulated a Rogerian psychotherapist. Carl Rogers had developed a person-centered therapy where the therapist reformulates the patient’s statements without claiming to understand everything. This approach perfectly matched the limited capabilities of Joseph Weizenbaum’s program. Faced with “I am sad”, ELIZA responded “I am sorry to hear that you are sad” or “Can you explain what makes you sad?” The program maintained the illusion through reformulation, open-ended questions, and requests for clarification.

Joseph Weizenbaum did not anticipate the impact of his creation. Users attributed genuine understanding to ELIZA. Some requested private sessions with the program. This reaction troubled the researcher, who realized that humans projected their own interpretations onto the machine’s mechanical responses. They themselves constructed the meaning of the exchange. This spontaneous anthropomorphization still resonates in our relationships with voice assistants and other chatbots. ELIZA revealed how elementary mechanisms generate the illusion of intelligence, raising lasting questions about the nature of understanding.

The program’s technical weaknesses are glaringly obvious: no memory of past exchanges, no model of the interlocutor, no knowledge of the real world. Responses emerged from syntactic transformations blind to meaning. But these limitations do not erase ELIZA’s historical importance, which established the foundations of automatic language processing: keyword recognition, identification of minimal contexts, selection of appropriate transformations, generation of default responses.

Beyond the technical aspects, ELIZA crystallized our reflection on artificial intelligence. Joseph Weizenbaum observed this paradox where, once the mechanisms had been explained, the “magic” evaporated and the program revealed itself as a collection of trivial procedures. Yet the effect on users persisted. ELIZA teaches us that the perception of intelligence in our exchanges with machines depends as much on our projections as on the actual capabilities of the systems.

The questions raised by Joseph Weizenbaum’s program about the nature of understanding, the boundaries of simulation, and the ethical stakes of artificial intelligence have lost none of their relevance. ELIZA remains that foundational moment when computing began to explore the ambiguous territories between mechanical language processing and genuine intelligence.