Zum Inhalt springen

ELIZA: The Chatbot That Horrified Its Creator

Zusammenfassung

ELIZA (1966), the first chatbot, was designed by MIT computer scientist Joseph Weizenbaum specifically to demonstrate how shallow and superficial human-computer “conversation” was. He built it as a proof that humans would anthropomorphize any sufficiently responsive program, regardless of whether it understood anything. He was horrified when people took ELIZA seriously — when his own secretary asked him to leave the room so she could speak with it privately, when psychiatrists proposed using it for therapy, and when academics suggested it demonstrated machine understanding. His book Computer Power and Human Reason (1976) was a direct response to the consequences of his own invention.

The Design

ELIZA ran on an IBM 7094 at MIT between 1964 and 1966, programmed by Weizenbaum in MAD-SLIP. The most famous script, DOCTOR, simulated a Rogerian psychotherapist — a style of therapy where the therapist reflects the patient’s statements back as questions, encouraging elaboration without providing analysis.

The implementation was simple: ELIZA matched input text against patterns and transformed matched patterns into responses. If the input contained “I am [X]”, ELIZA might respond “How long have you been [X]?” If the input contained “my mother”, ELIZA might ask “Tell me more about your family.” If no pattern matched, it used generic responses: “Please go on” or “That’s very interesting.”

ELIZA understood nothing. It performed pattern matching and string substitution. It had no model of the conversation’s content, no understanding of the patient’s situation, and no persistence between conversations (it could not remember what had been said earlier in the same session without specific programming).

The Reactions

Weizenbaum expected the simulation to be transparently hollow — a demonstration tool. The reactions surprised and disturbed him:

His secretary’s privacy request: The secretary who typed in the lab asked Weizenbaum to leave the room while she conversed with ELIZA. She knew it was a program. She still wanted privacy for her interaction.

Psychiatric proposals: Several psychiatrists seriously proposed that ELIZA could provide large-scale, affordable psychotherapy. Weizenbaum found this appalling: the program had no understanding; it could produce responses that appeared empathetic while missing the actual content of what the patient said.

Academic interpretations: Some academics argued that ELIZA demonstrated that machine communication was more human than previously thought. Weizenbaum argued the opposite: it demonstrated that human communication required less understanding than previously assumed.

Computer Power and Human Reason

Weizenbaum’s 1976 book argued that computers should not be used for applications that required human judgment, genuine compassion, or ethical reasoning — not because computers couldn’t simulate these behaviors, but because simulation without understanding was dangerous. A computer that appeared compassionate while lacking compassion was worse than a machine that appeared to be merely a machine.

The book was controversial. Marvin Minsky criticized it; AI researchers argued Weizenbaum was overstating the risks. But the concern Weizenbaum raised — that humans would attribute understanding to programs that exhibited none — has proved accurate in ways that extend far beyond ELIZA. The same dynamic plays out with modern language models, whose outputs are often interpreted as evidence of understanding that the models’ architecture does not support.

The pattern Weizenbaum identified — anthropomorphization of responsive machines — is also discussed in Expert Systems and the First AI Winter and the AI history articles.

📚 Sources