Chatbot
Definition
A chatbot is a program that conducts a conversation with a person in natural language, through text or speech.
History
ELIZA (1966), written by Joseph Weizenbaum at MIT, is the origin point. Its best-known script, DOCTOR, imitated a Rogerian psychotherapist by reflecting the user's statements back as questions, using pattern matching on keywords and almost no model of meaning. Weizenbaum was disturbed rather than pleased by the result: people who knew exactly how the program worked, including his own secretary, formed attachments to it and asked to be left alone with it. He named the phenomenon and it kept his name — the ELIZA effect, the human tendency to attribute understanding to a system that produces plausible responses.[1] His 1976 book is an argument, from the person who built the thing, that there are decisions computers should not be asked to make.
PARRY (1972), by the psychiatrist Kenneth Colby, simulated a person with paranoid schizophrenia and was more convincing than ELIZA — psychiatrists shown transcripts could not reliably distinguish it from a patient. The two programs were connected to each other over ARPANET.
The subsequent lineage runs through ALICE (1995) and the pattern-matching AIML tradition, the Loebner Prize competitions, scripted commercial assistants, and from the 2010s the shift to statistical and then neural language models.
How They Work
Three broad architectures, in historical order:
- Rule-based. Hand-written patterns and responses. Predictable, auditable, brittle, and incapable of anything not anticipated.
- Retrieval-based. Selects the best response from a fixed set using a similarity model. Safer than generation, limited to what is in the set.
- Generative. Produces text token by token from a model trained on large corpora. Fluent and general, and correspondingly prone to producing confident text that is not true, since fluency and accuracy are separately determined.
The Turing Test and What It Measures
Alan Turing's 1950 "imitation game" proposed conversational indistinguishability as an operational substitute for the question of whether machines can think. ELIZA's reception is the standing objection: people are far easier to convince than Turing anticipated, which means the test measures human credulity at least as much as machine capability. Weizenbaum's point was precisely this — that the interesting finding was about the users.
Social Dimensions
Chatbots are the clearest case of technology that recruits social reflexes people cannot switch off. Reeves and Nass's "media equation" research established that people apply politeness norms, gender stereotypes and reciprocity to computers even while denying that they do. This makes conversational systems unusually powerful and unusually easy to misuse: an interface that speaks is granted trust that an interface with buttons is not.
Related questions concern disclosure (whether a system must say it is not a person), companionship products and their effects on people who become attached, and the labour that a convincingly conversational system displaces or conceals.
Relevance to Cyborg Anthropology
The chatbot inverts the field's usual configuration. Elsewhere the machine is joined to the body; here it is joined to the social reflexes, occupying the position of an interlocutor. Weizenbaum's finding — that knowing how it works does not stop it working — is one of the most robust results in the whole literature on people and machines.
Related Reading
References
- ↑ Weizenbaum, Joseph. Computer Power and Human Reason: From Judgment to Calculation. San Francisco: W. H. Freeman, 1976.
