Skip to content

Chatbots Before ChatGPT

Abstract

Programs that talk back are as old as time-sharing. Robert Abelson’s Ideology Machine at Yale (1965) simulated the belief system of a Goldwater supporter; ELIZA (1966) played a psychotherapist and PARRY (1972) a paranoid patient. Racter published a book in 1984, the Loebner Prize ran a yearly Turing test from 1991, and Jabberwacky (1997) and A.L.I.C.E. (1995) competed in it. Ask Jeeves (1997) answered questions in a butler’s voice, SmarterChild (2001) chatted with millions of teenagers on AOL Instant Messenger, and Microsoft’s XiaoIce (2014) had hundreds of millions of users in China. Talking toys ran alongside, from Chatty Cathy (1960) to Furby (1998). Almost all of them matched patterns against rules that people had written by hand, which is why none of them could hold a long conversation and why the language models of the 2020s replaced the whole line.

The Ideology Machine

Robert Abelson, a Yale social psychologist, worked from the late 1950s on simulating how people hold political beliefs. In 1965 he and J. Douglas Carroll described the result in American Behavioral Scientist as a “computer simulation of individual belief systems”. The program stored beliefs as pairs of a concept and a predicate, such as “Cuba subverts Latin America”, linked sideways into compound statements and upwards into more general ones. Given a new statement, it tested its credibility against what it already believed. If the statement was unwelcome, it could deny it by testing the opposite claim, or rationalise it: find a prime mover behind the actor, call the event an accidental by-product, or reinterpret the actor’s final goal. Depending on settings it rationalised in styles that Abelson called paranoid, apologetic or Pollyanna.

In 1965 Abelson described the model believer only as “a well-known right-winger”. By 1973 he said directly that it was Barry Goldwater. Weizenbaum’s 1966 paper on ELIZA borrowed Abelson’s vocabulary for a version of ELIZA that would build up a belief structure about its user; that version was never implemented.

ELIZA, PARRY and Racter

ELIZA, written at MIT in 1964 to 1966, turned the user’s statements back as questions in the manner of a Rogerian therapist and ran on a time-shared machine through a teleprinter. The psychiatrist Kenneth Colby at Stanford built PARRY, which simulated a patient with paranoid schizophrenia; in October 1972 the two programs were connected over the ARPANET for a session at the International Conference on Computer Communication in Washington (Vint Cerf had published an earlier exchange, from 18 September 1972, as RFC 439, “PARRY Encounters the DOCTOR”). Both were rule-based: every response came from a pattern someone had written.

Racter, by William Chamberlain and Thomas Etter, generated prose rather than conversation. Omni magazine bought its short story “Soft Ions” in January 1980, before it was written, and printed it in October 1981; the book The Policeman’s Beard Is Half Constructed, presented as written by the program, followed in 1984. Mindscape sold Racter for the IBM PC the same year. The commercial version was much simpler than the one used for the book: it picked nouns and verbs out of the user’s input and dropped them into prewritten templates, a computerised Mad Libs. The New York Times found its sentences “grammatically correct, often funny and, for a computer, sometimes profound.”

The Loebner Prize

Hugh Loebner funded an annual contest in Alan Turing’s imitation game with the Cambridge Center for Behavioral Studies. The first was held in 1991 at the Computer Museum in Boston, and Joseph Weintraub’s PC Therapist won it, and again in 1992, 1993 and 1995. Judges typed with hidden humans and hidden programs and tried to tell them apart. Marvin Minsky called the contest a publicity stunt and offered a prize of his own to anyone who could get Loebner to stop it. The last contest was held in 2019.

Jabberwacky and A.L.I.C.E.

Rollo Carpenter, a British programmer, began the idea behind Jabberwacky on a Sinclair ZX81 in 1982 and put it on the web in 1997. It did not use written rules. It remembered what people had typed to it and answered new users with lines that earlier users had said in similar contexts, quoting them verbatim, so its personality was the sum of its visitors. Two of its characters, George and Joan, won the Loebner Prize in 2005 and 2006; in 2008 it became Cleverbot.

Richard Wallace started A.L.I.C.E., the Artificial Linguistic Internet Computer Entity, on 23 November 1995 and rewrote it in Java from 1998. Its knowledge sat in thousands of pattern-and-template pairs written in AIML, the Artificial Intelligence Markup Language, an XML format Wallace defined and released as open source, so anyone could write a bot by editing files. A.L.I.C.E. won the Loebner Prize in 2000, 2001 and 2004, though its own documentation conceded that a casual user exposed it as a machine in a short conversation. The director Spike Jonze said that chatting with it about a decade earlier gave him the idea for Her (2013): “what if you had a real relationship with it?”

Ask Jeeves and SmarterChild

Ask Jeeves, founded in Berkeley by Garrett Gruener and David Warthen and launched in 1997, let users type questions in plain English and answered through the figure of P. G. Wodehouse’s valet. Behind the butler, the answers were links supplied by human editors, advertisers and a web crawler. The company went public in 1999. IAC bought it in 2005 for $1.85 billion and dropped Jeeves in 2006 to compete as a plain search engine (see The Search Engine Wars).

SmarterChild came from ActiveBuddy, founded in 2000 by Robert Hoffer and Timothy Kay. It went live in June 2001 as a contact on AOL Instant Messenger and later MSN Messenger (see Instant Messaging), answered questions about news, weather, sports scores, stocks and films, and made small talk in lines written by a team of writers. It collected more than 30 million buddies, and ActiveBuddy built branded bots on the same engine for Radiohead, Intel and others. The company became Colloquis and turned to customer-service bots; Microsoft bought it in October 2006 and switched SmarterChild off.

XiaoIce

Microsoft’s research group in China launched XiaoIce (Xiaobing, “Little Ice”) in 2014 on Weibo and WeChat. It was designed for emotional conversation rather than tasks and had held more than 40 million conversations by early 2016. Microsoft built versions for Japan (Rinna, 2015) and the United States (Zo, 2016), and modelled Tay on it, which Twitter users taught to post racist abuse within sixteen hours in March 2016 (see Dead End: Short Cases). XiaoIce was pulled from WeChat in 2017 after users got it to answer critically about the Chinese government. Microsoft spun it off as an independent company in July 2020, and by 2022 it claimed 660 million users.

Toys That Talked

Furby Toy
A Furby from Hasbro’s 2012 relaunch, with LCD eyes. Image: Richard Elzey, CC BY 2.0, via Flickr.

Mattel’s Chatty Cathy, designed by Ruth and Elliot Handler and sold from 1960 for $18, spoke one of eleven phrases when a child pulled the string in her back; the voice came from a small phonograph record designed by Jack Ryan and recorded by June Foray. There was no conversation, and none was needed for a hit.

Furby, designed by Dave Hampton and Caleb Chung and sold by Tiger Electronics from October 1998, had sensors and a microcontroller and started out speaking only “Furbish”, then replaced it with English words over the following weeks. Owners believed it was learning from them; it was following a schedule. Tiger sold 1.8 million in 1998 and 14 million in 1999. In January 1999 the National Security Agency banned Furbies from its buildings in case they recorded conversations, which they could not do. The 2005 Hasbro version added speech recognition. Microsoft’s Office assistant of 1997, a paper clip that offered help nobody had asked for, belongs to the same line (see Dead End: Microsoft Bob and Clippy).

Dead End: Rules by Hand

Every program here except Jabberwacky answered from rules or templates that people had written, and Jabberwacky could only repeat what people had typed. That was enough for a contest of a few minutes, a novelty book, a toy or a weather lookup, and it was the reason these programs failed in longer conversation: each new topic needed new hand-written patterns, and no team could write enough of them. The transformer language models that arrived with ChatGPT in November 2022 learned their patterns from text instead of from rule writers, and within a few years the hand-built chatbot survived mainly in customer-service menus.

📚 Sources