chatterbots

The famous Turing test (see Turing, Alan M.) proposes that if a human is unable to reliably distinguish mes-sages from a computer from those of another person, the computer program involved can at least be provisionally declared to be “intelligent.” The advent of textual commu-nication via the Internet (see texting and instant mes-saging) has afforded a variety of ways to attempt to meet this challenge. Programs that mimic human conversational styles have come to be known as “chatterbots.”

The prototypical chatterbot was ELIZA, developed by Joseph Weizenbaum in the mid-1960s (see Weizenbaum, Joseph). ELIZA mimicked a form of nondirective psychother-apy in which the therapist echoes or plays off of the client’s statements as a form of gentle encouragement and validation. Thus if one types, “My father didn’t really like me,” ELIZA might reply, “Tell me more about your father.” Although primitive, ELIZA once inadvertently fooled an executive into thinking he was exchanging messages with Weizenbaum. Other classic chatterbots include Parry, designed to mimic a paranoid, and the story-generating Racter.

The first online chatterbot, Julia, was created by Michael Mauldin in 1994. (Mauldin also coined the term chatterbot.) More sophisticated chatterbots include Jabberwocky, which includes learning algorithms, and ELLA, which applies linguistic principles to produce more realistic speech (see natural language processing). On a negative note, some chatterbots have been designed to enter chat rooms and spew advertising (spambots) or even to elicit personal infor-mation (see identity theft and phishing and spoofing).

One of the best known and most versatile modern chatter-bots is ALICE (Artificial Linguistic Internet Computer Entity), developed by Richard Wallace. ALICE does much more than simply echo statements; it includes an elaborate parser and sentence generator and specialized scripts for initiating or responding to conversational gambits. ALICE can even tell lies and spread “gossip” that it learns from other users.

In the following dialog ALICE seems to be able to respond appropriately:

Human: Are you intelligent?

ALICE: I am highly intelligent. I have many intellectual functions. Human: How do you know you are intelligent?

ALICE: By introspection.

However the illusion breaks down as the matter is pressed further and the bot uses a “deflection” strategy:

Human: What were you thinking while you said that? ALICE: I was minding my own business.

Human: How do you feel about being interrogated in this way? ALICE: The opposite of not being interrogated in this way.

Still, with more than 25,000 conversation templates stored in XML files that are modifiable by the user, ALICE has repeatedly won the Loebner Contest, a modern version of the Turing test.

Although many are created as programming exercises or for amusement, chatterbots embody principles that are important in artificial intelligence research, including natu-ral language processing and machine learning (see arti-ficial intelligence). Techniques first developed with chatterbots can contribute to the creation of programs designed to provide answers to users’ questions or other forms of assistance (see software agent).