Sign In

Communications of the ACM

Communications of the ACM

Staffing the Web with Interactive Characters

Like their human counterparts in the real world, interactive characters will soon play helpful roles, enabling customers of Web enterprises to do the things they go to Web sites to do. They will make it easy for people to interact with them through natural forms of conversation and gesture. They will add interest and warmth to the online experience through their distinctive personas, projecting images and styles designed to appeal specifically to the customers they serve. They will build relationships with customers based on familiarity, affection, and trust.

Interactive characters also will surpass their human counterparts in several important ways. They will be able to exploit the Web's vast information resources. They will deliver desired behavior consistently 24 hours a day, seven days a week. They will scale flexibly in number as needed to give immediate one-to-one service to each and every customer. They will remember important features of past interactions, making customers feel known and valued. They will set a new standard for customer service.

We have built several such interactive characters, including Max, who guides visitors at Extempo's Web site; Merlin, who guides interactive neighborhood tours at Haight Street on the Web; and Jennifer James, who works as a spokesperson at Samson Motors on Extempo's Web site. All are built on the component-based architecture in Figure 1. The Extempo Improvisational (Imp) engine runs character mind software, providing basic cognitive functions for multiple characters and character instances running simultaneously. Application programming interfaces (APIs) support integration with various knowledge sources. For example, a Web-guide character accesses information from a tour database or from a site's back-end systems. It constructs a personalization database to record and retrieve information it has learned about customers. And it may use the Internet to retrieve information for its dialogs or to send and retrieve email.

The mind of an Imp character consists of three pairs of plug-and-play competence/content components: a unique and persistent persona; one of several alternative roles; and the knowledge needed to perform its role in a specific application. Persona competence and content components involve all facets of a character's identity: backstory; appearance; manner of moving, gesturing, and speaking; voice; knowledge and opinions; and social and emotional dynamics. A character's persona is revealed through manner, dialog, and behavior. Role competence and content components are the knowledge and skills the character needs to perform his or her job. For example, a character acting as a Web guide has to greet each customer, determine the details of the right "tour," act as escort to the prescribed series of destinations, and offer value-added commentary. Application components involve the knowledge needed to perform this role in a given context, as in guiding tours at vs. at

The Extempo Mind-Body Controller mediates two-way communication between a character's mind running in the Imp engine and his or her body running on a user's desktop machine. Related APIs allow alternative body media, user input modes, and integration with various browsers or applications.

Extempo Character-Configuration Tools allow authors to build new characters by creating, selecting, and automatically configuring stock components from a library. Software developers create competence components; writers and artists create content, knowledge, and media components.

Our work on interactive characters is related to work in several disciplines. For example, our characters are a type of intelligent agent differing from traditional agents in their multifaceted cognitive capabilities, embodiment, interactive style, and relationship to the user, as well as the criteria used to evaluate their success [1].

Our component-based architecture has its roots in an architecture we developed previously for traditional intelligent agents operating in various domains. Our approach to natural language processing is a conceptual descendant of the Eliza program [5] and a cousin of the more recent Julia program. However, our conversation model is integrated with models of character role and persona, employing the cognitive structures and skills in our underlying character architecture. Our characters demonstrate a kind of applied psychology designed to manifest aspects of personality, intelligence, linguistic style, and social skills [3, 6].

Our characters continue the pioneering efforts of animators Walt Disney and Chuck Jones to create "believable" characters who "appear to think and make decisions and act of their own volition" [4]. Our characters also follow certain principles of improvisational acting and have job skills for providing the one-to-one service required in customer-oriented enterprises [2].

The Turing test. There is a long history behind the question of whether machines can "think" or achieve "human-level intelligence." In 1950, Alan Turing, the British mathematician and computing pioneer, proposed that we replace logical debate, which often regresses into arguments on the nature of thought or intelligence, with a behavioral test. A computer passes the Turing test if a person conversing with it via a keyboard believes he or she is conversing with a human being. However, passing the Turing test is not the goal of our work on interactive characters. It is not necessary, sufficient, nor even desirable for our characters to pass the Turing test in order to do their jobs well. In fact, people might feel uncomfortable in a world populated by automata masquerading as humans.

Instead, we propose a new behavioral test to measure the success of interactive characters. A character passes the Extempo test if a person interacting with the character over a communication medium willingly suspends disbelief that he or she is interacting with a human being. Our test differs from the Turing one in all three of its critical elements: It replaces the purely linguistic conversation of Turing (and today's Loebner Prize) with a broader class of interactions between the person and the character that may include natural language conversation as well as the actions, gestures, and facial expressions or body language. Our test replaces the teletype of Turing's test with any communication medium, permitting multimodal exchanges through, say, visual and auditory perception of both words and images. Most important, our test replaces the belief that the character is a human being, as specified by Turing, with a willing suspension of disbelief that the character is a human being.

In other words, the human user might well know that the character is not a human being and say so if asked. But the user would not care and would not intentionally or significantly alter his or her own behavior on that account within the context of the interaction. In this respect, our test resembles the artistic criteria we apply to evaluate characters in books and animated films and to actors playing roles in films, plays, and TV shows. People know these characters are not real human beings, but when such characters are rendered or acted well, the audience willingly suspends its disbelief in order to have the pleasure of responding to them as if they were real. Similarly, interactive characters pass our test when people willingly suspend disbelief in order to have the pleasure, convenience, or other desirable consequence of interacting with them as if they were real.

Toward a commercial mass medium. Interactive characters offer an unprecedented opportunity for developing two-way learning relationships between enterprises and customers. They apply humanlike skills to engage customers in rich conversational interactions. And they apply electronic technologies to record complete transcripts of these interactions. Like any computer-mediated transaction, there is a potential for abuse of recorded information and a need for both real and perceived protection of privacy. However, these records also support the efforts of enterprises that want to improve their understanding of their customers and their effectiveness in serving them.

Interactive characters also present a special opportunity to exploit the premier branding vehicle—characters. By interacting directly with customers, branded characters can help transform the Web from a high-tech playground into a commercial mass medium competing aggressively with TV, radio, and print. Characters appear in all of these media; they will delight customers by coming alive to interact with them on the Web.

Back to Top


1. Elliott, C., and Brzezinski, J. Autonomous agents as synthetic characters. AI Mag. 19, 2 (Summer 1998), 13–30.

2. Peppers, D., and Rogers, M. The One to One Future. Doubleday, New York, 1993.

3. Reeves, B., and Nass, C. The Media Equation: How People Treat Computers, Televisions, and New Media Like Real People and Places. Cambridge University Press, New York, 1996.

4. Thomas, F., and Johnston, O. The Illusion of Life: Disney Animation. Hyperion Books, New York, 1981.

5. Weizenbaum, J. ELIZA—A computer program for the study of natural language communication between man and machine. Commun. ACM 9, 1 (Jan. 1966).

6. Zimbardo, P., and Leippe, M. The Psychology of Attitude Change and Social Influence. McGraw-Hill, New York, 1991.

Back to Top


Barbara Hayes-Roth ( is founder and CEO of Extempo Systems, Inc. and director of the Adaptive Agents Project in the Computer Science Department at Stanford University.

Vaughan Johnson ( is engineering manager at Extempo Systems, Inc.

Robert van Gent ( is technology manager at Extempo Systems, Inc.

Keith Wescourt ( is director of product management at Extempo Systems, Inc.

Back to Top


This work is supported by NIST Advanced Technology Program Contract #70NANB5H1161.

Back to Top


F1Figure 1. Extempo component-based Imp architecture

Back to top

©1999 ACM  0002-0782/99/0300  $5.00

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.

The Digital Library is published by the Association for Computing Machinery. Copyright © 1999 ACM, Inc.


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account
Article Contents: