Building Character for Conversational Agents: Ethos, Credibility, and Believability

[extended abstract of invited talk]

Most computer interfaces use human language to communicate with users and this fact has consequences that I feel have not been explicitly drawn out but that I think are at the core of current ethical concerns in the development of more loquacious interfaces, such as conversational agents. One consequence, which is the focus of this talk, is that computer communications, because they are acts of language, cannot be divorced from the ethics of rhetoric, and in selecting this term I am harkening back to the Greek derivative of this word, eiro, meaning "I say." Rhetoric is thus concerned with the relationship of discourse (what is said) to the character of the speaker (the I), or what the Greeks called ethos. Ethos is a complicated term, often treated as a synonym for character, reputation and persona, in the classical literature, and for subject and self, in more modern discourse. Ethos is the manifestation of character in language. It is a necessary component of dialogue because an evaluation of a person's character enables the listener to judge the truth of what is being said.

It is my thesis that the concept of ethos enlarges our understanding of what it means to construct a believable character for conversational agents. I begin by exploring the disintegration of ethos (the ground of credibility and trust) when a human speaker is entirely removed from the words that are being spoken, as is the case with conversational agents. I then suggest that re-constructing ethos within artificially generated dialogue requires more than giving an agent a discernable personality: it requires that we design agents that have the ability to perceive and to shape the ethos of their human conversational partners, that we make transparent within the agent the ethos of the underlying agency employing the agent, and that we make an effort to build a reputation for the agent that is centered in responsible design.