What is Conversation,
And How Can We Design For It?
Dubberly Design Office | firstname.lastname@example.org
CyberneticLifestyles.com | email@example.com
[ 1] Dubberly, H., U.
Haque, and P. Pangaro.
“What Is Interaction?
Are There Different
Types?”, interactions 16,
no. 1 (2009): 69-75.
July + August 2009
Interaction describes a range of processes. A previous “On Modeling” article presented models of
interaction based on the internal capacity of the
systems doing the interacting [ 1]. At one extreme,
there are simple reactive systems, such as a door
that opens when you step on a mat or a search
engine that returns results when you submit
a query. At the other extreme is conversation.
Conversation is a progression of exchanges among
participants. Each participant is a “learning system,” that is, a system that changes internally as
a consequence of experience. This highly complex
type of interaction is also quite powerful, for conversation is the means by which existing knowledge is conveyed and new knowledge is generated.
We talk all the time, but we’re usually unaware
of when conversation works, when it doesn’t, and
how we can improve it. Few of us have robust
models of conversation. This article addresses the
questions: What is conversation? How can conversation be improved? And, if conversation is important, why don’t we consider conversation explicitly
when we design for interaction? This article hopes
to move practice in that direction. If, as this forum
has often argued, models can improve design, we
further ask, what models of conversation are useful for interaction design?
We begin by contrasting “conversation” with
“communication” in a specific sense. We then offer
a pragmatic but not exhaustive model of the process
of conversing and explore how it is useful for design.
What Isn’t Conversation?
Claude Shannon developed a rigorous model of a
transmission channel used to convey messages
between an information source and a destina-
tion. While his context was analog telephones
with wires highly susceptible to noise, Shannon
produced a model that applies to a wide range of
In Shannon’s model an information source
selects a message from a known set of possible
messages, for example, a dot or a dash, a letter
of the alphabet, or a word or phrase from a list.
Human communication often relies on context to
limit the expected set of messages. If you receive
a call from a friend (the source) arriving by train,
you expect to hear “I’m getting on the train,” or
“I’m on the train,” or “the train is late,” and so
on—messages that are drawn from a set of possibilities known to both of you. The channel is effective if it enables you (the destination) to select
which of the possible messages is currently being
transmitted. (Voice communication is more than
sufficient for this, and Shannon’s interest was
highly encoded transmission. But this simplified
example draws useful distinctions for the discussion that follows.)
Communication in the sense of distinguishing
among possible messages known in advance is
important for much of our daily life. It allows us to
synchronize a wide range of actions with others.
But it has limits. Shannon’s model captures a fundamental limit of nearly all human-to-computer
interaction: Our input gestures can only activate
an existing interface command (select a message)
from the preprogrammed set. While we can automate sequences of existing commands, we can’t
ask for something novel. If our software application does anything novel, we file a bug report!
In Shannon’s model, how can we say something
novel to one another? The answer is, we can’t. It’s