Sociologist and conversation analyst Bob Moore and artist and UX designer Raphael Arar — researchers at IBM Research — are working at the forefront of an emerging field: Conversational UX design.
Their goal is simple: to get AI systems to converse more like humans. I had the pleasure of interviewing Moore and Arar on their conversational design research and what goes into it.
In part one of my interview series, Moore and Arar gave a rundown of the conversational design basics, which includes a library of sociology-backed conversation patterns. In part two, the pair described how they’re designing AI agents that actually understand and retain context — unlike many AI agents on the market today.
What struck me the most from speaking with Moore and Arar was just how uncharted the field of conversational design is today. I wasn’t the only one surprised by this.
Here, the two collaborators talk about what it’s like to be blazing a path in a field where new discoveries are being made every day and best practices are yet to be defined. One (ironic) reality: Many of the diverse teams working on building conversations for AI agents lacked the language to describe what they’re doing.
What did you find most surprising in your work with conversational design?
Raphael Arar: For me it was the lack of buy-in from the community. The way that people are approaching these types of problems is really all over the place. So it’s exciting in that respect because the terrain is completely open. At the same time there’s a lot of room for figuring out how we as a community can move this discipline forward to really make it something a lot better than it currently is.
Bob Moore: One thing that surprised me is that, when I set out, my goal was first to learn how to create a conversational interfaces so that I could apply conversation analysis and also come up with a set of patterns. I thought designers and developers for conversational systems could benefit from a set of patterns for things like how to open a conversation and how to close a conversation. By that I mean everything from “Hello, how are you?” or “I’ve got to go, goodbye” to asking and answering questions or telling stories or jokes — all the kinds of things that conversation analysts find interesting. I thought it would be useful for a designer to have a set of these patterns so they don’t have to reinvent the wheel.
That was and is my big goal, but what I found in then sharing these patterns with all kinds of different people — developers, designers and stakeholders — is that even before we get to those patterns, they lacked a vocabulary for talking about the parts of a conversation. And that kind of surprised me. Different teams were calling [the parts of a conversation] different things. Is this a turn, is this an utterance, is this an input, is this a topic, is this an activity?
So then I realized that just the terminology — the vocabulary that conversation analysts use — is useful, even if you don’t provide any patterns. I could see the design team struggling to describe their design. They could show it in a mock up, but in talking about the parts [of a conversation] they lacked the words and different teams were using different terminologies. So that surprised me. In retrospect, it shouldn’t have because I remember when I first started studying conversation analysis back in undergrad I was in the same boat.