Designing for “Smart” Agents Among Us

Earlier this week, Google demonstrated Duplex, an astonishing advance in human-computer interaction. If you haven’t seen the relevant part of Sundar Pichai’s presentation, please watch it now:

If you understand how computers work, you’ll know how difficult it is for computers to do what Duplex is doing in this video. The system seems to be forming accurate models of the evolving contexts it’s participating in. It also shows nuance in communicating back to its human interlocutors, injecting “ums” and “ahs” at the right moments. Again, all of this is very difficult. (That’s why the audience laughs at several points in the presentation; they know how improbable this thing they’re hearing is.)

It’s worth noting this demonstration doesn’t suggest an artificial general intelligence like HAL 9000 or C-3PO; Duplex seems to be modeling a relatively narrow area of human interactions. (Namely, making an appointment.) Still, the system sounds convincingly human, and that raises deep questions. The (human) interlocutors seem not to be aware that they’re talking to an artificial entity. Are they being manipulated? (Google has said the system will be transparent to the people who interact with it, but this didn’t come across in the demonstration.) What would widespread availability such technology do to relations between human beings, to our ability to empathize with others? What would it do to social inequity?

We’re not far from the day when interactions with convincingly-sounding artificial agents are commonplace. We will both deploy agents to do our bidding, and interact with agents that have been deployed by others to do theirs. Both scenarios will play out in information environments. What affordances and signifiers are required? How will we balance transparency and seamlessness? (And how will this balance evolve as we become accustomed to engaging with these agents?) How will we structure the information environments where these human-agent encounters happen so they augment (rather than erode) human-human interactions? How will we know such erosion isn’t happening? Interesting challenges ahead.