A persistent mismatch has existed between the characteristics of
AI-driven features, current interface design-languages, and end-user
mental models. In simpler times we could joke that UI's were writing a
check the AI couldn't cash. AI systems couldn't quite behave in the
way UI designers wanted—or promised—they would. In this talk, I'll
illustrate how the problem has also flipped: AI designers are now
overpromising things interfaces can't deliver. Decision-theoretic
constructs such as `mixed-initiative' provide broad strategies—but not
tactics—for addressing this problem. While we have evolved a set of
examples for masking, controlling, and explaining, as ways to manage
the AI-UI bridge, a more generic design language (patterns & UI style)
is elusive. As more AI-driven features become integrated into
user-facing systems, the development of a modern design language is
critical. I'll try to lay out the value (and challenges) of such a
language. If done correctly, both end-users navigating the
uncertainty/noise of AI-features and data-hungry AI subsystems will
benefit.
Eytan Adar is an Associate Professor in the School of Information &
Computer Science and Engineering at the University of Michigan (though
he’s currently on sabbatical at Adobe and UW). He works broadly at the
intersection of HCI and IR/Data Mining and ranges from empirical
studies of large-scale online behaviors to building new systems, tools
and methods. He completed his doctoral work in the Computer Science at
the University of Washington and has a Masters and Bachelors degree
from MIT. Before going back for his PhD, Eytan was a researcher at HP
Labs and Xerox PARC for a number of years (spinning out a company
called Outride somewhere in there). Eytan is co-founder of ICWSM and
has served as general chair for ICWSM and WSDM. His website is at
http://www.cond.org