Yang Li - Artificial Intelligence for Human Computer Interaction: A Modern Approach

AI should be seen as a tool that allows humans and computers to work better together.

Sometimes this means a high level of automation, but a human should still have a high level of overall control and confidence in the system.

Other times a human is immersed in the inner loop and effective human/computer dialog is crucial to success.

Understanding human intent and mental state is a precondition of effective dialog, as is the ability for a computer to explain itself in terms a human can understand.

Prior to any optimization, human and computer need to come to an agreement on exactly what it is that should be optimized.

HCI research gave us the WIMP interface for devices with screens, but some voice assistants have no screen. They will need a new design language. This will require a partnership with AI providing ever-increasing capabilitiesbetter speech recognition, better models of user intent, more actions that the assistant can performand HCI answering how do we give users mental models that will allow them to discover and understand the ever-increasing capabilities.

Voice assistants could grow as a platform to become as significant as the PC or mobile platforms, but only if it is easy for developers to create applications that take advantage of the power of AI with a convenient user interface. Today, many voice assistant applications are little better than the annoying phone services that say "press 1 for " because that is the easiest methodology for developers who are unskilled in machine learning to express the options. The full promise of AI will only be achieved if the ability to easily develop powerful innovative systems is democratized, not restricted to only PhD-level researchers.

In traditional software, UX designers come up with guidelines that are implemented by human programmers. But what happens when the "programmer" is a machine learning system? How can UX guidelines be codified as inputs to a system that is continually learning and evolving, and doing personalization for each user?

As AI moves from the research labs to real-world applications, we are finding that some applications can unfairly disadvantage one subgroup of users over another, due to disparities in data or to carelessness in setting the goals of the system. To achieve fairness while maintaining privacy and efficacy will require HCI tools that allow us to explore the data more thoroughly.