SXSW Day 7 session 1: AI - Actually still terrible

Kate Darling, MIT Media Lab and Nilesh Ashra, Creative Director at Wieden+Kennedy Lodge.

There is a lot of hype from companies around what AI can and can't do, focusing on successes, less on failures.  The discussion around AI has moved beyond research and practicality into the domain of hype.  This hype is potentially distracting from real valuable applications.
There is also confusion between AI, insofar as analytics, and human-like interaction.  The latter is a much harder problem than the former.  It's much easier to build AI that mimics the analytical workings of the human mind, than the interactive side of it.  At this, we are still relatively far.
As an example, call center speech bots can still be very frustrating; they are script based, and can navigate a conversation in a very limited domain.  Small changes in speech patterns can confuse the AI and so interaction with it can be very frustrating.  Humans have better referential context capabilities ("the other one"), and are better at taking and reacting to interruptions.

Being able to improve the way AI interacts with humans will require a lot more research in that domain.  A few examples of this type of interaction research:

• Hitchbot - an early experiment in having a robot interact in an open environment with humans.  Hitchbot was created as a social experiment in a Canada university.  The Robot, which had very basic humanoid form (basically a largish bucket with head, hands and legs), with very rudimentary communication skills, and a mission to hitchhike first across Canada, then other places.  The research here was very rudimentary, just GPS-tracking Hitchbot, and tracking people's posted reactions to it in social media.  After going across Canada, Germany and the Netherlands, it was vandalized while trying to cross the US.
• Needybot - Wieden Kennedy advertising agency built a robot and that has basic mobile but can't really do anything by itself, and needs to ask everyone around it.  This one was more sophisticated, having a camera and including face recognition, so it could identify people it met.  Wieden Kennedy set it loose in their offices.  The experiment focused on interaction impacts, such as that people like being recognized by robots or that silent robots don’t generally get help.