AI teaching assistants and the future of machine learning

A boy looking at the monitor of the computer.

By Dan Heuskin

Published March 8, 2017 This content is archived.

In Spring 2016, students at Georgia Tech didn’t realize that the teaching assistant answering their email inquiries was powered by artificial intelligence. What could this mean for research universities like UB and beyond?

Print

Photo

Dan.

Daniel Heuskin (UB Student, Class of 2017) is originally from Long Island, NY. He is studying English at UB with aspirations to become a teacher or writer. In his free time, he enjoys playing bass guitar, doing nature photography, writing songs and reading.

Meet Jill Watson. She—or, more accurately, it—is a virtual TA based on IBM’s Watson platform.

Jill’s life began as a potential solution to the problem of insufficient teaching support in a Massive Open Online Course (MOOC) on Knowledge Based Artificial Intelligence (KBAI) at the Georgia Institute of Technology.

“We created Jill as a way to provide faster answers and feedback,” explains Ashok Goel, the Georgia Tech professor who teaches the course, in a 2016 news article by Science Daily.

With thousands of local and online enrollees, the MOOC elicits some 10,000 forum messages from students every semester--and without Jill, Professor Goel and his eight human TAs couldn’t possibly answer every student question in a timely manner.

Goel and his TAs started working on Jill in 2015, feeding her every question that had ever been asked in KBAI’s online discussion forum along with accurate solutions. They wrote code allowing Jill to answer routine queries, such as where particular readings and assignments could be found. 

A slow climb

When Jill was first implemented as a virtual TA in January 2016, her responses to student inquiries were often irrelevant to the questions being asked.

“Initially her answers weren’t good enough because she would get stuck on keywords,” explains Lalith Polepeddi, a graduate student who contributed to Jill Watson’s development. “So we learned from mistakes… and gradually made Jill smarter.”

With some modifications, Jill was soon answering student questions with 97 percent certainty. By March 2016, she was responding to students without any intervention from human TAs.

Most students didn't know they were interacting with an AI until April 2016, when Goel revealed Jill’s true identity. The revelation was met with both astonishment and enthusiasm. Some alumni of the KBAI course even organized a forum to keep up on Jill’s latest developments, while others developed an open source initiative designed to replicate the virtual TA. 

What does it mean?

If other research universities, like the University at Buffalo, were to design similar AIs, the field of applied artificial intelligence might advance at an increasingly rapid pace.

And why not UB? "Here at UB, we have the faculty and IT staff expertise to innovate in the fields of AI and robotics," said J. Brice Bible, UB Vice President and Chief Information Officer. "I expect to see pilot projects in the near future from interested UB faculty in exactly these areas."

But development as swift as Jill’s would be possible only under certain circumstances.

“An important element in Jill’s rapid development was the database of thousands of previously answered queries from which to learn,” explains Peter D. Scott, retired professor of Computer Science and Engineering, Electrical Engineering, Physiology and Biophysics at UB. “This could only be replicated by MOOCs and other courses with very large enrollments. In smaller courses, AI-based TAs would learn much more slowly.”

It isn’t hard to imagine a future in which AI has become deeply integrated in students' everyday lives. Imagine a team of virtual TAs, each with a unique name and email address, fielding student questions while constantly improving their own comprehension and question-answering capabilities.

“Just as with evaluation of human TAs, some virtual TAs would be graded more favorably and some less favorably by the students they’ve worked with,” explains Professor Scott. “This would further accelerate their learning of student needs, and reveal how students like to interact with AI-TAs.”

Eventually, the use of AI in similar roles might expand into the world of business and public service. Forget those annoyingly unnatural automated phone systems that can’t listen to what you say; in the not-so-distant future, your phone calls might be answered by sophisticated, intuitive AIs—and you might just mistake them for human.