Prof's Teaching Assistant Isn't Human


Georgia Tech professor Ashok Goel and some graduated students created an AI teaching assistant they named Jill Watson by feeding the AI about 40,000 forum posts from previous years, as well as the answers. Cultura RM Exclusive/Frank and Helena/Getty Images
Georgia Tech professor Ashok Goel and some graduated students created an AI teaching assistant they named Jill Watson by feeding the AI about 40,000 forum posts from previous years, as well as the answers. Cultura RM Exclusive/Frank and Helena/Getty Images

Imagine going to a college class and realizing one of your professor's teaching assistants was a former champion on "Jeopardy!" The champion that was a computer, that is. That's what happened in professor Ashok Goel's classroom at Georgia Tech.

Goel routinely teaches knowledge-based artificial intelligence and receives help from eight human teaching assistants and now one famous AI program: IBM's Watson. Watson is the computer that beat two returning "Jeopardy!" champions in a special series of episodes. It also acts as a platform for medical research and can recommend some pretty weird recipes for your next dinner party.

The professor decided to use Watson to answer questions posted to an online forum. The TA, known as "Jill Watson," helped out by responding to student queries. This helped solve a problem: Students were leaving more questions than Goel and his human assistants could answer efficiently.

Along with some grad students, Goel created Jill Watson by feeding the AI about 40,000 forum posts from previous years, as well as the answers. It turns out that although the number of questions increases with the number of students in a class, the types of questions would tend to fall into the same categories. In other words, you just end up with more people asking the same things, something an AI program can deal with fairly easily.

It didn't start out that smoothly, though. At first, Jill's answers were a little suspect. Fortunately, Goel and his team had all the answers publish in a forum that wasn't accessible by students. This way, the team could monitor Jill's performance and make adjustments. Jill's early responses focused too narrowly on keywords, but with time the team improved the AI's ability to answer questions.

In time, Jill's ability to create a relevant response hit a 97-percent success rate, and the team let the AI loose. The students who interacted with Jill weren't told the TA was a computer program until late April 2016.

Their reaction? They loved it.