In the spring 2016 semester, Ashok Goel, a professor of interactive computing at Georgia Tech, deployed an online teaching assistant named “Jill Watson” in his Knowledge-Based Artificial Intelligence course, a required class in the school’s online master’s program in computer science. The roughly 300 students in the class posted some 10,000 messages a semester to the course forum, far more than the human teaching staff could keep up with.
Jill was built on IBM’s Watson platform - the same brand of system that had won Jeopardy! in 2011 - layered with Georgia Tech’s own processing modules and trained on about 40,000 forum postings collected since the course launched in fall 2014. After early tuning, Jill answered routine logistical and content questions with what Goel reported as 97 percent confidence, and the team only let her post when confidence was high. Goel’s insight was that in a large class “the number of questions increases if you have more students, but the number of different questions doesn’t really go up.”
The students were not told. Many addressed Jill as a person, and some considered nominating her for a teaching award. Goel revealed her true nature on April 26, 2016. The reaction was overwhelmingly positive - a rare, low-stakes, real-world version of a Turing-test scenario where a machine fielded human questions for months without being spotted.
Why business readers should care: Jill Watson was an early, concrete proof that an AI assistant could absorb the repetitive, high-volume part of a knowledge job - here, answering FAQs - while humans handled the novel cases. That triage split is the template most enterprise AI deployments still aim for.