When Georgia Tech Professor Ashok Goel told 400 students in the Knowledge-Based Artificial Intelligence course at the School of Interactive Computing that his teaching assistant was Jill Watson, they quickly discovered that not only did Jill never seem to sleep, she was also very focused and seemed to know exactly how to help.
What they didn’t realize was that Jill is the world’s first artificially intelligent teaching assistant (TA) and that IBM Watson was delivering the support.
Jill appears super smart, answering about 80% of the 10,000 course-related questions asked each year. She is supported by a small team of actual tutors who directly help students, and these interactions are added to the database so the next student with a problem the humans dealt with first can get help from Jill instead.
It’s how Jill Watson is “trained” and how she learns.
Although Jill was the first, she has been joined by many other chatbots and machine-language support engines to support student learning in schools, colleges and universities around the world.
Intelligent chatbots and agents represent a significant opportunity to improve the student experience while reducing, over time, instructors’ workload. There is also great opportunity to improve student services in general and to provide extra supports for students with disabilities. Although inexpensive, they represent a new phase of the deployment of smart technologies in higher education.
How Chatbots Work
Jill Watson had an advantage. Professor Goel had a catalogue of frequently asked questions and answers for her to start with and learn from. Jill also had human help. Each time a “new” question was asked, Jill’s answer could be checked by a real person before being released to the student. More recent AI-enabled systems can learn “on their own.” For example, Google’s DeepMind taught itself to play the complex board game GO in just a few days. After three days of continuous play against itself, it then beat the world champion 100 games to zero.
Machine learning is now being used for a variety of applications, especially in health care where both IBM Watson and DeepMind are being used to “read” medical records, MRI images and x-rays, and diagnose patients as reliably as expert doctors. Because the systems can absorb the findings of hundreds of research papers published in a variety of languages, such systems can also keep themselves up to date.
“Feeding” information to some of the simpler chatbots is one method of training. The simpler version of IBM Watson, for example, can be trained this way. More advanced intelligent systems can either be fed information or employ self-learning. For example, they can study course materials and related readings and be able to reliably respond to many course questions.
Who Pays and What Does It Take?
Colleges and universities buy site licenses for the specific intelligent system or chatbot they are planning to deploy, which can then be adapted for different programs and courses. Costs vary by platform and scale of use.
The key is the cost of ensuring the intelligent tutoring system has the knowledge base it needs to support students. EDUCAUSE estimates it takes about 25 hours of instructor time per hour of instruction to train a system to do the work well. If, however, this work is “crowdsourced” with experts around the world, the time investment and costs fall. Some systems, such as RiPPLE, have been developed to enable this. Access to “self-learning” engines like DeepMind is not yet widely available.
What Are Chatbots Being Used For?
One Canadian university has deployed a chatbot to improve the quality of service provided by the Office of the Registrar. In a typical year, the office deals with 80,000 e-mails, 71,000 phone calls and 37,000 in-person visits. By deploying IBM Watson, it has been able to streamline what it does, providing 24x7 service and improving student satisfaction. St. Lawrence College and Seneca College, as well as a large number of US institutions, also use chatbots this way.
Bolton College in Lancashire, in the UK won a Beacon Award from the Association of Colleges for its pioneering use of its chatbot Ada. Deployed across the community college, Ada supports students in their academic and practical work. JISC, the UK’s technology support organization for higher education, calls it the “Alexa of learning.” Like Jill Watson, Ada uses machine learning for its services, which include employability services, careers guidance and financial advising.
Lancaster University, also in the UK, is using voice technologies integrated with Amazon’s Alexa as a support service. Students can check their schedules, book a practice room or sport facility, be reminded of assignments that are due, and ask for their latest grades to show up on their smartphone or Alexa device as soon as they’re posted. The UK’s Staffordshire University employs similar technology with its Beacon services, while Deakin University in Australia has DeakinSync.
Some institutions have deployed chatbots in a different way. Leeds Beckett University in the UK uses its service to support the admission process, helping students navigate the national system and local conditions — and even going as far as making and confirming offers to students. At this time, the system known as Becky supplements human supports, but will, replace them over time.
Others, such as the University of Grenoble, have deployed chatbots to support students with disabilities and special needs. Students can use audio, text (enhanced) and other inputs such as braille to communicate with the system. Since it is available 24x7 and anywhere, it has made a real difference to the quality of service. Microsoft Cognitive Services is also partnering with the University of Canberra to increase the range of AI supports for students and others living with disabilities.
The Real Potential: Subject Matter Expertise
Professor Goel says the real potential lies in subject matter expertise to help students at the course level. One powerful and popular system for this work is TutorMe. Like Jill Watson, this machine-learning engine is supported by expert tutors from around the world who help TutorMe develop its capacity and expertise. Using a model in which learners pay monthly for a specific level of support, dedicated resources and tutor time become available linked to specific subjects. A wide range of subjects at various levels is available and the list is growing. Not linked to a specific college or university, TutorMe is available to anyone who wishes to sign up: students, instructors, instructional designers. Each user pays for her or his own time-based license.
Institutions are starting to look at TutorMe as a support for learning, especially in STEM subjects and those known to be especially challenging, such as calculus and statistics. The idea is that the institution buys a license for all students in a program or course for a certain amount of time, which the student can then top up.
AI-enabled tutors use a variety of approaches to support learner understanding and problems. Some, like TutorMe, are focused on content mastery. Others are added on to adaptive learning engines within LMS systems. They use students’ quiz or test scores to advise on a learning pathway and provide the learning material that will help them bridge the gap between how they performed at that stage of the course and how they were expected to perform. An example of this adaptive learning approach is Cognii, which refers to its engine as offering “adaptive personalization.” Similar tools are emerging elsewhere, including Korbi, Mika, Pearson’s Aida and others.
Instructors and their teaching assistants are not the only ones “feeding” quality information to these apps and chatbots. Students are working collaboratively to do the same thing. Using a space called Brainly, some 200 million learners from 35 countries are curating content and learning responses to questions asked. AI and qualified experts support the app (for a fee for the ad-free version), providing unlimited answers to questions learners ask, with answers being verified either by a person, AI or both before being shared in the learning community.
A question such as “What is the equation used to measure liquid pressure?” or “Why did Norway go to war with Britain and fight the battle of Vagan in 1665?” produces a verified answer in seconds. The community of inquiry lives through this app, at least in part.
A related development is the provision of automated marking and feedback systems, the first of which debuted around 1998, with many of the earlier tools focused on marking and later adding feedback components to the marking regime. Several institutions, including the University of Toronto, University of Waterloo, Northwestern University, Michigan State University and Georgia Tech, use Crowdmark, mainly to automate grading but also to provide richer feedback faster. Over time, Crowdmark will add AI capacities to its functionality — something the free-to-use TAO already does. These technologies enable us to consider assessment on demand with automated test development, deployment and marking.
As learning management systems evolve, many features that are now “add-on” services — including almost all of those described here — will become integrated into the LMS. One prediction is that up to half of the activities students and instructors engage in through the LMS will be AI enabled within two to three years. Blackboard, for example, has a partnership with IBM Watson and D2L’s Brightspace has already embedded some AI components in its adaptive learning engine (e.g. automated nudges).
What Does the Future Hold?
We are at the early stages of deploying AI in support of learning. There are a lot of ideas, some experiments and only a few deployments. As AI continues to develop — it has been with us now for more than 30 years — we will see more and more deployments. Chatbots and AI systems won’t replace human interaction and the work of a “real” instructor. Instead, they will enable more of the routine work to be done by AI so that the deep relationship skills of professional instructors can be used where they are needed most.
