The team’s work builds upon ASSISTments, a free math homework tool that helps teachers track student progress, developed over the last couple decades by WPI computer science professor Neil Heffernan. The platform is now used by about 20,000 teachers and 500,000 students in 20 countries, according to its website.
ASSISTments’ new AI feature, dubbed Quick-Comments, makes use of machine learning and natural language processing principles to suggest what comments a teacher could leave for students, according to Heffernan, who led the research.
Heffernan said the AI focuses largely on written-answer math problems for students, which demand a more nuanced approach than simply marking answers correct or incorrect. The AI works like Google Smart Reply in function, allowing teachers to send more feedback in less time based on the AI’s understanding of each answer.
According to Heffernan, written-answer questions have become more commonplace in math courses as K-12 schools increasingly embrace digital learning models. He said the AI gives teachers a choice between several comments for feedback and explanations about their correctness.
“Now, we’re grading them and suggesting comments to the teacher,” he said. “We’re telling the teacher, ‘Here are three options for a comment you should send to the student.‘”
Heffernan said the feature is being beta tested by a handful of teachers for about 5,000 students, mainly in elementary and middle school grade levels. He said the tool has so far proven useful, and feedback on the platform has increased as researchers work to improve the AI moving forward.
“We’re now slowly releasing it to teachers who are using the system and giving them access to it,” he said of the rollout that began in the spring. “We’ve been scaling it out as we convince ourselves, ‘People like this.’”
Ed-tech companies have invested in the development of rudimentary AI and machine learning features on platforms used for exam proctoring to detect cheating during virtual exams, as well as grading and content suggestions for course material based on aggregated data.
While AI often conjures up images of a dystopian society in which the “machines take over” and jobs are automated, Heffernan thinks AI will be more likely to serve simple functions such as these. He said there are still limitations to what AI can and can’t do in terms of feedback.
Heffernan said he’s concerned about the ethical implications of relying too much on AI in education, noting that there’s still much work to be done to eliminate biases that could impact students of color and other marginalized groups. He emphasized that the goal is to help teachers, not replace them.
“When I think of AI, I think of the small ways AI can do little things,” he said. “What AI can do well [now] is very simple things, but we can do better. We’re not good on all types of responses, but we’re very good at when the child gives an answer that’s very good or really poor.
“I don’t want the AI to do much. I just want it to help read the kids’ homework and make some suggestions for teachers,” he added. “As long as it’s doing that simple thing, then I’d feel comfortable using that.”
