Growing up in Kenya, Chris Piech saw firsthand the dearth of educational opportunity.
“You can imagine how many people in my neighborhood would have access to a professor who knows computer graphics or someone to sit down with them and explain the scientific method,” he says.
Now, decades later, Piech, an assistant professor of computer science, is part of a team at Stanford that aims to do something about that gap, by developing an AI-based system to help students learn how to complete open-ended tasks such as coding. The system could multiply teachers’ impact or even offer tutor-type assistance.
The team, experts in reinforcement learning, human-computer interaction, pedagogy, and other areas, includes Stanford researchers Emma Brunskill, Jennifer Langer-Osuna, James Landay, Dan Schwartz, and Noah Goodman. The group received one of Stanford Institute for Human-Centered Artificial Intelligence’s inaugural Hoffman-Yee grants, which will fund their work.
Worldwide, education systems are seeing a shortage of teachers and other educational resources.
According to U.N. estimates, about 69 million teachers will be needed to achieve 2030 sustainable education goals. Today, more than 260 million children and youth do not attend school, and just 14 percent of global citizens complete upper-secondary education in low-income nations.
While online educational tools can help, the lack of resources is particularly challenging for open-ended tasks such as crafting paragraphs, learning scientific inquiry, and writing code.
“To create the open-access, equitable education system that a lot of us dream of, we need to make some roles teachers play easier — especially understanding how to help students work through open-ended tasks,” Piech says.
“For students without access to a teacher, this could open up a whole world of learning.”
At the heart of the team’s proposal is an AI-based engine that “understands students,” which could have a long tail of positive impacts. As Piech suggests, these include expanding the impact of teachers and even taking the role of a one-on-one tutor in some scenarios.
A separate project of Piech’s illustrates the proposed idea’s value.
“When the COVID-19 lockdown started, a colleague and I wanted to ‘open-source’ our Intro to Computer Science courses to make them available free to people worldwide,” he says. “So we planned to recruit 900 teachers to teach small groups of students and get high-quality education to 10,000 people. But we realized there was no solution to give teachers feedback on how their students were learning.”
In one version of the team’s proposed AI tool, a “Super Teaching Assistant,” could provide the volunteer teachers automated, detailed reports on what exactly students are struggling with at the collective and individual levels: “It might say, ‘Hey Teacher, these five students are struggling with this one idea.’ That will help the teachers understand their students deeply and deliver a better education, multiplying their impact.”
Indeed, such a system would enable teachers to spend less time on grading and more effort on teaching or creating exciting assignments to inspire students.
Another version could offer customized feedback directly to students to promote their learning with a fast-feedback loop. “If you can have humans involved in the way education is delivered, you should,” Piech says. “But there are many situations where that’s not an option,” such as in the Kenya-based community in which he grew up.
The proposed system would first focus on helping students learn the scientific method and coding. Both disciplines involve many open-ended tasks but also offer structured-learning opportunities, as compared with, say, poetry. The scientific method, for example, is about generating specific hypotheses and testing them with experiments using collected data. “There are great online tools that enable you to practice experimentation and see results,” Piech says. “But there’s no tool to look at your process of learning experimental methods, assessing your understanding, and giving you feedback.”
Coding involves similar learning opportunities and broad application. At this point, the system wouldn’t be able to help people research a novel or write poetry, Piech notes. “That may be possible in the future, but there’s plenty of middle ground to explore now.”
While some large leaps forward in technology might benefit those who already have the most access, as Piech suggests, his team envisions a much more inclusive process when building their system, one that includes explicit design for students from different backgrounds and geographies.
“The system should be adapted to diverse learning needs and contexts,” he says, “and can help train new teachers, multiplying its effects and lowering the barrier of creating scaled human-centered education.”
In this way, the tool could help create a more fair world where more learners have access to a high-quality, skills-focused education.
“Think about how many people want to learn STEM,” Piech says. “There’s a real hunger especially right now for very practical skills. It’s a shame when someone wants to put in the work to learn and make a contribution but no teacher can show up. Our system works toward addressing that gap.”