In 1966, Patrick Suppes—a leading philosopher, professor, and winner of the National Medal of Science—predicted that educational technologies would, “in a few years,” lead to millions of students having access to “the personal services of a tutor as well-informed and responsive as Aristotle.” In 2015, the CEO of Knewton, an adaptive learning platform, described his company’s product as “robo tutors in the sky.” While technology has made large impacts on education in the past 75 years, it's fair to say that it has not unlocked the world of easy access to high-quality, personalized tutoring that proponents have envisioned. Now, in 2023, we have entered the age of generative AI and large language models, such as ChatGPT. Is technology finally ready to deliver on the promise of computer-powered tutoring as a scalable technological solution?
To help address that question, let’s unpack why there is such a focus on tutoring as a use case. In terms of educational interventions that have been tested rigorously, in-person tutoring is one of the most effective practices there is, with meta-analysis showing robust effects. How do tutors produce such positive effects? Research has identified a number of elements, but, for simplicity, we can boil them down to:
- Social connection and motivational support.
- Cognitive scaffolding that unlocks a students’ own thinking.
Are those things generative AI can do well?
In terms of social connection, it is true that tools like ChatGPT can adopt particular tones and strategies that build up rapport, and they can respond in ways that are sensitive to frustration and other negative emotions, offering encouragement and normalizing struggles. For example, the image below is what was produced in a tutoring dialogue (on ChatGPT, Aug. 3, 2023, version), in response to a frustrated entry from the student.
So, it can potentially handle the moment-to-moment motivational support. However, it is not as appropriate for the longer-term social connection that comes from having a caring mentor who is invested in your success and who checks in with you—although new tools that emerge leveraging generative AI may soon function in that way.
Improvements to core functionality are going to be needed to help students build their own understanding. Soon after ChatGPT was unveiled, Khan Academy announced Khanmigo, a ChatGPT-powered tool that, among other things, can be used to tutor students. One of the key elements that had to be developed, according to Kristen DiCerbo, Khan’s chief learning officer, was the prompting happening on the back end. If you directly ask ChatGPT the kinds of questions a student might (e.g., “What is 10-(-9)2?”), it will tell you the answer. It likely even gives you the kind of step-by-step instructions you may see in a textbook. But it will not stop and ask you to generate each of those steps to understand where you may get stuck. It will not try to help figure out what you do and don’t already understand. Tools that build upon ChatGPT, like Khanmigo, can expand on the core capabilities of the generative AI—they generate effective prompts on the back end so that the experience is much more constructive for the learner.
To summarize, as the title of an excellent paper put it, “Learning is Not a Spectator Sport.” Having tools to tell you the right information or even show you how to do something won’t be sufficient to create learning. That can only be done when the learner works on things themselves and constructs their own internal understanding. And having social connections that help you, both by keeping you accountable and by rooting for you over the longer term, are critical. Perhaps a productive model for the future involves a human “meta-tutor” that helps guide a student’s interactions with their personal robot tutor, with the human helping build your personalized path and stopping to help you engage with a generative AI tutor when you get stuck.