ChatGPT and Other AI Tutors: Potential and Pitfalls

ai tutors
ChatGPT-type technology has launched a new era of AI tutors. (Image credit: Image by Donate PayPal Me from Pixabay )

Artificial intelligence has long been built into education apps but recent advances in large language models that power the popular ChatGPT have opened the door to a next-generation digital tutor that can interact with students and respond to their questions on an individual basis. 

Previous generations of AI technology might tailor preexisting questions and examples to a student’s perceived skill level in math, but rather than merely curating existing content, large language models can generate new answers and examples in response to queries from learners. 

“The thing about generative AI that is so different from past AI is it is designed to be creative,” says Gillian Hayes, dean of the graduate division at the University of California Irvine and co-head of the CERES (opens in new tab) edtech research initiative. “If you ask it a question, you're going to get one answer. If you asked the exact same question again, you're going to get a different answer.” 

This creativity is what makes AI tutors so spectacular and potentially beneficial to educators. ChatGPT-style tools can provide real-time, on-demand coaching for students, generating new differentiated examples with the click of a button. 

However, AI still generates inaccurate responses and we don’t quite know what the impact on anyone is, no less children, of interacting with something that seems so human but is certainly not human. 

Potential Benefits of AI Tutors 

Khan Academy has partnered with OpenAI, the developer behind ChatGPT, to utilize the most advanced version of its AI technology GPT-4 to power digital tutors on their respective platforms. Sal Khan told T&L (opens in new tab) that while he hadn’t been convinced previous versions of the technology were good enough to serve as a tutor, GPT-4 sold him. “It was able to do things that seemed like science fiction before that, like drive a nuanced conversation,” Khan said. “I actually think that 4, if it's prompted right, feels like it passes the Turing Test. It really feels like a caring human on the other side.”

Khan is using GPT-4 to power Khanmigo, an interactive learning assistant that is being piloted to select students and educators on the Khan Academy platform. 

The language learning app Duolingo now uses GPT-4 to power multiple features available to Duolingo Max (opens in new tab) subscribers. One GPT-4 feature will explain answers to a student generating unlimited explanations and examples on demand. Another allows learners to have a simulated conversation in the language they are learning – the digital tutor’s prompts are both spoken and written, and the learner can speak or type their responses.  

“Both these features are a great step toward our vision or dream of allowing Duolingo Max to be more like a human tutor in your pocket,” says Edwin Bodge, Senior Product Manager at Duolingo. 

GPT technology is also being used to power other learning apps and other similar AI models will likely be incorporated into more and more learning tools. Philip Oreopoulos, professor of economics at the University of Toronto, who has studied (opens in new tab) the impact of traditional human tutoring, is optimistic about the new era of AI tutors. Prior to the launch of Khanmigo, Oreopoulos had been interested in existing tools from Khan Academy to simulate human tutoring. “The only problem is that students just don't like to do it on their own,” he says. 

Over the last two years, Oreopoulos has been working with colleagues to explore how teachers might use Khan Academy more effectively. He believes Khanmigo could jumpstart these efforts. “You can have it beside you while you're going through a problem and it says all the things that you would want a good tutor to be saying,” he says. 

However, the technology is still not perfect. Khanmigo is currently text-based, which is not a natural way for younger students to communicate, Oreopoulos says. “The potential is there and it’s not a big stretch to kind of see where it could go,” he says. “These little details have to be worked out and that's where the research comes in. A lot of education technology is getting ahead of the research and so we don't have a good sense of which things work best.”

Potential Pitfalls of AI Tutors 

The unknown questions about the technology are the heart of concerns surrounding AI tutors. 

“Despite the many years of really interesting research and the sort of leaps and bounds that we've taken in generative AI recently, my biggest concern with any edtech is that I just don't think we have the level of evidence yet to really know how good these things are,” Hayes says. “We haven't done the kind of large-scale trials that you would see in the pharmaceutical industry to really understand how does AI tutoring do compared to human tutoring in terms of learning outcomes, student engagement, student well-being, and any of the things that you might care about and want to measure in kids.” 

Hayes is concerned about biases that are inherently built into complex AI models because these models reflect the biases in society. Though she believes generative AI tools will ultimately increase educational opportunities, there may be questions around equity in the short term. For instance, most access to GPT-4 at this time is subscription-based, which limits its audience. 

Another prime concern around AI tutors is the well-publicized tendency to say inaccurate or offensive things. This issue is why Khan Academy has rolled out Khanmigo in pilot form first. 

“We feel like there's so much value here for educators and for students, and we just don't want bad things to happen that turn people sour on all the positive things. So that's why we're being very careful,” Khan said. Though, he added, when these tools do go off the rails it’s usually because people are actively trying to steer them there. 

The Duolingo team is also cognizant of this and uses a second AI model to monitor the GPT-4 enabled conversations in order to ensure the conversation doesn’t go off-topic, which could make the AI tutor more likely to say weird or offensive things, Bodge says. 

Ultimately, Hayes says these tools don’t have to be perfect to be helpful and she hopes that AI tutors will increase educational opportunities for all. “Personally, I think the large-scale democratization of these technologies is likely to help us but I tend to be optimistic about these things, and that hasn't always been [warranted] historically, so we'll see,” she says. 

To share your feedback and ideas on this article, consider joining our Tech & Learning online community here (opens in new tab)

Erik Ofgang

Erik Ofgang is Tech & Learning's senior staff writer. A journalist, author (opens in new tab) and educator, his work has appeared in the Washington Post, The Atlantic, and Associated Press. He currently teaches at Western Connecticut State University’s MFA program. While a staff writer at Connecticut Magazine he won a Society of Professional Journalism Award for his education reporting. He is interested in how humans learn and how technology can make that more effective.