AI Goes to School: Exploring AI’s Impact on Personalized Learning

I’ve recently read a few articles listing the pros, cons, and questions^{1 2 3} that come up when we think about the changes education will and is already experiencing with the introduction of AI into learning environments. It seems this topic of personalized learning powered by AI warrants a bit of digging from a fellow concerned adult who is curious about the changing landscape of educational technologies.

Going into this work, I wondered: What does personalized learning infused with AI capabilities actually look like, and is it as revolutionary as it’s been presented? What can parents, educators, and adults who care about young people’s future in an AI-infused world listen to and advocate for when we see these innovations integrated into schools and children’s devices? To gain more insight, we spoke with representatives from three companies: Dr. Klinton Bicknell, who leads Duolingo’s AI group; Dr. Kristen Dicerbo, Chief Learning Officer at Khan Academy; Dr. Kristen Huff, Head of Measurement at Curriculum Associates and Dr. Amelia Kelly, CTO at SoapBox Labs, now part of Curriculum Associates. These experts and their respective companies are hard at work thinking about how they will integrate AI into learning environments and products for use in the classroom or other settings. Here’s what we learned from these conversations, and what we’ll be keeping in mind going forward:

Let’s start with some mythbusting: this is not the first time AI has been used in learning products. Curriculum Associates, for example, has already been “using the power of data and algorithms to adapt the assessment to each student and to understand where each student’s strengths are in math and reading, and where there are still gaps to close to get to grade level”, Dr. Huff explains. i-Ready, Curriculum Associate’s main product, is a digital curriculum and assessment for reading and math programs aimed at students K-8. The integration of SoapBox Labs’ technology will allow the product to use voice AI to help young kids learn to read in noisy environments like classrooms and improve reading comprehension.

The real innovation of the moment is AI’s ability to use natural language to engage with learners about content in new ways. For example, an AI-driven learning system might be able to better understand a student’s capabilities and deliver interventions based on their needs. That may mean making connections between what a system knows about a student’s capabilities and the interventions digital tutors can make to further their learning based on that information.

Duolingo, the popular language learning application, has also been integrating AI into many of their products. Dr. Bicknell explains:

“By taking the millions of points of data we have about each learner that uses the app, we create a detailed signal about what this person knows right now, what kinds of mistakes they’re making, and what they know well and over time. Then this gives us the whole picture of how people are learning across all our different courses…. AI then gives us a way of taking all that data and transforming it into ways of personalizing and optimizing the learning experience for everyone.”

More specifically, generative AI has been especially helpful for Duolingo as they are developing more nuanced features for conversation practice – a skill that’s necessary to learn how to speak a new language. Dr. Bicknell further explained:

“Historically, it was a big limitation of any kind of electronic learning tools that they weren’t very interactive. There are basically [only] right answers and wrong answers. And you can go to the next question. Without AI, you’re really limited in the kinds of ways that you can create things or give people feedback about things because it really has to be stuff that you hard code essentially. Duolingo is trying to build really high quality education and scale it, making it available to everyone. AI is really the only way of achieving both of those things at once.”

What are the potential benefits?

Across the three interviews, we heard several clear benefits of integrating AI capabilities into learning products and platforms.

First, companies believe that by integrating generative AI capabilities, we may be able to provide student assistance not available from already over-stretched teachers in the classroom. When using Khanmigo in a classroom setting, for example, Dr. Dicerbo explains that one main benefit is “the ability to provide the help to a student when they’re working on independent work, [since] a teacher in a classroom just can’t help 25 kids at the same time as they’re stuck. The teacher doesn’t have to keep helping the kids that are doing independent practice now because Khanmigo can.”

Similarly, when envisioning the use of iReady’s speech recognition capabilities in the classroom, Dr. Kelly describes how AI can be a useful tracking tool for a teacher and act as a data source to accelerate children’s learning: “Crucially, the teacher will now have time to do this, because they are not engaged in listening to hours and hours of children’s reading, and having to grade that reading by hand.”

Dr. Bicknell reflected on data showing many language instructors use Duolingo with their students to fill in downtime between in-class assignments, and found its personalized speed and manner helpful for managing different skill levels and learning speeds in one classroom, saying that teachers may choose to use Duolingo in different ways.

“Sometimes it’s: ‘If you finish the assignment early, do some exercises [on Duolingo]. Duolingo is like this personalized learning app and the assignment you just did is probably not personalized. And so this allows the teacher to go help the students who are struggling, and the students who are just doing great can go do more advanced stuff on their own, which is cool.”

Using personalized learning environments might also impact students’ engagement with the material alongside giving teachers signals about a student’s progress. Dr. Huff at Curriculum Associates explained that “Students are motivated to read aloud and be scored with AI-powered voice technology because of the immediate feedback loops for them as students as well as for their teachers.”

Using AI to create different learning contexts according to a learner’s unique interests also increases their engagement, especially for Duolingo, according to Dr. Bicknell:

“[We] found this very effective both in increasing the amount people learn and also in helping increase engagement as well. If you can make things really relevant for people, then they will both learn more and be more into it. Otherwise, [we risk showing] people things that they already know, and they’ll get bored and leave. Or if you show people things that are too hard for them, they’ll get frustrated and leave. So it’s a nice sweet spot that we’ve [achieved by using AI]”.

Perhaps one of the more interesting benefits AI can bring when it comes to personalized learning, especially for adults, is the very fact that it is not human – and thereby poses less of a risk of the learner feeling judged or hurting the learner’s self-esteem and confidence. For example, Dr. Bicknell reflects that Duolingo has “experimented with [practicing conversations] through human tutors who would connect with [learners] – but that never got [to] big scale. What we found is that a lot of people have a really hard time talking to a human in a language that they are not confident in. They’re very embarrassed and shy, and feel like this other person is judging them right when they make mistakes. And so one really cool thing about AI is that you don’t have that. It’s just a computer, it’s not judging you.”

Finally, many of the experts we talked to reflected that  integrating AI capabilities into personalized learning can have a positive effect on the type and timing of feedback, both important aspects that can affect learning trajectories and milestones throughout the learning process. With the integration of AI capabilities in Khanmigo, for example, Dr. Dicerbo spoke of “being able to get more of that immediate help, the scaffolding and the quick feedback that we see as kind of the biggest thing that we think is going to improve learning.” Inside Duolingo, using AI unlocked the app’s ability to “give people feedback on their answers. Obviously, we can’t exactly say whether this is right or wrong, but we can at least say something like, ‘Oh, here’s how to fix your grammar, or here’s how to say that better’.”

What are some of the challenges when integrating AI into learning environments?

Of course, it’s not entirely smooth sailing. As the experts described the benefits of generative AI, they voiced concerns about its impacts as well. Integrating AI into learning products is expensive; the high cost of this technology means it is not readily accessible to all children and learners. Dr. Bicknell reflected:

“It’s a feature we think is really helpful for learning, and we wish we could give to everyone, but for now we can’t [because of the high cost].That is a tension of using AI for education; there can be some inequity, considering who has access and who doesn’t. Now, all that said, prices are falling very fast, and so we are hopeful that it won’t be very long before more and more people can get access to this.”

Others raised concerns about the way this technology can be used in schools, looking at both the accuracy and how they provide guidance to students. Modeling educators, Dr. Dicerbo reflected that they don’t want their product Khanmigo to be “a tool that just answers questions. We want a tool that’s going to help students get to the answer themselves. And so we started learning how to prompt engineer to basically give instructions to the model of how to act like a tutor.” Moreover, sometimes an AI agent might provide wrong answers (though that can also happen with human educators). Students and teachers using the tools need to be aware of this potential.

The mistakes that AI tools make often show up in math curricula, for example. Dr. Dicerbo reflected that “It’s known that math accuracy is an issue. But it’s not just doing the calculation –  actually one of the bigger issues is evaluating the students’ responses and saying they’re right when they’re right or wrong when they’re wrong. And we found that the models have a positivity bias. So we’ve done a lot of work on how to make it more accurate on the evaluation side, and we also released an evaluation data set so that anyone who’s building a model can test it on how good it is at evaluating student work”. 

What should companies keep in mind as they develop AI-powered learning experiences?

An emphasis on maintaining human connection
The experts we spoke with emphasized the importance of creating a tool, rather than a replacement for the teacher – as “AI will keep getting better at being able to explain things and being able to personalize instruction…But it’s quite far from replacing that human connection, the critical part of what [a] teacher’s job actually is,” says Dr. Bicknell.

With that being said, products that guide students like Khanmigo are built in a way that emphasizes and prioritizes human connection. Dr. Dicerbo explains: “There are so many things that teachers do that AI is never going to do. We have a phrase that we use: human high fives beat computer confetti… it’s always going to be more motivating for someone to be like, ‘Hey, awesome job,’ as opposed to like Khanmigo showing virtual confetti.”

Echoing that sentiment, Dr. Kelly said: “We believe that AI, or really any computer assisted technology, is just a tool for the teacher. Teacher judgment, teacher expertise must be centered in the learning experience if it’s to be most effective for the student.”

Focusing on education-specific learning tools
One key point to keep in mind as we think about AI-powered tools aiding learning is the original purpose, and whether they were designed for learning in the first place. Dr. Dicerbo reminded us that something that might not be clear to educators and parents when thinking about personalized learning “is the difference between education specific [products] versus ChatGPT, for example. If you ask ChatGPT a question, it’s just going to give you the answer, leaving the students to their own devices. In contrast to not just Khan Academy, but any of these education apps which are designed to encourage students to work in them.”

The emphasis on learning mechanisms and user fit is reminiscent of age-appropriate design initiatives that have been grown in popularity in recent years. In essence, as Dr. Kelly explains:

“Taking ChatGPT and other large language models (LLMs) off the shelf and sticking a wrapper on it and giving it to children won’t work. They’re not looking at how a voice tutor, for example, would respond to children from different regions, or if it’s the child’s first language or second language. There’s no consideration put into how it’s built and what it can do. It’s not being deployed using data and using evidence-driven research.”

She went on to elaborate:

“It’s for that reason that I am very concerned about the types of technology that may make it into the school system; because at the end of the day, AI-powered technology is coming to the classroom. It’s coming everywhere, and we can either let it in without testing it, or we can test it and make sure it works. Make sure that it’s built with privacy by design. Make sure it’s built with transparency. Make sure it’s built with equity at the forefront. and if we can be totally upfront with that, show people how things are built, we can make a real difference.”

After having these conversations, I have become cautiously optimistic about the future of AI-powered personalized learning being integrated in classrooms and learning experiences in the coming years. Yes, there is still a lot to learn, and yes, we should keep questioning and reassessing the decisions we make with this technology and how it affects young people’s chances of success, equity, and teachers’ autonomy in the classroom. Nonetheless, thoughtful design decisions and ongoing conversations like the ones recorded above mean that we may yet avoid delving into a world where we use AI for AI’s sake, and consistently reiterate towards the goal of educating the next generation with empathy and attention to their uniqueness and potential.

 

¹https://www.commonsensemedia.org/sites/default/files/research/report/2024-the-dawn-of-the-ai-era_final-release-for-web.pdf

²https://www.the74million.org/article/is-ai-in-schools-disruptive-or-overhyped-potentially-both-new-reports-suggest/

³https://boldscience.org/how-can-educators-use-ai-to-support-their-students-learning

⁴https://www.the74million.org/article/ai-fueled-testing-from-the-mouths-of-babes/

⁵ https://childrightsbydesign.5rightsfoundation.com/

 

 

Rotem Landesman is a PhD student in the Information School at the University of Washington, co-advised by Dr. Katie Davis and Dr. Amy J Ko. Her research looks at developing metrics for well-being design in digital spaces, alongside fostering youth’s technological wisdom. She is currently a graduate intern at the Joan Ganz Cooney Center.