Math teacher Al Rabanera has a new tool in his decades-long quest to ensure his students, most of whom come from low-income families and marginalized backgrounds, see themselves in the curriculum.
Artificial Intelligence.
Many of of Rabanera’s students at La Vista High School in Fullerton, Calif., an alternative high school, struggled to learn in a traditional setting. But they care deeply about how their classwork connects to the working world.
So, when Rabanera taught rate of change—a statistical concept—he asked a large language model to create an assignment that would help his students better understand a highly relevant topic: the job market.
“Personalizing” lessons to students’ interests—especially in math, a subject that many middle and high schoolers find dry and irrelevant—has been held out as one of the most promising potential upsides of generative AI, the technology that powers large language models like ChatGPT and Gemini. Educators and researchers see a path to capitalize on the technology’s power and get at the elusive goal of customizing lessons for individual students interests, even if years of failed and overhyped attempts at innovation offer reasons for skepticism.
Rabanera’s approach was to put a prompt into an AI tool citing his students’ curiosity about the workforce. It responded with U.S. Department of Labor data showing correlations between education level, gender, and median weekly income.
Then it helped Rabanera brainstorm a list of questions to help his students dig into the numbers, using the statistical strategies they’re working to master.
For instance, students were shown the first and third quartiles of a linear graph and asked to estimate what the second and fourth quartiles might look like, justifying their responses.
When Rabanera asked the class to find themselves in the data, one female student quickly grasped what the income numbers met meant for her: “Whoa, Mr. Rab! I’m gonna get paid less ‘cause I’m a girl?” Rabanera recalls her saying.
AI has offered Rabanera—an experienced educator who has won multiple accolades for his teaching—a tool to “deepen and refine” the kind of lesson he always strives to deliver, he said.
“When I design these types of lessons without AI, it takes me hours,” Rabanera said. “I’ve done it before, looking at all the trends and themes, grouping and coding them, and then sharing it back with the kids.”
Even teachers who might not have Rabanera’s technological know-how may be able to harness AI tools to make math more relevant and exciting, some district officials say.
David Miyashiro, the superintendent of the Cajon Valley Union school district in southern California, envisions word problems that incorporate the names of an individual student’s friends, or a homework assignment that asks a baseball-fanatic to answer math questions about how far the ball traveled when San Diego Padres player Fernando Tatis, Jr. hit his latest home run.
“They’re learning the same math content, but it’s with vocabulary and names of people that they’re interested in,” Miyashiro said. AI and personalized learning is “going to really accelerate what you can do for student engagement.”
But the reality of creating these made-to-order assignments isn’t as easy as 1-2-3.
The process takes teachers time and know-how. And there are technical challenges to creating AI tools that can quickly deliver questions that incorporate student interests, but also assess math skills appropriately, and make realistic sense.
There are also big questions about whether the strategy actually improves student outcomes.
Khan Academy, an early leader in incorporating generative AI into student learning nixed a feature of its Khanmigo chatbot that sought to incorporate students’ personal interests into tutoring, in part because it wasn’t seeing clear benefits in either students’ academic progress or engagement.
Generative AI has also been touted as a way to customize math lessons for students’ academic abilities—not just their interests—potentially by creating new questions that target skills students need extra support to master. That prospect comes with its own set of technical hurdles and has been tough for companies that provide academic resources to crack.
Lackluster student engagement is a ‘significant’ challenge to teaching math
Students’ lack of interest in academics in general—but particularly in math—presents a hurdle for teachers, according to an EdWeek Research Center of 729 educators conducted from Jan. 28 to March 5.
More than half of teachers, 55%, cited poor student engagement in academics, including math, as a significant challenge. And more than a third, 36%, reported that students are less engaged in math than in other subjects.
Research has shown that linking math concepts to students’ interests—whether that’s TikTok influencers or Minecraft—can make an often-abstract subject come alive.
“A lot of kids will see math as not relevant, not connected to things that they do, and, as a result, not interesting,” said Candace Walkington, a professor in the teaching and learning department at Southern Methodist University in Dallas. “So we have to attack all three of those things at once by reimagining how we teach mathematics.”
A lot of kids will see math as not relevant, not connected to things that they do, and, as a result, not interesting.
Candace Walkington, professor, Southern Methodist University
The best personalized math problems look a lot like the ones teachers like Rabanera have crafted with an assist from AI: Deeply connected to what students know or care about but also representative of how math is used in the real world.
One problem is that the technology often struggles to come up with questions that feel authentic to the way math is used in the real world, said Walkington.
“AI is very good at knowing about interest areas. It knows all about Vampire Diaries or K-Pop Demon Hunters,” she said. “It’s great at that, but it’s not very good at connecting a topic to an academic area in a meaningful way.”
Trying to put a personalized spin on a math question through AI can result in a problem that doesn’t make any real-world sense.
For instance, an AI-generated question for students who spend their weekends going to see bands might ask about an event where the music level reached 400 decibels, a physical impossibility. Or it might suggest that nine people turned out to see music superstar Olivia Rodrigo perform.
Walkington, who received a grant from the National Science Foundation to explore personalization and math, has had some success in employing a specially designed “realism bot” to review potential math problems.
But she’s had a tougher time working through another challenge: Sometimes the math assignments AI conjures up may nominally incorporate a student’s interest but ask them to calculate something that would never come up in the real world.
For instance, AI may come up with a homework question that asks students to figure out how many “pins” people in the audience are wearing three minutes into a concert featuring music from K-Pop Demon Hunters, compared to five minutes in.
Pin-wearing at a concert, though, is “something nobody would keep track of for any reason,” Walkington said.
Trying to trick students into eating their broccoli
When Khan Academy’s Khanmigo chatbot tried to incorporate students’ personal passions in its tutoring, the tool ran into a different set of challenges.
Initially, the bot allowed students to input up to 10 interests that the tutoring tool could use to help the student wrap their mind around a new concept. If a knitting enthusiast had trouble understanding unknown numbers—which are foundational in algebra—the bot could describe the idea in terms of counting stitches.
The problem? It took Khanmigo a bit longer to come up with these more tailored responses.
That time lapse made a big difference to students, said Kristen DiCerbo, Khan Academy’s chief learning officer.
If the bot took more than 5 seconds to respond, students would just “drop off a cliff in terms of their usage,” DiCerbo said.
Though that may have been fixable with some technical tweaking, Khan Academy also found that spicing up responses with references to basketball or the Netflix series Stranger Things didn’t appear to improve kids’ grasp of material—or even their engagement with the tool, DiCerbo added.
With personalization, it’s possible that “the scenario that you make up to try to link [an academic concept] to an interest is just too contrived,” DiCerbo said. “Kids see through that, and they’re like, ‘Yeah, I think you’re trying to get me to eat my broccoli.’”
DiCerbo also wonders if asking students only about topics they are interested in—even as a spoonful of sugar to help the math medicine go down—is too limiting for young people.
“They’re still growing and they maybe don’t know all their interests,” DiCerbo said. “Education should be exposing them to new things that they might not know they’re interested in and that might spark their interest.”
AI may produce ‘cringy’ math problems
Getting past some of these obstacles may require technology beyond what’s currently widely available.
Walkington is testing a tool she’s developing with NSF support that allows middle school teachers to revamp the word problems included in their curriculum by incorporating topics, people, or even locations their students care about.
Teachers then review the personalized problems to ensure they make sense and address the right math skills.
Leslie Brown, a 7th -grade math teacher at Genoa Junior High in Texarkana, Ark., who is helping to test the tool, turned a problem about calculating distances from a cellphone tower into one about figuring out the best fishing spots in relation to a buoy. The new version capitalized on a popular pastime for her rural students.
The tool isn’t always perfect, Brown said. One confusingly crafted problem that asked students to calculate the circumference of a donut—and also for some reason referenced walking laps—was clearly “not gonna work at all,” she said.
And Brown’s students flagged a failure in a superhero problem that featured Marvel Comic book characters.
“These boys were like, ‘look, Thor’s hammer doesn’t do that. That’s not how it works, Ms. Brown, you need correct this,’” the teacher said. “It was just funny to me. I was like ‘thanks! I wouldn’t have known!”
Teachers don’t want AI to problems that are wildly off base
One way around that challenge: Allowing students themselves to pick the topics they most want to see show up in their math problems.
Burcu Arslan, a research scientist at ETS, a nonprofit organization, is working on a student-facing tool that not only incorporates students’ passions but allows for serious specificity.
The idea is to give students problems based on Freya Skye (a buzzy folk singer) as opposed to just “music.” Or about the Boston Bruins as opposed to just “sports” or even “hockey.”
Otherwise, if students “select music, and they get a problem about music, it can be still something that they are not interested in,” said Arslan, who is collaborating with Walkington on the student-facing tool.
After conversations with teachers, Arslan is also working to ensure that students are limited to safe, appropriate topics, as opposed to those focused on things like sex or illicit drugs.
Students choose: Witch’s Potion Lab or Friendly Pumpkin Patch
To be sure, it will take time for platforms like the ones that Walkington and Arslan are hoping to help develop to reach schools.
In the meantime, teachers are turning to AI platforms designed specifically for K-12 education—such as Brisk and Magic School AI— to add a personal twist or give students more choice in assignments.
For instance, Rebecca Sheeley, a math teacher at Weston Ranch High School in Stockton, Calif., used Magic School AI to create a Halloween-themed choose-your-own adventure style “magic math quest” to help students practice adding and subtracting polynomials, in preparation for an upcoming quiz.
Kids could choose among a friendly pumpkin patch (which Sheeley described as “light and fun”), a haunted mystery (billed as “creepy and spooky”) or a witch’s potion lab (“dark and enchanting”).
Here’s a look at what Sheeley’s students were presented with:
The seasonal twist helped hook students, Sheeley said.
“I could have just given it to them plain,” she said, “but then they just would have been like, ‘I’m done,’ and not really paid attention to it—not really [focusing] on what they needed to do.”
