Can AI Finally Personalize Math Learning?
Teachers have long wanted to connect math concepts to what students actually care about. AI math personalization promises to make that easier, but early results suggest the technology is far from delivering on the hype. A survey of 729 educators by the EdWeek Research Center found that 55% consider weak student engagement a significant obstacle, and more than a third reported that math draws even less interest than other subjects.
Khan Academy Pulled the Plug on Interest-Based Tutoring
Khan Academy tried letting students feed up to 10 personal interests into its Khanmigo chatbot so the tool could frame concepts around hobbies like knitting or basketball. The feature was quietly dropped after the organization saw no meaningful improvement in learning outcomes or student engagement in math.
Response time played a role. Kristen DiCerbo, Khan Academy's chief learning officer, noted that students abandoned the tool when replies took longer than five seconds. But latency was only part of the problem. Sprinkling pop-culture references into tutoring sessions did not help students grasp material any better.
DiCerbo also raised a philosophical concern: students are still developing, and restricting content to known interests could actually narrow their educational exposure rather than broaden it.
Generative AI Still Struggles With Realism
Even when generative AI in education tools successfully weave a student's hobby into a word problem, the results often lack real-world logic. Candace Walkington, a professor at Southern Methodist University studying personalized math learning with support from the National Science Foundation, offered telling examples. One AI-generated question described a concert reaching 400 decibels, a physical impossibility. Another placed just nine audience members at a major pop star's show.
Walkington has built a "realism bot" designed to screen out nonsensical outputs, but a subtler issue persists. The technology frequently asks students to calculate things nobody would ever measure in practice, such as tracking how many pins concertgoers wear at different points during a show.
Leslie Brown, a 7th-grade math teacher in Texarkana, Ark., who is testing one of Walkington's tools, confirmed this gap. One problem tried to combine donut circumference with walking laps in a way that made no sense. Her students also caught a superhero-themed question that misrepresented how a well-known comic book weapon works.
Where AI Is Already Helping Teachers
Despite these limitations, some educators are finding practical value. Al Rabanera, a math teacher at La Vista High School in Fullerton, Calif., used an AI tool to build a rate-of-change lesson around U.S. Department of Labor wage data. His students, many from low-income backgrounds, explored how education level and gender correlate with weekly income. One student immediately spotted the gender pay gap in the numbers.
Rabanera emphasized that he could have built the lesson without AI, but the technology compressed hours of data research and question design into a much shorter workflow. Other teachers are using K-12 focused platforms like Magic School AI and Brisk to add thematic variety to assignments, from Halloween-themed polynomial practice to choose-your-own-adventure math quests.
Researchers like Burcu Arslan at ETS are also developing student-facing tools that allow learners to select highly specific interests, such as a particular musician or sports team, rather than broad categories. The goal is to produce problems that feel genuinely relevant rather than superficially decorated.
For now, the most effective uses of AI in the classroom appear to rely on experienced teachers reviewing and refining what the technology produces. Fully automated personalization at scale remains an unsolved challenge, and the gap between the promise and current capabilities is wide. Read the full report from Education Week.