Rebecca Koenig Interim Senior Editorial Director | EdSurge Research
When students solve math problems, they are often required to show their work so teachers can identify errors in reasoning and ensure a proper understanding of mathematical concepts. New artificial intelligence (AI) projects are being developed to automate this process by training machines to detect and predict the mistakes students make while studying math. The goal is to provide teachers with real-time insights into student misconceptions.
Sarah Johnson, CEO at Teaching Lab, an organization that provides professional development for teachers, stated, "For the first time ever, developers can now build fascinating algorithms into products that will help teachers without requiring them to understand machine learning."
Some initiatives trace back to Eedi Labs, a U.K.-based edtech platform. Since 2020, Eedi Labs has organized coding competitions aimed at exploring AI’s potential in improving math performance. The most recent competition was conducted earlier this year and focused on using AI to identify misconceptions from multiple choice questions and accompanying student explanations. The event utilized data from Eedi Labs but was managed by The Learning Agency, a U.S.-based education consultancy firm. It was a joint project with Vanderbilt University and used Kaggle as its data science platform. Support came from the Gates Foundation and the Walton Family Foundation, with teams competing for $55,000 in awards.
Eedi Labs reported that the latest competition achieved “impressive” accuracy in predicting student misconceptions in math.
Researchers and edtech developers hope these advancements will lead to practical AI applications in math classrooms—a field where adoption has lagged compared to English instruction. There is debate about whether training algorithms on common student errors can result in effective tools for targeting instruction.
Despite significant investment in AI across industries—including education—some leaders remain cautious about inflated expectations around generative AI’s impact on teaching. Johnson explained that early attempts involved connecting educational platforms directly to large language models like ChatGPT but found such approaches lacking unless tailored specifically with educational data.
Eedi Labs currently offers an AI-powered tutoring service for math that uses a “human in the loop” model: human tutors review messages generated by the platform before sending them to students. Through efforts like its recent competition, Eedi aims to further train machines to detect student errors more efficiently.
However, developing machine learning algorithms capable of identifying common math misconceptions remains challenging.
Jim Malamut, a postdoctoral researcher at Stanford Graduate School of Education who is not affiliated with Eedi Labs or The Learning Agency’s competition, commented on the approach used: he noted that relying on multiple choice questions may be limiting when open-ended responses could provide deeper insights into student thinking—a strength of large language models.
Malamut also observed that while multiple choice formats have been popular due to scalability concerns, advances in AI could allow for broader use of more complex assessment types aligned with current research trends toward conceptual skills testing rather than rote knowledge checks.
Simon Woodhead, cofounder of Eedi Labs, emphasized their comprehensive approach: "Eedi Labs blends multiple choice questions, adaptive assessments and open responses for a comprehensive diagnosis." He added that incorporating student responses allowed deeper analysis during their latest competition but acknowledged trade-offs between assessment time and insight depth.
Woodhead also pointed out that broader definitions of “misconception” were used during the contest compared with typical company standards but said they were impressed by the accuracy achieved nonetheless.
There are differing views among researchers regarding whether such competitions truly capture underlying misunderstandings or if better-formed questions would yield greater insight into how students think about mathematics concepts.
Interest—and funding—in these efforts continues as policymakers look for ways technology might address declining U.S. scores on international assessments. Federal support under recent administrations has directed resources toward advancing AI strategies within education; Digital Promise recently allocated $26 million toward bridging best practices between education research and AI tool development.
A randomized controlled trial conducted by Eedi Labs alongside Google DeepMind showed positive results: integrating Eedi’s “human in the loop” tutoring boosted learning outcomes among 11- and 12-year-olds in the U.K., according to study findings cited by Eedi Labs. In the United States, their platform is currently used by nearly 5,000 students across dozens of K-12 schools and colleges; another randomized trial is planned for 2026 with Imagine Learning as a partner.
Teaching Lab is also piloting an AI model based on shared data from both Eedi and Anet companies; this model is currently being tested with students according to Johnson.
Meg Benner of The Learning Agency suggested possible classroom applications include providing teachers feedback about specific misconceptions or even triggering chatbot-led lessons designed around those gaps in understanding.
Johnson described it as "an interesting research project," noting future work must determine if improved diagnostic capability leads directly to better interventions for teachers and students alike.
Skeptics argue chatbots alone may not significantly enhance learning outcomes since previous studies have shown only top-performing students tend to benefit substantially from digital math programs; instead they stress ongoing teacher involvement remains crucial when addressing misconceptions revealed through computer analysis.
Still others see promise amid uncertainty: "I’m cautiously optimistic," said Malamut at Stanford. He noted formative assessments exist today but lack automation—if properly implemented using high-quality data sets aligned with classroom needs—AI tools could potentially close persistent gaps between assessment convenience and instructional value.
Alerts Sign-up