Should maths classrooms change in the AI era?
AI could help teachers take a more discussion-based approach
Author
In many classrooms, maths instruction looks the same as ever. The teacher explains, students practice, some keep up, but many do not.
Nearly 75% of 15‑year‑olds in Latin America lack basic proficiency in maths, finding it challenging to perform everyday tasks like reading a simple graph or calculating a percentage discount in a shop.
“Nearly 75% of 15‑year‑olds in Latin America lack basic proficiency in maths.”
Jose Rafael Espinosa, Co-Founder and CEO of MentuLabs, an education organisation working in Colombia and the Dominican Republic, has been tackling this challenge for years. According to Espinosa, improving learning outcomes at scale in Latin America requires rethinking what happens inside classrooms.
Students learn more when maths lessons focus on reasoning, discussion, and conceptual understanding and not just on speed and repetition. Evidence-based approaches such as Stanford University’s YouCubed initiative promote a ‘maths mindset’: the idea that all students can learn maths when classrooms value multiple solution strategies, and mistakes are not seen as a source of anxiety.
As Espinosa puts it, YouCubed “assumes that every student can be good at math. What you need to create is a positive environment for that”.
The challenge is implementing this approach in classrooms. Planning discussion-based lessons takes time, and facilitating them requires confidence. In many schools, teachers can struggle with large classes, limited preparation time, and uneven internet connectivity. In some classrooms, the only reliable device is a teacher’s smartphone.
What maths classrooms could look like with AI
Mentu developed an AI-supported tool called Shaia, based on the YouCubed approach, to assist teachers with planning and reflection. Instead of assigning a long list of exercises for students, teachers are encouraged to run ‘maths talks’, which Espinosa defines as “structured conversations around a specific math problem”.
“Students see that there are different ways of doing math and that their own way was actually good or interesting.”
The teacher presents a problem, such as 18 × 5, and invites students to think of different ways to solve it. Some break the numbers apart, while others double 5 and halve 18 to get to the answer 90. The teacher collects the strategies and discusses them with the class. “Students see that there are different ways of doing math and that their own way was actually good or interesting. They discuss it and it creates a very positive environment”, Espinosa explains.
The goal is not speed or finding a single correct method, but making mathematical thinking visible and inclusive.
Shaia helps teachers plan these discussions, anticipate student responses, and reflect afterwards on what worked and what did not. However, Espinosa is clear about the limits of AI in this context, pointing out that “knowing your students shouldn’t be delegated to the AI”. While AI can help organise, suggest and structure, teachers remain responsible for pedagogical decisions around how to support students in their learning journey.
Early evidence from classrooms
To understand whether this approach leads to measurable gains, Mentu is running a randomised controlled trial with 58 public secondary schools in Bogotá, Colombia. Teachers in the intervention receive access to Shaia alongside professional development focused on maths pedagogy. Those in the control group continue teaching maths as they always have.
Preliminary results halfway through the trial are encouraging. According to Espinosa, there appears to be a moderate to large positive impact on learning outcomes in the intervention group, “depending on the different methodologies that researchers are using to measure the impact”. There are also early indications that Mentu may reduce the gap between boys and girls in maths.
The study is ongoing, with final results expected later this year. These early findings should therefore be interpreted cautiously.
Implementing learning science with AI
I take several broad lessons from this early evidence. First, AI may be able to reduce the practical barriers to implementing evidence-based practices. The tool is designed to lower planning effort while supporting teachers in facilitating structured classroom discussions around maths concepts.
Second, technology alone is unlikely to be sufficient for improving maths at scale. In the trial, access to the tool was combined with professional development that emphasised the underlying pedagogy. This combination may be critical for sustained change.
Third, early data from Mentu suggest that even when teachers do not use the AI tool constantly, occasional use, when paired with training, may lead to lasting changes in classroom practice. It seems that digital tools don’t need to be used daily to have impact.
“AI may be most effective not when it attempts to automate teaching… but when it helps teachers apply learning science.”
Espinosa is excited about AI as a tool for helping teachers integrate best practices into their lessons. AI may be most effective not when it attempts to automate teaching – for example providing a long list of practice questions – but when it helps teachers apply learning science.
Author