Differentiation and Scaffolding in Math Teaching With AI
Excerpt from Integrating AI in Math Teaching: A Comprehensive Guide for K-12 Teachers
[Note: Tomorrow being July 4, there will be no AI in Edu weekly digest.]
In a 2023 message to fellow National Council of Teachers of Mathematics members, NCTM President Kevin Dykema proclaimed that the mathematics education community “must consider how to best integrate AI” into mathematics teaching and learning. As Dykema sees it, math educators “have a lot to learn” about how to best utilize artificial intelligence in the mathematics classroom. He writes that if math teachers are truly interested in meeting the needs of all students, they should not avoid or ban the use of AI.
Dykema’s message appears to have fallen on deaf ears. According to an EdWeek national survey, more than a third of math teachers don’t want any training on integrating AI into math teaching. And RAND reports that only 21% of math teachers have attempted integrating AI into math instruction. Put simply, most math teachers don’t think AI should be used to teach math to students.
Part of this resistance stems from lingering misconceptions about AI’s mathematical capabilities. Many teachers still assume that AI tools are poor at math, largely due to early experiences with models like GPT-3.5, which often made fundamental mistakes in arithmetic and struggled with multi-step reasoning.
But we’ve moved far, far beyond GPT-3.5.
Four months ago, Google’s AlphaGeometry 2 surpassed the average gold medalist’s performance on the International Mathematical Olympiad (IMO). According to Scientific American, this specialized AI system has solved 84% of geometry problems from the past 25 years of IMOs—some of which are so complex that even human experts can’t consistently solve them.
“Google’s AI Can Beat the Smartest High Schoolers in Math”
--Scientific American, February 2025
Advanced math models like Alpha Geometry may not be available to the public for a couple of years. But, as I’ve written, even publicly available tools like GPT-4o already demonstrate strong proficiency in key areas of math instruction—particularly Pre-Algebra, Algebra, and foundational Geometry, where procedural thinking and step-by-step problem solving are essential.
Another part of the resistance is that math teachers view AI as a “cheating machine”. The reality is that AI can ably support productive struggle—a core component of deep math learning. When used well, AI can help students wrestle with problems rather than shortcut their thinking. It can guide rather than give away, probe rather than tell. It can give students agency to explore, test, and reflect, which is vital for building conceptual understanding.
It’s time to move past outdated assumptions. AI can support strong math instruction—in most commonly taught math subjects. AI will make errors at times, but used skillfully, AI can augment thinking rather than replace it and augment teachers, not replace them.
This kind of AI-enhanced instruction doesn’t happen by accident. It requires clear vision, intentional design, and thoughtful implementation.
To that end, here is an excerpt from my upcoming guidebook, Integrating AI in Math Teaching: A Comprehensive Guide for K-12 Teachers.
Differentiation and Scaffolding with AI
Meeting the needs of struggling learners and those ready for more challenge is a constant challenge in math education. AI tools can be a powerful assistant by providing adaptive scaffolds and tailored content for varied levels, often in a time frame that a human teacher alone could never match.
Adaptive Practice and Tutoring: Khanmigo adjusts its support based on how students interact in real time. If a student is breezing through a problem, Khanmigo might challenge them with a follow-up question, ask for a deeper explanation, or offer a more complex version of the problem. If a student is struggling, it can provide hints, offer step-by-step scaffolding, or suggest simpler sub-problems to build confidence. While not an adaptive learning system in the traditional sense, Khanmigo mimics the responsiveness of a human tutor—tailoring the conversation to meet students where they are.
These systems essentially undertake one-on-one differentiation continuously. In a typical class, you might use such a tool for independent practice: each student works on a device with the AI tutor while you monitor and spend your time assisting those who need human intervention. The AI handles the fundamental help queries (“I’m stuck on step 2”), so you can address the critical misconceptions or provide enrichment.
Research from Stanford has shown AI tutors can be a “tremendous thought partner” for teachers in creating tiered lessons that reach diverse skill levels. That same partnership extends to students using the tutor. It’s as if each student gets a personalized sequence of tasks. Importantly, the AI tutor encourages students through the process. Khanmigo’s standout feature is guiding sequential steps and prompting reflection rather than giving the solution. This helps nurture metacognition and perseverance, two key goals in differentiation. In other words, we want struggling students to adopt strategies, and advanced students to push further.
Tiered Assignments: AI can help teachers create different task versions for varying readiness levels. For example, you might ask an AI: “Generate three versions of a geometry problem about finding the area of a trapezoid: one basic with numbers given, one intermediate requiring converting units, and one advanced integrating algebra (solving for a missing dimension).” Within seconds, you have a scaffolded set of problems.
Similarly, suppose you have students who need review vs. those who need extension. In this case, AI can produce a remediation task focusing on prerequisite skills for some and an enrichment project for others. Adopted this way, AI can drastically reduce the time to prepare multiple versions. Before the advent of AI, teachers might not have differentiated as much simply due to their workload. Now, AI removes that barrier by doing most of the busy work.
Targeted Hints and Explanations: Hints in math textbooks and software programs are one-size-fits-all. With AI, students can get immediate hints at their level. Suppose a student is working through a geometry proof and is unsure how to start. They could ask the AI a focused question like, “What’s the first step to prove these triangles are congruent?” The AI might respond with a hint like, “Consider which congruence postulate might apply. Have you identified three pairs of matching parts?” In other words, the AI provides a nudge, but not the complete answer.
Training students to ask for hints (versus answers) is critical to the success of this strategy. Some AI platforms have a built-in hint button. In the absence of that, you can role-play how to query AI for help: e.g., “I don’t need the solution, just give me a hint regarding property X.” This builds a scaffolded support structure that students control. Each student can get help when needed, and as needed, which differs greatly from a traditional one-size-fits-all instructional strategy.
Challenge Extensions: For those students who finish early or need more challenge, AI can generate extensions: “Great, you solved that standard problem. Ask the AI to give you a harder one or to twist it.” For instance, a student mastered solving a simple linear equation> Here’s a way to challenge them: “Have Gemini give you a word problem that leads to a linear equation to solve.” Or for geometry: “Ask the AI to pose a creative real-world geometry problem involving the concept we learned.” This keeps advanced students engaged without you having to prepare a whole extra set of materials. They can basically request their own challenge from the AI, albeit with your guidance on the type of challenge. In this way, AI differentiates and nurtures independence and exploration for advanced learners.
Differentiating by Learning Modality: Some students absorb better by reading, others by listening or doing. AI can help differentiate in modality. For example, a student who struggles with reading word problems might use a text-to-speech AI (or even the read-aloud feature of an AI tutor) to read the problem to them. Also, tools like Wolfram Alpha provide visual graphs that can turn an abstract function into a concrete picture. Let’s say you have a student who just isn’t getting the concept of solving equations abstractly; you might use an AI to show the balance scale image or the line-by-line algebra, and simultaneously have them do it with algebra tiles physically. You can balance an AI contribution with a physical experience, another way to aid differentiation.
Best Use Tips: Differentiation is most effective when done proactively (planning different paths) and reactively (adjusting as you see student responses). Fortunately, AI supports both, since you can can plan tiers ahead of time, but also respond in the moment.
It’s important to keep tasks equitable in interest. Avoid consistently giving some students “drill” and others “fun projects.” Instead, try to make all tiers engaging, just pitched at different levels. AI can help by adding context to simple problems to keep them interesting or by simplifying the numbers in a complex project to make it accessible.
Also, train students that different work doesn’t mean “smart vs dumb” – rather, everyone is getting what they need to progress. Perhaps use neutral language like “Level 1, 2, 3 problems” or let students choose their challenge level at times (with guidance). AI could even generate a leveled set and you allow students to pick one level to attempt, moving up if successful.
Any suggestions to improve this newsletter? Please message me or leave a comment below!