Few questions divide educators and families right now like the one about screens.
Pediatric guidance has moved away from simple screen time limits and now places more emphasis on what children are doing on a device, why they are using it, and the human interaction that surrounds the experience. This reflects something early childhood development researchers have long understood: language and reasoning grow primarily through rich human interaction, and time on a screen too often displaces face-to-face talk and play that build these skills.
At the same time, the educational technology landscape is crowded with tools that promised transformation and delivered distraction, leaving many teachers and administrators rightly skeptical of whatever arrives next.
Artificial intelligence has poured accelerant on all of it. The debate that once centered on screen time and app efficacy now runs straight into harder questions about data, bias, and what it means for a child to interact with a system that always has an answer.
Even the U.S. Department of Education, in its guidance on AI in education, lands on a deceptively simple principle: keep humans in the loop, demand transparency, and treat equity as a design requirement rather than an afterthought. That framing reorients the whole conversation. The question is not whether to use AI in elementary math. It is whether the humans in the loop (the teachers) end up empowered or sidelined.
At PowerMyLearning, we designed MathVoice to use AI in service of teachers, not in place of their judgment.
Discourse is data
We start with a belief that sounds modest but is truly revolutionary: what students say to each other is as important as what they write down.
Most math tools are built around right and wrong answers. A correct answer tells you a student reached the destination, but not how they got there, what strategy they used, or whether they could apply the same thinking in a new context. When a child explains their reasoning out loud, in their own words, understanding becomes visible.
Even in an era of edtech integration, many products are still essentially digital worksheets. Whether on paper or on a screen, a static artifact of student work limits what teachers can learn. A correct answer may signal understanding, or it may reflect procedural mimicry. An incorrect answer may reveal a misconception, or it may mask sound reasoning interrupted by a small error. Student discourse gives teachers another layer of evidence.
A student who explains why she combined two groups, or who confuses a nickel and a dime while otherwise reasoning soundly, is giving a teacher far more to act on than a right-or-wrong answer ever could. This is the same conviction that runs through our Foundations of Numeracy framework: proficiency is not a single skill, but a braid of competencies, content, ways of thinking, and the motivation to keep going. Much of that braid reveals itself through talk.
Student discourse gives teachers a window into that deeper understanding. We have to create more opportunities for them to explain their reasoning.
What MathVoice Is—and Isn't
In MathVoice, technology sets up the experience, but it does not drive the learning.
Students access a short, curriculum-aligned math activity, then move into offline play with a partner or small group. They might partition shapes, compare quantities, or build equivalent fractions before returning to the device to record a brief video explaining what they figured out.
That sequence is intentional, and the collaborative offline work is key. Children are not interacting with AI or a screen during the activity. They are talking to each other: negotiating who has more, disagreeing about how to partition a shape, trying out a justification and revising it when a partner pushes back.
That peer interaction is where young learners rehearse mathematical language before they are asked to produce it formally. It is exactly the kind of human, conversational learning that pediatric and developmental guidance says screens should support rather than replace.
Rather than your typical warm-up that precedes the learning, the game is the learning that lays the foundation for everything that comes next. You cannot generate honest insight about a child's mathematical thinking if no thinking was ever made audible.
Only after the talk happens does technology enter, and even then it never touches the student. The recorded explanation is transcribed, the transcription is analyzed for proficiency on the lesson's learning goals and underlying skills and for the depth of reasoning behind it, and the results surface in a teacher dashboard that groups students by what they demonstrated and suggests next steps linked to freely available curriculum lessons. For edtech products used by young children, keeping the AI teacher-facing is a core design responsibility.
Four design commitments for responsible AI
The way MathVoice captures student thinking and turns it into teacher insight is guided by four design commitments that shape how we use AI with young learners.
- Surface student understanding, not just a score.
Most math tools simply provide a score or proficiency label, a small window into student understanding. MathVoice is designed to analyze students’ explanations and strategies in their own words, helping teachers identify common misconceptions, patterns, and opportunities for support. - Strengthen teacher judgment, not replace it.
AI should empower teachers as a tool to inform instruction, not prescribe it. MathVoice surfaces patterns in student thinking and suggests possible next steps, but nothing is auto-assigned. Every recommendation is a starting point for the teacher’s professional judgment, weighed against everything that teacher already knows about the child. The tool surfaces time-saving insights; the teacher decides what to do next. - Work in the background, not directly with students.
For young learners, responsible AI protects both the learning experience and student data. MathVoice is teacher-facing by design: students never chat with or receive responses from AI. Instead, students collaborate with their peers, practice math using hands-on activities, and strengthen math discourse. The AI works behind the scenes to analyze transcriptions and provide teachers with time-saving insights. Student data is encrypted and information is never sold or shared beyond its educational use. - Build for real student voices, not ideal conditions.
The way young learners learn and explain math is nuanced. They use developing speech, informal language, and varied ways of describing what they understand. Most speech recognition is trained on adult voices in quiet rooms, and research shows that these systems often produce higher error rates for children, multilingual learners, and speakers of non-dominant dialects. That matters because a system that mishears a six-year-old who says “free” for “three” does not just produce a typo; it risks misjudging the mathematics underneath. MathVoice is purpose-built for young learners’ math discourse, not wrapped around a general chatbot or one-size-fits-all AI model. Designing for the way children actually speak, reason, and explain math is both a technical decision and an equity decision.
Why design choices matter for equity
It would be easy to treat equity as one principle among several. It is better understood as the reason the others matter.
Consider language first. In MathVoice, a multilingual student once explained, "the sticks on the clock point to the one and the three, so it shows 1:15." She never reached for the academic word hands, but her description is accurate and her mathematical reasoning is sound, and the analysis recognized it as proficient.
Separating a student's grasp of the mathematics from her command of academic English is something most assessments quietly fail to do, penalizing children for still-developing vocabulary when the reasoning is already there. Research on translanguaging in content assessment makes the same case: letting students draw on their full linguistic repertoire reduces irrelevant language barriers so that assessment captures what they actually understand. Allowing a child to explain her thinking in her own voice encourages discourse and positive math mindsets.
Then there is the longer history. Math instruction in the United States was built more than a century ago in part to sort students into those considered "math people" and those who weren't, and it has worked well for relatively few of them. The results still show it: only about a fifth of twelfth graders in the country are proficient in mathematics.
That history shapes achievement data and also how students see themselves. Student submissions can reveal not only what children understand, but whether they are beginning to see themselves as doers of mathematics.
As one MathVoice teacher in Houston said, “I’ve just seen her blossom into a real mathematician. I have not seen them blossom ever before like I have this semester using MathVoice.”
Teachers inherit this system. Because teaching is among the few professions we watched performed thousands of times before we ever did it ourselves, we tend to replicate the practices we absorbed, including the inequitable ones. Equitable instruction pushes the other way: it gives students multiple access points to both concepts and procedures, often through the Concrete Representational Abstract (CRA) progression that grounds abstract symbols in things a child can see and hold. The offline games are built in that spirit, and the analysis layer is designed to honor it.
This is where AI earns its place. It should not decide what a student does next or interact with the student directly. Instead, it should give teachers a clearer, fairer, faster picture of what just happened so their expertise can do more with it.
A partner teacher in Santa Clara, CA said, “I think the most valuable thing is that these videos let me feel like I’m getting a 1:1 conferring session with the kids, when I’m really not, because there’s no time. So it’s nice to be able to understand how they understand.”
Pacing, grouping, intervention, language support, the quality of math talk: these are the questions skilled teachers are already asking after any formative assessment. Instead of inventing a new workflow, the tool honors teachers’ judgement and gives them back one of the most valuable commodities: time.
At PowerMyLearning, research-backed pedagogy and equitable outcomes are at the heart of everything we do. AI is simply a tool to help teachers do what they do best and extend those outcomes to more students. In an elementary math classroom, its greatest potential is helping teachers make their next instructional move with greater confidence and clarity.
Continue the Conversation
How should AI support great teaching? We'd love to hear your perspective.
