Open navigation menu. Close navigation menu.

When AI is the Messenger: How AI-Enabled Tools Can Increase Motivation, Engagement, and Persistence for Students Learning Math

Insights about AI-enabled tools for students from Math Narrative Project field tests 

This article features results of AI-focused field tests supported by the Math Narrative Project and the Gates Foundation. These field tests were led by: EdLight, Goblins, OKO, and The Emancipation Group. 

In this article from the Math Narrative Project, you will learn: 

  • How organizations used messaging recommendations from the Math Narrative Project to train AI-enabled tools to effectively reach students learning math; and
  • Several lessons that organizations learned about how AI-enabled tools can increase students’ motivation, engagement, and persistence. 

Learning and teaching math is a deeply emotional experience for students, parents, and teachers alike. Far too often, students encounter pervasive, deficit-oriented narratives about learning math that can lead to negative emotions like stress or frustration, and lead to disengagement. Emotions are part of the human experience. The good news is that emerging evidence from the Math Narrative Project is showing that it is possible for messages delivered by humans and AI-enabled tools, such as AI agents and chatbots, to change beliefs and behaviors in a way that supports students to engage in cognitively demanding work, to persist through productive struggle, and to succeed in learning.

The Math Narrative Project, supported by the Gates Foundation, is advancing evidence-based messaging and narrative change strategies to improve math instruction and outcomes for 6th to 10th grade Black and Hispanic students of all incomes as well as all students from lower-income households. The Math Narrative Project uses narrative as a tool to improve students’ math learning so that more students feel motivated and supported to learn more math. 

To test the effectiveness of evidence-based messaging recommendations for students, the Math Narrative Project initiated a series of field tests; a subset of the field tests focused on understanding the effectiveness of AI-enabled tools in serving as direct messengers of positive math narratives to students, as part of a community of practice. The student-facing AI-enabled tools that were tested varied widely, including: 

  • EdLight’s AI-powered platform where students scan their handwritten math work and interact with a chatbot, Ember, that delivers responsive feedback
  • The Emancipation Group’s (TEG) web-based video game that features an AI-driven non-player character who teaches higher-level math through Black history
  • OKO’s AI-enabled digital platform that facilitates collaboration among small groups of students working together to solve math problems
  • Goblins’ multimodal math tutor that allows students to type, draw, or speak with an AI agent

By integrating targeted narrative interventions—including language, messages, and framing—into the core mechanics of AI-enabled tools and platforms, these organizations made progress in shifting students’ beliefs and behaviors. The results from these field tests indicate that AI-enabled tools show immense promise as effective messengers that can deliver nuanced, frequent, and well-timed positive math narratives to students in low-risk environments and increase motivation, engagement, and persistence when learning math. The field tests show potential for driving narrative change among students at scale. 

Leaders of the field tests learned important lessons about the efficacy of AI-enabled tools for delivering positive math narratives to students. The following insights emerged as a result of the field tests conducted during the 2025-2026 school year:  

1. AI-enabled tools can be prompt engineered to credibly deliver nuanced messaging, aligned to the Math Narrative Project messaging recommendations, that students trust

Developing an effective AI messenger that can deliver messaging aligned to the Math Narrative Project recommendations in a tone that students find credible and trustworthy requires much more than simply instructing a Large Language Model (LLM) to be “supportive” or “affirm mistakes.” Organizations discovered that relying on abstract instructions for their AI models often resulted in generic platitudes that students felt were dismissive of their experience of struggle. To establish a credible tone, developers had to provide the AI model with guardrails and a lot of examples, including explicitly good examples and counter examples (i.e., “few shot” examples). For example: 

  • OKO learned that the AI needed clear examples of what to do (e.g., “I see you tried a different strategy there. That mistake actually reveals a really interesting part of the math”) versus what not to do (e.g., “Great try! You’ll get it next time!”) to effectively affirm the value of mistakes, in alignment with the Math Narrative Project recommendations. OKO hard-coded these values into their “Student Collaboration Score,” ensuring the AI specifically rewarded pro-social and help-seeking behaviors rather than just cheering for correctness.
  • EdLight engineered its chatbot, Ember, using a dual-model architecture where one LLM generated raw output and a second revised it to strictly adhere to guardrails aligned to the Math Narrative Project messaging recommendations. This ensured the student-facing messaging from the AI affirmed partial understanding, validated effort, and offered multiple pathways for deeper understanding. Notably, a student who initially strongly opposed asking a chatbot for help (rating it 1 out of 7) shifted to a 4 out of 7 after having 16 supportive, question-based interactions with EdLight’s AI-enabled Ember chatbot. EdLight helped establish the credibility of their AI-enabled tool early on by hosting co-design sessions where students directly shaped the chatbot’s tone and emotional resonance.

2. AI-enabled tools offer opportunities for students to make mistakes and struggle productively in environments that lower the risk of embarrassment or social comparison

Because learning math is an inherently emotional experience, moments of struggle or failure can trigger negative emotions that suppress a student’s motivation or engagement. In classroom settings, students may hesitate to ask for help in front of their peers due to the fear of embarrassment or shame. In contrast, AI-enabled tools and platforms effectively create a private, non-judgmental environment where wrong answers carry no social cost. Lowering the stakes of making mistakes, in combination with the available support of an AI-enabled tool, enables students to persist through challenging math problems. 

  • The Emancipation Group observed this persistence phenomenon in their video game field test, noting that the AI non-player character (NPC), Thabo, functioned as a trusted messenger whose non-judgmental responses kept students in a zone where continued effort felt worthwhile. Remarkably, the AI-driven NPC even managed to convert initial hostility—such as students using profanity or declaring “math sucks”—into genuine mathematical engagement within a single session.
  • Similarly, OKO witnessed persistence by students who engaged with their AI-enabled digital platform that facilitates collaboration among small groups of students working together to solve math problems.  Initially, students exhibited fear or avoidance behaviors, asking virtually no substantive questions of the AI or each other. However, as OKO’s pro-social AI consistently validated their negative emotions as normal and reframed mistakes as valuable learning opportunities, it cultivated opportunities for students to struggle productively and seek help when they were getting stuck. This led to a significant behavioral shift. OKO documented an upward trend in “Questions Asked By Students Per Minute” from approximately 0 to .4, essentially moving students from passive avoidance to active participation because they felt safe enough to take risks.

3. AI can effectively deliver messaging “just in time” at moments of struggle 

Teachers, parents, or tutors—no matter how skilled—cannot always identify or be available to intervene at the precise moment every individual student experiences struggle with math learning or the negative emotions and beliefs that may accompany experiences of struggle. As a result, messaging aimed at reframing struggle is often delivered either preventively or long after the student’s window of receptivity to messaging has closed. AI-enabled tools, however, can be programmed to detect behaviors that signal struggle—such as multiple wrong attempts, consecutive messages without progress, or specific error patterns—and intervene in real-time. By delivering precisely timed support, AI-enabled tools sustain students through their frustration, showing promise to both shift students’ beliefs about the value of making mistakes in math learning and shift students’ behaviors related to seeking help and persisting through struggle. 

  • Goblins found that building LLM classifiers to proactively trigger interventions based on student behavior was essential. They discovered that intervening when the AI explicitly “noticed struggle” (e.g., three consecutive messages with no progress) was twice as impactful in encouraging help-seeking behaviors compared to simply offering a reactive “Need Help?” button after three minutes of inactivity. Additionally, affirming help-seeking after a student has already sought help had a smaller impact on boosting students’ belief in the value of asking for help or stated intention to ask for help in the future.
  • When The Emancipation Group’s AI-driven non-player character was programmed to deliver scaffolding language—tailored messages and questions that prompted students to reflect or modeled how to work through a problem—to help learners master new skills or content exactly when students provided incorrect answers, it fostered remarkable resilience. Nearly a third (29%) of analyzed transcripts revealed “productive persistence,” with students continuing through four to seven wrong attempts until reaching the correct solution.

4. AI-enabled tools can increase the dosage of exposure to positive math narrative messaging, creating a positive feedback loop that strengthens impact 

Narratives are developed over time through repeated exposure to stories and messages. Accordingly, meaningful shifts in deeply held beliefs that are shaped by those narratives—such as students’ beliefs about their capability to learn math or motivation to ask for help or persist through challenges—are unlikely to occur after exposure to a single new story or message, whether that is from a human or an AI agent or chatbot. Narrative interventions designed to shift beliefs require sustained, repeated exposure over time. AI-enabled tools, embedded in the classroom experience, offer an opportunity to increase the number of times students are exposed to positive math narrative messaging. Moreover, the continuous feedback loops inherent in AI-enabled tools keep students engaged during difficult tasks when they might otherwise drop off, thereby naturally increasing the dosage of positive messaging they receive. 

  • EdLight’s field test documented a clear pattern reflecting a response to the dose of messaging students encountered. They found that “super users”—those who interacted with the Ember chatbot six or more times—showed the most consistent positive movement across targeted belief pathways. These “super users” demonstrated a +1 average increase (on a 7-point scale) over 9 weeks in viewing mistakes as helpful feedback and strongly rejected the belief that making mistakes means you are bad at math. One notable EdLight student case study revealed that after 16 interactions, the student experienced a massive shift from strongly disagreeing that mistakes help, to recognizing their distinct value. Conversely, students with limited or no exposure to the Ember chatbot showed little to no change in their beliefs about the value of mistakes in math learning.
  • OKO observed that greater exposure to positive math narrative messaging through AI interactions increased students’ persistence in the face of challenging math content. Their field test found that the AI-delivered messaging moved students past a “difficulty dip” in November, where the difficulty of the material had increased since the beginning of the year causing students’ avoidance behaviors to increase (students’ “Math Proficiency Levels” dipped to “Needs Intervention”). By December, students using their AI had achieved a bounce-back in “Math Proficiency Levels.” Consistent exposure to a high dosage of messaging is critical when students are engaging with more difficult math; the behavioral data OKO observed in the platform matched the increase in students’ self-reported positive beliefs about persistence in survey data.  

Ultimately, AI-enabled tools present a profound opportunity to scale high-quality, deeply personal interactions that support students’ engagement, motivation, and persistence in learning math. By engineering AI to act as an empathetic, non-judgmental, and perfectly timed messenger, organizations have successfully disrupted deficit-oriented narratives and fostered genuine productive struggle. However, a critical next step is further integrating these AI tools—like chatbots, video games, and digital math tutors—directly into classroom environments where teachers are also actively engaged in the narrative shift. As OKO emphasizes, achieving lasting mindset shifts requires “instructional coherence,” where teachers are equipped to actively deliver similar messaging to students as those embedded within the AI-enabled tools. Amplifying and surrounding students with positive math narratives from human and AI messengers can ensure that every student can engage productively and persist through challenges they encounter in math learning.