Bridging Digital and Linguistic Divides: Piloting AI-Powered Personalized Learning in Rural India

• In Event: Investigating Generative AI Use and Integration in International Contexts: Opportunities, Challenges, and Insights

Abstract

This paper investigates the potential of generative artificial intelligence to drive pedagogical advancement and educational equity in under-resourced rural settings. We present findings from a field deployment of TALI (Teaching At Learner's Interest), an AI-powered learning companion designed to evaluate handwritten work and generate personalized, rubric-aligned learning plans for Marathi-medium students in Maharashtra, India. We assess how AI-driven personalization, designed for local contexts and languages, can address chronic challenges such as teacher workload and diverse learner needs by providing immediate, accurate formative feedback. (Yuskovych-Zhukovska et al., 2022).
In the May 2025 pilot across 12 rural Zilla Parishad schools (n=225 students), TALI used a custom Gemini-based AI model to process and evaluate student handwriting in both Marathi and English. The system demonstrated the ability to grade a 50-mark exam in under 32 seconds with 95% accuracy relative to teacher judgment, and reduced teacher grading workload by up to 40%. Teachers reported enhanced insights into class- and student-level learning gaps, enabling data-driven instruction and targeted support. Notably, 92% of surveyed students reported feeling clearer and more confident in their learning progression, and 12% of scores shifted when AI grading was compared with teacher grading, indicating increased assessment precision. Importantly, by working with paper-based student responses, TALI mitigates the risk of digital exclusion and supports equitable AI integration even in low-connectivity environments.
This research is grounded in sociocultural theories of learning and digital equity (Vygotsky, 1978), emphasizing the importance of culturally and linguistically responsive pedagogy in bridging opportunity gaps (Au, 2018; Paris and Alim, 2017). We also draw on frameworks of AI for social good (Bender, 2024), which highlight the need for human-in-the-loop systems that support teachers through augmentation rather than replace or deprofessionalize them (Luckin et al., 2016), and facilitate inclusive, differentiated instruction at scale. We interrogate the human-centred AI framework that considers the impact of AI interventions not just on the direct user (teachers), but on the classroom, community, and society in terms of second-order effects (Schmager et al. 2025). Finally, we draw on theories of impactful formative assessment and feedback, with evidence that delayed or generic feedback can reduce learning gains (Shute, 2008) to investigate the role of AI in this use case.

Our study extends the global conversation on AI in education by centering the needs of marginalized learners and local language classrooms. It addresses the underexplored intersection of generative AI, handwriting recognition, and low-resource educational contexts in the Global South, offering practical strategies for sustainable, scalable impact (Yuskovych-Zhukovska et al., 2022; Bender, 2024). Integration with teachers’ workflows fosters professional agency and builds foundational AI literacy among educators, aligning with contemporary calls for co-creation and ethical implementation of educational technologies. This work contributes new evidence for the promise and challenges of generative AI in advancing UN Sustainable Development Goal 4 by expanding access to responsive, high-quality education for all. We discuss the implications for future research and policy, including design considerations for inclusive AI and the importance of participatory innovation in international contexts.

Authors

• Zainab Azim, Harvard University

• Ranjitsinh Disale, Harvard University