Safe and Effective AI for Learning: A Chatbot for Afghanistan

• In Event: Ensuring quality and standards in AI-EdTech Part 1

Proposal

Learning challenges in Afghanistan are unique. Girls are banned from formal education from grade 7 onwards. But even schoolgoing students face low-quality education, dominated by lecture-based lessons with little interaction and poor learning outcomes.

Lapis developed a tutor chatbot delivered by WhatsApp. The chatbot complements the extensive portfolio of roughly 2,000 curriculum-based TV shows, radio broadcasts , and digital learning platforms that offer a comprehensive suite of remote learning modalities.

An appropriate, usable and safe chatbot for Afghanistan should meet several criteria, ensuring: safety for girls and for the implementing organisation; a cultural and linguistic fit; curriculum-alignment and relevance for learners’ study journey; pedagogical soundness; and student access despite bandwidth constraints.

In Afghanistan, safety goes beyond privacy. A poorly designed system can expose both users and implementers to real risks. The development team conducted extensive red-teaming to avoid conversations on sensitive non-curricular subjects, such as religion, conflicts, politics, or history. At the same time, the chatbot was designed to comply with the UN Human Rights framework and to not block responses to genuinely relevant non-curricular queries.

Until recently, it was not evident that large language models (LLMs) could adequately function in Dari and Pashto, Afghanistan’s primary national languages which are both underrepresented in LLMs. However, several models in our evaluation matrix now perform well enough to be deployed in these languages. Multiple models respond in Dari or Pashto, but only two currently respond to pre-defined prompts with meaningful responses and acceptable low levels of hallucination. For these two models, costing has been the main consideration for selecting a preferred model, with the other model as a fallback in the engine. Lapis adopted an out-of-the-box LLM with a Retrieval-Augmented Generation (RAG) approach and ensured curricular and cultural relevance. The RAG pipeline drew on a corpus of roughly 15 million words of high-quality curriculum-aligned text content in Dari and Pashto.

The system is designed to minimise cognitive offloading by guiding learners step by step through problem-solving. Progression requires a relevant user response after each step, especially in mathematics and other procedural subjects.

Lapis relies on several metrics to measure quality and effectiveness. Online analytics track user retention, frequency of visits, conversation length and user engagement. Students seeking support are referred to a human responder via the same WhatsApp channel used for other educational queries and complaints. Focus Group Discussions with students and teachers provide insights into the ease of use, curriculum alignment, and cultural fit. Finally, regular surveys collect structured feedback and suggestions for improvement.

Author

• Alex Thier, Lapis