Showcase from the AI in the Global South seminar: Project #2 - Malam: A Low-Bandwidth AI Math Tutor for Afghan Girls

Last semester, Dr. Vikram Kamath Cannanure from I2SC organized the AI in the Global South Seminar where students from Saarland University worked together to discuss case studies, brainstorm the challenges in AI adoption, and build AI applications to support the Global South. In this series of blog posts, we showcase the projects created by the seminar participants. 

Project: Malam: A Low-Bandwidth AI Math Tutor for Afghan Girls

Team: Tedis Cico, Ahmad Fadel, Mahalakshmi Raveenthiran, Désirée Wiltzius

After the Taliban enforced an education ban for girls after sixth grade in late 2021, a lot of girls and women have turned to online studying resources. Our project explores how generative AI (GenAI) can provide safe, motivational, and culturally responsive math tutoring for Afghan girls. To better understand this dreadful problem, we have read papers, watched interviews, listened to podcasts and, most importantly, talked to Hamyoon Behmanush, an Afghan doctoral researcher at Saarland University.

1 Problem and Context

During the seminar presentations, we have discussed multiple topics regarding AI in the global south. We have also seen that there are still problems in the global south that need to be addressed, such as insufficient financial resources, unqualified teachers or inadequate classroom facilities [13]. One of the topics discussed was LLMs (Large Language Models) for learning. LLMs offer a large range of possibilities, e.g. an AI Math Tutor in Ghana or GPT Tutor [16, 10].

After reading the latest published paper that Vikram Kamath Cannanure (who held this seminar) worked on, our focus shifted to the learning situation for girls after sixth grade in Afghanistan [12]. Since late 2021, education has been banned for girls after the age of 12 or after sixth grade in Afghanistan [18].

In 2021, the global literacy rate¹ was 86.93% [1]. In Afghanistan, the overall literacy rate recorded was 37%. For men, the literacy rate was 52.1% and for women it was 22.6%. [2]. As we can see, the difference between the global literacy rate and the literacy rate in Afghanistan is substantial. We can only imagine the damage this ban on education will have on literacy rates, especially for women, in Afghanistan.

This ban has led females to search for alternative opportunities to educate themselves and find potential work. This is why, for our project, we focused on establishing an LLM for learning math for girls in Afghanistan after sixth grade.

2 Understanding Context

Our team's motivation for this topic emerged from both personal and academic reasons. For

Désirée and Mahalakshmi, the issue resonated deeply on a personal level. As women currently pursuing higher education, they could not imagine being denied this opportunity solely because of their gender. Ahmad, having grown up in a country that also experienced authoritarian rule, related to the fear and oppression faced by those who resist or challenge such regimes. Tedis, who closely follows international affairs, had been tracking the situation in Afghanistan since the Taliban takeover and was keen to understand its long-term societal implications.

To move beyond personal empathy and gain a grounded understanding, we explored multiple sources. We reviewed UNESCO reports, recent education papers, and especially the work of Behmanush et al. [12], which studies how Afghan women use online learning and GenAI tools under educational bans. We also watched interviews and first-hand video accounts on platforms such as YouTube, while remaining cautious about possible biases. Most importantly, we spoke directly with Hamayoon Behmanush, an Afghan doctoral researcher at Saarland University, who offered valuable insight into the realities of women's online learning in Afghanistan today.

This combination of personal reflection, literature review, and direct conversation helped us grasp not only the technical barriers such as unreliable internet access and restricted platforms, but also the emotional and social dimensions of learning under surveillance. These insights informed our decision to design a low-bandwidth AI math tutor that offers both educational and motivational support in Dari, Pashto, and English, ensuring accessibility for a broader range of Afghan learners.

3 Literature Review

Our work all started from the paper Online Learning and GenAI: Supporting Women's Aspi-

rations Amid Socio-Political Instability in Afghanistan. This paper investigates how women in

Afghanistan, in the face of extreme sociopolitical instability and education bans, are using online learning and GenAI (generative AI) to pursue introductory programming education. The authors find that, even though internet access is often restricted, establishing GenAI tools could be an interesting idea to help women in Afghanistan study [12].

GenAI is fundamentally transforming the educational landscape [14]. Research emphasizes the idea of using GenAI to support marginalized learners, particularly women in Afghanistan [12, 11]. These learners deal with persistent challenges, such as inconsistent internet access, language barriers and the absence of live tutoring or mentors [11, 12]. Through a study, one paper can show that incorporating a GPT-4 based conversational agent into online class discussions significantly improved the engagement of the students [15]. Other authors mention advantages such as personalized and interactive learning, automated essay grading and language translation [8]. Other authors mention disadvantages such as lack of human interaction, biases in training data and lack of creative thinking (note: Advait Sarkar's paper on creative thinking [17]). In recognizing this foundational shift in education, other authors emphasize the necessity for collaboration among researchers, educators, technology experts and policymakers to ensure safe usage [8]. We can conclude that the most listed disadvantages for using GenAI for students in countries from the global south were:

• reliance on GenAI and loss of critical thinking
• no stable internet connection

We will tackle both of these problems in our project. We want to highlight that the current government in Afghanistan has, on multiple occasions, cut off internet and phone networks, leading to failures in using GenAI [3].

4 Comparing neighboring apps and models

The market for our product would be girls in Afghanistan from age twelve on. We would like to highlight that even though we have to look at our 'competitors' we don't see them as such, as our objective is to help out girls in Afghanistan and not be 'the biggest' in our market.

4.1 Solax

Solax is an initiative that helps Afghan girls and young women continue their education online. They offer whatsapp-based micro-lessons and are simple and accessible [4].

Figure 1: SolaX logo

4.2 Sahar

Sahar offers free online courses for Afghan women in subjects like computer science, business, social sciences and language courses. They offer low-data access and download features [5].

Figure 2: Sahar logo

4.3 SolaTeach

SolaTeach offers free online courses for Afghan women in subjects like English, digital literacy and professional skills. Solateach has won the Global Recognition Award (GRA) in 2025 for its commitment in digital education and for empowering girls in Afghanistan [6].

Figure 3: SolaTeach logo

4.4 Right to learn Afghanistan

Right to learn offers online high school courses through 'Darakht-e Danesh' (knowledge tree in Dari). They offer courses in subjects such as English, Dari and Pashto [7].

Figure 4: Right to Learn Afghanistan logo

4.5 Rori

Rori is an AI-powered math tutor developed by Rising Academies in Ghana that delivers over 500 short math lessons through WhatsApp [16]. Each lesson includes explanations, scaffolded exercises, and adaptive hints based on the Teaching at the Right Level and

Intelligent Tutoring System approaches. Our product builds on Rori's model of low-cost, mobile-based tutoring but adapts it for Afghan girls by adding language support in Dari, Pashto, and English, motivational messaging, and offline access.

Figure 5: Interaction examples with Rori, the AI-powered math tutor on WhatsApp [16].

4.6 What they are doing well and what our product can fill

The reviewed platforms SolaX, Sahar, SolaTeach, Right to Learn, and Rori are all making important contributions to female education in Afghanistan and other low-resource regions. They successfully provide accessible learning through low-data platforms like WhatsApp and browser-based lessons, and some even include certification or downloadable content. Rori, in particular, demonstrates how conversational AI can improve math learning outcomes in low-bandwidth environments. However, most of these tools still rely on the stable internet and present static or text-heavy content with limited adaptability and engagement features. They rarely address the motivational and emotional needs of learners who are isolated from formal education.

The gaps our product aims to fill include making learning more interactive and engaging by gamifying the progress, providing offline functionality for areas with restricted internet, integrating multimedia and audio support for low-literacy learners, and embedding motivational texts and examples of successful Afghan women to inspire continued learning. We made the replies from Malam short, so the girls would not feel overwhelmed by so much text.

5 Our Product Plan

Our model, named Malam, aims to support the Afghan community in mathematics education. Malam acts as both a teacher and a textbook, offering self-created "pre-recorded" lessons in Dari, Pashto, and English. At the beginning of the learning journey, Malam conducts an entry evaluation - similar to Rori to assess each student's level and recommend a personalized learning path [16]. Lessons are divided into chapters, each ending with a quiz to ensure understanding and address the issue of "solving without learning." Students can only proceed to the next topic after passing the quiz, promoting mastery-based progression. The lessons are gamified to make learning more engaging and, whenever possible, highlight stories of successful Afghan women to motivate learners.

We first developed Malam as a custom ChatGPT prototype to explore conversational flow, math explanations, and motivational feedback in a controlled environment. After validating the interaction design and content structure, we implemented an early working prototype as a Telegram chatbot. The Telegram version delivers short math explanations and multiple-choice questions while offering hints, encouragement, and explanations in the learner's preferred language - Dari, Pashto, or English. Telegram was chosen because it is widely used in Afghanistan, functions well on low-end Android phones, and can operate under low-bandwidth conditions. This multilingual chatbot format ensures that Malam remains accessible even for learners with limited English proficiency or unstable internet connections. The screenshots below show the current interaction flow.

A publicly accessible version of Malam is available as a custom ChatGPT prototype. It demonstrates the same conversational flow, quiz logic, and motivational tone explored during development. You can access it here: Malam ChatGPT Prototype.

Figure 6: From Custom ChatGPT Prototype: Malam Welcome Text

Figure 7: From Custom ChatGPT Prototype: Malam Intro points

Figure 8: Malam Telegram chatbot: the learner begins the math journey and can switch between Dari, Pashto, or English.

Figure 9: Lesson roadmap and example of Malam explaining the same content in Pashto, showing multilingual support.

5.1 Adapting Existing Models

Malam adapts existing LLM-based tutoring approaches such as Rori by offering personalized, chapter-based lessons accessible via WhatsApp or Telegram. While Rori focuses on math tutoring through text conversations, Malam extends this idea with built-in quizzes, progress tracking, and motivational elements. It uses a similar structure of short micro-lessons but adapts them for learners in low-connectivity regions, supporting interaction in Dari, Pashto, and English to ensure inclusivity and accessibility for a wider Afghan audience.

Figure 10: From Custom ChatGPT Prototype: Malam Quiz

Figure 11: From Custom ChatGPT Prototype: Malam additional resources

5.2 Learning from Context

From our conversations with Hamayoon Behmanush and review of the Afghan education situation, we learned that learners face barriers such as unstable internet, limited English proficiency, and the lack of safe educational spaces [12]. Therefore, Malam is designed to function in multiple languages (Dari, Pashto, and English), operate on minimal bandwidth, and provide a private, phone-based learning experience. The initial evaluation ensures that learners start at a comfortable level, reducing frustration and dropout risk.

5.3 Adaptation from Literature

Research on GenAI for marginalized learners shows that AI tutors can enhance learning when they are adaptive, motivational, and culturally sensitive [14, 11]. Malam incorporates these insights by adjusting its explanations based on learner performance, encouraging reflection after each quiz, and offering positive reinforcement. It also follows the Teaching at the Right Level (TaRL) approach, ensuring that the lessons match the learner's ability instead of fixed grade levels [9].

5.4 Adaptation from the Market

We drew inspiration from educational tools like SolaX and Sahar, which offer online micro-lessons for Afghan learners [4, 5]. However, these platforms rely heavily on continuous internet access and static lessons. Malam differentiates itself by combining adaptive learning with offline functionality and gamification. It also includes motivational storytelling elements that highlight Afghan female role models, something currently missing in existing solutions.

6 What More Can Be Done

While our current focus is on mathematics, the long-term vision for Malam extends far beyond a single subject. In future iterations, Malam could be made fully downloadable and periodically updatable offline, allowing students in areas with no internet access to continue learning without interruption. To ensure accessibility for all learners, an initial comprehension test could help match students to the most suitable language mode. We also envision creating multiple subject-specific versions of Malam, such as Malam Logic or Malam Language, each focusing on core domains of the Afghan curriculum. Finally, community and NGO partnerships could help distribute devices preloaded with Malam, provide teacher training, and collect feedback from learners to continuously improve the system.

References

[1] https://www.macrotrends.net/global-metrics/countries/wld/world/literacy-rate. Accessed: 19/10/2025.

[2] https://worldpopulationreview.com/country-rankings/literacy-rate-by-country. Accessed: 19/10/2025.

[3] https://edition.cnn.com/2025/10/03/asia/afghanistan-internet-shutdown-intl-hnk-dst. Accessed: 19/10/20255.

[4] https://www.solax.org/. Accessed: 20/10/2025.

[5] https://www.sahareducation.org/. Accessed: 20/10/2025.

[6] https://solateach.com/. Accessed: 20/10/2025.

[7] https://righttolearn.ca/. Accessed: 20/10/2025.

[8] D. Baidoo-Anu and L. O. Ansah. Education in the era of generative artificial intelligence (ai): Understanding the potential benefits of chatgpt in promoting teaching and learning. Journal of AI, 7(1):52-62, 2023.

[9] A. Banerjee, R. Banerji, J. Berry, E. Duflo, H. Kannan, S. Mukerji, M. Shotland, and M. Walton. Mainstreaming an effective intervention: Evidence from randomized eavluations of 'teaching at the right level' in india. CEPR Discussion Paper Series, 2016.

[10] H. Bastani, O. Bastani, A. Sungu, H. Ge, ¨O. Kabakcı, and R. Mariman. Generative ai can harm learning. The Wharton School Research Paper, 2024.

[11] H. Behmanush. Supporting marginalized learners with genai. In Proceedings of the

AAAI/ACM Conference on AI, Ethics, and Society, volume 8, pages 2848-2849, 2025.

[12] H. Behmanush, F. Akhtari, R. Nooripour, I. Weber, and V. K. Cannanure. Online learning and genai: Supporting women's aspirations amid socio-political instability in afghanistan. In Proceedings of the ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies, pages 401-417, 2025.

[13] R. Chandra. Challenges and vision in educating the global south. Int J Res Anal Rev, 9(3):277-280, 2022.

[14] P. Denny, S. Gulwani, N. T. Heffernan, T. K¨aser, S. Moore, A. N. Rafferty, and A. Singla. Generative ai for education (gaied): Advances, opportunities, and challenges. arXiv preprint arXiv:2402.01580, 2024.

[15] J. A. Haqbeen, S. Sahab, and T. Ito. Assessment of the llm-based chatbots on student engagement and learning outcomes in afghanistan. In Conference on Digital Government Research, volume 1, 2025.

[16] O. Henkel, H. Horne-Robinson, N. Kozhakhmetova, and A. Lee. Effective and scalable math support: Evidence on the impact of an ai-tutor on math achievement in ghana. arXiv preprint arXiv:2402.09809, 2024.

[17] H.-P. Lee, A. Sarkar, L. Tankelevitch, I. Drosos, S. Rintel, R. Banks, and N. Wilson. The impact of generative ai on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI conference on human factors in computing systems, pages 1-22, 2025.

[18] A. Q. Sarwari and H. M. Adnan. Analysis of discourses in the islamic world about the ban upon women's education by the taliban in afghanistan. Issues in Educational Research, 33(3):1148-160, 2023