Shiksha Copilot: Teacher-AI Collaboration for Curating and Customizing Lesson Plans in Low-Resource Schools

This study evaluates Shiksha Copilot, an AI-assisted tool deployed in Karnataka, India, which successfully reduced teachers' administrative burdens and stress while fostering activity-based pedagogy through a human-AI collaborative workflow, though its broader pedagogical impact was limited by systemic challenges like staffing shortages.

The Gist

Imagine a teacher in a rural school in India. They are juggling 50 students in a crowded classroom, dealing with a shortage of textbooks, and facing a mountain of paperwork. Their boss demands a detailed "lesson plan" every day—a document proving they know what they are going to teach. But between managing the kids and the paperwork, they have almost no time to actually create those plans. They are exhausted, and their creativity is running on empty.

Now, imagine handing that teacher a super-smart, tireless assistant who can draft a perfect lesson plan in seconds, but with one catch: the assistant sometimes speaks a bit awkwardly or misses the local flavor.

This is the story of Shiksha Copilot, a new AI tool tested in Karnataka, India, designed to help teachers in low-resource schools. Here is how it works, broken down into simple concepts.

1. The Problem: The "Paperwork Trap"

In many schools in the "Global South" (developing regions), teachers are drowning in administrative work. They have to write lesson plans not just to teach, but to prove to the government that they are working. It's like a chef who has to spend 90% of their time writing a menu for the health inspector and only 10% actually cooking. This leads to burnout and less time for actual teaching.

2. The Solution: A "Co-Pilot" for Teachers

The researchers built an AI tool called Shiksha Copilot. Think of it not as a robot that replaces the teacher, but as a co-pilot in a cockpit.

• The AI (The Co-pilot): It flies the plane (drafts the lesson plan) using a massive library of textbooks and curriculum rules. It can write in English and the local language (Kannada).
• The Teacher (The Captain): They don't just sit back. They check the map, steer the plane if it goes off course, and decide where to land.

3. How It Works: The "Human-in-the-Loop" Kitchen

The system uses a clever three-step process, like a high-end restaurant kitchen:

• Step 1: The AI Chef (Drafting): The AI reads the textbooks and generates a draft lesson plan. It's fast and follows the rules (like the "5E" teaching model: Engage, Explore, Explain, Elaborate, Evaluate).
• Step 2: The Head Chef (Curation): Before the plan reaches the teachers, a team of expert human teachers (called "curators") reviews the AI's work. They fix mistakes, ensure the science is accurate, and make sure the language sounds natural.
  • The Catch: The AI is great at English. But when it translates to Kannada, it sometimes sounds like a robot reading a dictionary. The human curators had to rewrite about 95% of the Kannada content to make it sound like a real teacher speaking to real kids.
• Step 3: The Teacher (Customization): The teacher gets the "pre-cooked" meal. They can tweak it to fit their specific class. Maybe they swap a generic example for one about a local village festival. They can also ask the AI for extra help, like "Give me 5 quiz questions about photosynthesis."

4. What Happened? The Results

The study involved over 1,000 teachers. Here is what they found:

• Time Saved: Teachers saved about 2 hours a week. That's like getting an extra day off every month just for planning.
• Less Stress: The "paperwork trap" was broken. Teachers felt less stressed because they didn't have to stare at a blank page for hours.
• Better Teaching: Because they had more time, teachers started using more activities (like games, scavenger hunts, and group work) instead of just lecturing. The AI gave them a menu of fun ideas they could actually use.
• The "Bureaucracy" Win: The tool generated documents that looked official and satisfied the government inspectors. Teachers could print them or copy them by hand, but they didn't have to invent the content from scratch.

5. The Limitations: Why AI Isn't a Magic Wand

The paper also highlights some real-world hurdles:

• The Language Gap: While the English plans were great, the Kannada plans needed a lot of human polishing. AI still struggles with the nuance and "soul" of local languages.
• No Visuals: The AI is text-only. It can't draw a picture of a cell or a diagram of a machine. Teachers still had to find images on their own phones to make the lessons visual.
• Systemic Issues: The tool helped the teacher, but it couldn't fix the school. If a school is short-staffed (e.g., 3 teachers for 7 classes), even the best AI lesson plan can't fix the chaos of overcrowding. The tool reduces the teacher's burden, but it doesn't remove the system's problems.

The Big Picture

Shiksha Copilot teaches us that AI in education shouldn't be about replacing teachers with robots. Instead, it should be about partnership.

Think of it like a GPS. The GPS (AI) knows the fastest route and the traffic rules. But the Driver (Teacher) knows the road conditions, the passengers' needs, and when to take a scenic detour. When they work together, the journey is smoother, faster, and less stressful for everyone.

The study concludes that for AI to truly help in places like India, it needs to be:

1. Human-checked (to fix language and cultural errors).
2. Community-driven (so teachers can share their own tweaks and ideas).
3. Respectful of the teacher's role (as the expert who knows their students best).

In short: AI can do the heavy lifting of writing, but the teacher must still steer the ship.

Technical Summary

Here is a detailed technical summary of the paper "Shiksha Copilot: Teacher-AI Collaboration for Curating and Customizing Lesson Plans in Low-Resource Schools."

1. Problem Statement

The study addresses the critical challenges faced by teachers in low-resource schools in the Global South, specifically in Karnataka, India. These educators operate under severe constraints, including:

• Overcrowded classrooms and staff shortages: Leading to excessive workloads and burnout.
• Bureaucratic burdens: Teachers must create detailed lesson plans (LPs) for administrative compliance, often in a mandated "5E Instructional Model" format, which consumes significant time (45–90 minutes per plan).
• Resource scarcity: Limited access to high-quality instructional materials, digital infrastructure, and pedagogical support.
• Limitations of existing AI: While Large Language Models (LLMs) can generate content, they often lack factual accuracy, cultural sensitivity, and alignment with local curricula, particularly for non-English languages (like Kannada) and low-resource contexts.

The core problem is the need for an AI system that not only generates curriculum-aligned content but also integrates seamlessly into the existing workflows of teachers in multilingual, resource-constrained environments without increasing their cognitive load.

2. Methodology

The research employed a large-scale mixed-methods approach involving the design, deployment, and evaluation of "Shiksha Copilot."

• System Deployment:
  • Context: Deployed in government schools across Karnataka, India, from December 2024 to March 2025.
  • Participants: 1,043 teachers (Grades 5–10) and 23 human curators (subject matter experts).
  • Languages: English and Kannada (bilingual support).
• Data Collection:
  • System Logs: Analyzed 991 curated LPs and 5,544 teacher-generated LPs to track edit patterns, feature usage, and adaptation rates.
  • Surveys: Pre- and post-intervention surveys (n=406 matched responses) measuring teaching-related stress, time use, access to resources, and trust in AI.
  • Interviews: Semi-structured interviews with 43 stakeholders (17 curators, 22 teachers, 4 mentors) to understand qualitative experiences, adaptation strategies, and perceived barriers.
• Analytical Framework:
  • Quantitative: Statistical analysis of edit frequencies, time savings, and stress scores (using Teacher Stress Scale).
  • Qualitative: Thematic analysis of interview transcripts and open-ended survey responses, focusing on teacher-AI collaboration and communities of practice.

3. System Architecture and Design

Shiksha Copilot is a Human-in-the-Loop (HIL) AI system designed with a two-phase workflow: Curation and Distribution.

• Technical Stack:
  • Curriculum Ingestion: Textbooks (Science, Math, Social Science, English) were converted to text/images, chunked, and indexed using Retrieval-Augmented Generation (RAG) with Azure AI Search.
  • Translation Engine: A fine-tuned translation model (based on Sarvam-1 using LoRA) optimized for educational terminology to translate English LPs into Kannada.
  • LLM Orchestration: An agentic framework (using OpenAI GPT-4o) managing specialized agents:
    • RAG Agent: Retrieves curriculum context.
    • Lesson Plan Gen Agent: Constructs LPs following the 5E model (Engage, Explore, Explain, Elaborate, Evaluate).
    • Question Gen Agent: Creates assessments based on Bloom's Taxonomy.
    • Chat Agents: "Lesson Chat" (context-specific) and "Edu Chat" (general pedagogy).
• Workflow:
  1. Curation Phase: AI generates draft LPs; human curators review, edit, and validate them for accuracy and pedagogical soundness.
  2. Distribution Phase: Curated LPs are stored in a database. Teachers access them via a mobile-first web interface, create copies, and customize them for their specific classrooms.
• User Roles:
  • Curators: Full editing and validation rights.
  • Power Users (Teachers): Access to AI-assisted editing, regeneration, and chatbots.
  • Standard Users: Access to view, manually edit, and use the Question Generator.

4. Key Contributions

The paper makes three primary contributions to the CSCW and HCI fields:

1. Design of a Bilingual HIL System: Demonstrates a scalable architecture for generating and curating curriculum-aligned content in both English and a low-resource language (Kannada), bridging the gap between global LLM capabilities and local educational needs.
2. Empirical Evidence of Teacher-AI Collaboration: Provides data on how teachers in the Global South actively evaluate, adapt, and contextualize AI content, moving beyond the "passive user" model to an "active collaborator" model.
3. Workflow Transformation Analysis: Shows how AI tools can simultaneously reduce administrative burdens (record-keeping) and shift pedagogical practices toward activity-based learning, while highlighting the systemic constraints (staffing, policy) that limit full transformation.

5. Key Results and Findings

A. Quality and Adaptation

• English Content: High reliability. Only 4.1% of content blocks required editing. Errors were mostly minor formatting or assessment inaccuracies.
• Kannada Content: Significant linguistic challenges. 94.4% of Kannada content blocks required editing. Issues included literal translations, unnatural phrasing, and incorrect terminology (e.g., mathematical terms). This highlights the limitations of current translation models for specialized educational domains.
• Teacher Adaptation: Teachers rarely edited plans on the platform (only ~1.5% of plans were modified digitally). Instead, they engaged in "mental adaptation," using their professional expertise to modify plans offline or during class, relying on the AI output as a scaffold rather than a final product.

B. Impact on Workload and Stress

• Time Savings: Teachers saved an average of 2.02 hours per week on lesson planning. Active users saved up to 2.64 hours.
• Stress Reduction: Teaching-related stress scores decreased significantly (mean drop of 0.72 points on a 7-point scale), particularly regarding "teaching demands."
• Bureaucratic Efficiency: The tool successfully automated the creation of 5E-compliant documentation, reducing the time to generate submission-ready plans from ~60 minutes to <15 minutes.

C. Pedagogical Shifts

• Activity-Based Learning: The tool introduced a shift toward activity-based pedagogy. Teachers utilized the AI-generated activity ideas (e.g., "Clap and Tap," "Treasure Hunts") which required no digital infrastructure, making them suitable for low-resource classrooms.
• Assessment Generation: Teachers heavily utilized the "Question Gen" feature to create diverse assessments, moving away from rote memorization toward critical thinking questions.

D. Limitations and Challenges

• Visual Deficit: The text-heavy nature of the LPs was a major limitation, especially for Science and Math, where visual aids are crucial. Teachers had to search externally for images.
• Contextualization: While the content was regionally relevant (Karnataka), it lacked hyper-local relevance (specific neighborhoods), requiring teachers to mentally adapt examples.
• Systemic Constraints: The tool could not solve non-pedagogical issues like severe teacher shortages or excessive administrative data entry (e.g., APAAR ID requirements), which continued to consume instructional time.

6. Significance and Future Directions

• Human-in-the-Loop is Essential: The study validates that for low-resource, multilingual contexts, AI cannot replace human expertise. A curation layer (curators) is critical to ensure quality, especially for vernacular languages.
• Communities of Practice (CoP): The authors argue that AI tools must be embedded within existing professional networks (e.g., District Institutes, Cluster Resource Centers). Future iterations should enable teachers to share their adapted plans and corrections, creating a "living repository" of localized knowledge.
• Design for the Global South: Success in these contexts requires mobile-first, low-bandwidth designs and a focus on reducing bureaucratic friction, not just improving instructional content.
• Epistemic Justice: The paper calls for moving beyond translation-only approaches to developing native LLMs and content creation pipelines in local languages to preserve distinct epistemologies and cultural nuances.

In conclusion, Shiksha Copilot demonstrates that AI can be a powerful lever for reducing teacher burnout and enhancing instructional quality in low-resource settings, provided it is designed as a collaborative tool that respects local constraints and leverages human expertise for curation and contextualization.