An AI lesson-planning tool developed for public schools in southern India saved participating teachers about two hours a week, according to a study involving Cornell researchers. The results also showed how much human review the tool needed: nearly every Kannada-language lesson plan examined by curators required edits.
The study by Cornell and Microsoft Research India researchers examined Shiksha Copilot in fifth- through 10th-grade classrooms in Karnataka, India. Researchers analyzed use by 1,043 teachers and the work of 23 curators who reviewed AI-generated material before teachers adapted it for their classes. A Cornell Chronicle account published Oct. 8 said the team is now testing the system with about 30,000 teachers in the neighboring state of Telangana.
Teachers used plans for instruction and paperwork
Microsoft Research India initially developed the system, which turns textbook content into grade-appropriate lesson plans in English and translates them into Kannada. Curators review the plans for accuracy, language and local relevance, and teachers can customize the results. The system can suggest activities using materials available in schools with limited resources, such as collecting leaves to compare vein patterns.
Researchers studied the Karnataka deployment from December 2024 through March 2025 using system logs, interviews and surveys. They interviewed 22 teachers, 17 curators and four mentors. Of 601 teachers who completed a survey before deployment, 406 also completed the follow-up survey three months later; the researchers based their before-and-after analysis on those matched responses.
Those teachers reported an average reduction of 2.02 hours a week in lesson-planning time. The study also found a modest decline in reported stress related to teaching demands. Teachers described using the plans both to prepare instruction and to meet requirements to submit written lesson plans to school authorities.
The system was built for phones as well as computers. Platform records showed 87.6% of teachers accessed it on smartphones. Some could print completed plans for supervisors, while others still had to submit handwritten copies and used the generated material as a guide.
Teachers also described a shift toward more classroom activities instead of relying on lectures. The plans included local examples and exercises that did not require screens or specialized equipment. The researchers said staffing shortages and other administrative duties still limited how consistently teachers could put those ideas into practice.
Kannada language required much more review
The difference between English and Kannada material was substantial. Of 7,744 English-language content blocks reviewed, curators rated 96.13% suitable for classroom use, marked 3.53% for minor adjustments and rejected 0.34% because of technical errors. The 7,744 figure counts parts of plans, not whole lesson plans.
In a separate analysis of 26 Kannada-language plans, curators changed 94.4% of content blocks. They found literal translations, awkward phrasing and terms that did not match the language used in local textbooks. Teachers nonetheless rated about 81% of Kannada plans “very good,” compared with 87% of English plans.
The study said teachers often adapted material to their students without recording changes in the platform. Some wanted more visual aids, and some said plans labeled for beginners assumed skills their students had not yet learned. The researchers said state-level examples were useful but could miss the more specific circumstances of an individual classroom.
“With English, there are lots of resources available on the internet already, but that’s not the case for languages like Kannada,” Deepak Varuvel Dennison, a Cornell doctoral student and the study’s first author, told the Cornell Chronicle. He said experienced teachers reviewed the plans before they were distributed.
The study did not measure student learning
The authors said the study covered only part of one school year, one Indian state and grades five through 10. It relied on teacher reports, interviews and platform records; it did not test whether students learned more. The before-and-after findings describe participating teachers and cannot establish what the tool would do in other school systems.
The work brought together Cornell information science researchers, Microsoft Research India and the Sikshana Foundation, a nonprofit working with government schools. Microsoft supported the tool in collaboration with the AI Innovation and Inclusion Initiative at the International Institute of Information Technology Bangalore and the foundation. The research team says the Telangana expansion will test the system with a larger group of teachers. The Cornell Chronicle said the study received an honorable mention and will be presented at the ACM Computer-Supported Cooperative Work and Social Computing conference in Salt Lake City, Oct. 10-14.
FingerLakes1.com is the region’s leading all-digital news publication. The company was founded in 1998 and has been keeping residents informed for more than two decades. Have a lead? Send it to news@fingerlakes1.com.
