Shiksha Copilot Drafts Lesson Plans for Karnataka Teachers

What Happened

A team from Cornell University and Microsoft Research India has published the first large evaluation of Shiksha Copilot, an AI system that turns textbook chapters into ready-to-teach lesson plans for public school teachers in the Indian state of Karnataka. The paper, "Shiksha Copilot: Teacher-AI Collaboration for Curating and Customizing Lesson Plans in Low-Resource Schools," was presented at the ACM Computer-Supported Cooperative Work and Social Computing conference in Salt Lake City, which runs October 10 to 14, and it earned an Honorable Mention. Cornell described the work in a Cornell Chronicle story on October 8.

The tool is not a single chatbot. It is a collection of models that take material from a grade-level textbook, draft a lesson plan in English, and then translate it into Kannada, the language most of the students speak. Guardrails push the output toward examples that make sense locally. The release gives a simple illustration: a discussion of popular foods should reach for a dosa rather than a pizza. Human review sits on both sides of the machine. Curators check each plan for accuracy before it reaches classrooms, and teachers then adapt it for their own students.

The evaluation drew on more than 1,000 teachers of grades 5 through 10, using curator reviews, interviews and surveys. The first author is Deepak Varuvel Dennison, a doctoral student at Cornell, working with Aditya Vashistha and Rene Kizilcec of Cornell and Tanuja Ganu and colleagues at Microsoft Research India. The local nonprofit Sikshana Foundation, which works with government schools, was a partner, and the project was supported by the AI Innovation and Inclusion Initiative at IIIT Bangalore. The system was originally developed by Microsoft Research India, and its name comes from the Sanskrit word for learning.

Why It Matters

The numbers in the study show a tool that is strong in one language and shaky in another. Curators reviewed 7,744 English lesson plans. Only 3.53 percent needed minor adjustments, and just 0.34 percent were rejected for technical errors. Kannada was a different story. About 96 percent of the translated plans needed changes, mostly to terminology: curators swapped in "mirrors" where the software had written "glasses," and "timeline" where it had used "chronology." Even so, teachers rated almost 81 percent of the Kannada plans "very good," which suggests that the edits were the kind a working teacher can live with rather than a reason to abandon the tool.

That gap is the real lesson. Language models are trained on far more English than Kannada, and a system that looks polished in a demo can lean heavily on human editors once it meets a regional language. Shiksha Copilot was built around that reality instead of hiding it, with curators in the loop from the start. The team's wider research points the same direction. A companion framework paper, presented at the CHI conference in Barcelona in April, examined Shiksha Copilot alongside seven other AI systems built for social good in different countries. The authors found that adapting an application to local systems and needs can account for up to 80 percent of the development effort.

For education systems with large classes and stretched staff, the potential payoff is time. Teachers in the study said the tool saved them an average of two hours per week, and that it reduced paperwork such as submitting hard copies of lesson plans. In a profession where planning often happens at night and on weekends, two hours is not a small number, though it is worth remembering that it is a self-reported figure rather than a measured one.

Reaction

The reactions on record come from the teachers themselves, and they are largely positive. A science teacher with 20 years of experience called it "a burden-releasing concept," a notable phrase in a field where new technology has often meant more work, not less. Teachers also described a shift in how they run lessons. With a draft plan in hand, they said, they spent less time lecturing and more time on hands-on activities, such as sending students out to collect local leaves and compare their vein patterns.

It is worth being careful about what this feedback can and cannot show. The coverage does not describe a controlled comparison, a sampling method or statistical testing, and it does not report any change in student learning. Time savings and satisfaction come from surveys and interviews. No independent educators or outside researchers are quoted, so the picture is the one the study team and its partners present. That does not make the findings wrong, but it does mean the strongest claim available is that teachers like the tool and say it helps them, not that students learn more because of it.

What's Next

The next test is scale. About 30,000 teachers across the neighboring state of Telangana are using the tool this year, according to Cornell, which will be a much bigger and messier trial than a research evaluation in one state. The rollout will show whether the curator model, which depends on trained reviewers checking plans before they go out, can keep up as the number of teachers grows by a factor of thirty. It will also show whether the Kannada terminology problems are fixable through better models and glossaries, or whether they remain a permanent editing burden.

The researchers say they want to continue building culturally aware AI for education, healthcare, law and agriculture in the Global South. The honest question for that agenda is cost. If tailoring a tool to one place consumes most of the development budget, then every new language and region starts nearly from scratch. The paper and its companion study make that cost visible, which is useful for funders and governments deciding where to spend. Details of the work are also available in the Microsoft Research publication page, and the original release was republished by the National Tribune.

Closing Thoughts

It is easy to describe AI in the classroom as either a time machine or a threat. Shiksha Copilot is neither. It is a drafting assistant that works best when a person who knows the subject and the language stands on either side of it, and the study is candid that the Kannada half still needs plenty of that person's attention. The 96 percent edit rate and the 81 percent "very good" rating can both be true, and together they describe what useful AI often looks like in practice: rough in places, genuinely helpful, and dependent on people who know their context.

The most interesting finding may be the one about local effort. Technology tends to be designed for the places where it is cheapest to build and then exported. This work suggests that the hard part is not the model but the fit, the dosa instead of the pizza, the right word for mirrors, the curator who catches the mistake. If the Telangana rollout holds up, it will be a sign that careful fit can scale. If it does not, the 80 percent figure will explain why.

한글 요약

코넬대학교와 마이크로소프트 리서치 인디아 연구팀이 인도 카르나타카주 공립학교 교사를 위한 AI 수업 설계 도구 '시크샤 코파일럿(Shiksha Copilot)'의 대규모 평가 결과를 ACM CSCW 학회(10월 10~14일, 솔트레이크시티)에서 발표해 우수상(Honorable Mention)을 받았습니다. 이 도구는 교과서 내용을 학년에 맞는 영어 수업안으로 만들고 현지어인 칸나다어로 번역하며, 피자 대신 도사처럼 현지에 맞는 예시를 쓰도록 안전장치를 두었습니다. 평가에는 5~10학년 교사 1,000여 명이 참여했습니다.

영어 수업안 7,744건 중 경미한 수정이 필요한 비율은 3.53%, 기술 오류로 폐기된 비율은 0.34%였습니다. 반면 칸나다어 수업안은 약 96%가 주로 용어 문제로 수정이 필요했지만, 교사들은 거의 81%를 '매우 좋음'으로 평가했습니다. 교사들은 주당 평균 2시간을 절약했고 강의 중심에서 체험 활동 중심 수업으로 옮겨 갔다고 답했습니다. 다만 이 수치는 설문과 인터뷰에 근거한 자기 보고이며, 학생 학습 성과는 보고되지 않았고 대조군 비교도 공개되지 않았습니다.

올해 인접한 텔랑가나주에서 약 3만 명의 교사가 이 도구를 사용할 예정이어서, 검수 인력에 의존하는 방식이 규모를 감당할 수 있을지가 다음 시험대입니다. 연구팀의 관련 연구에 따르면 지역 현실에 맞추는 작업이 개발 노력의 최대 80%를 차지할 수 있습니다. 참고: Cornell Chronicle, Microsoft Research, The National Tribune.