Karnataka’s AI lesson planner had two jobs: paperwork and teaching

At Government Higher Primary School in Basavanahalli, Mahalakshmi Ashok found a way to move a lesson about the heart off the blackboard. Her pupils felt for their pulses, counted beats and compared what they found. The activity came from Shiksha Copilot, according to Microsoft’s reporting from an early Karnataka pilot, published in February 2024.[3]

An AI suggestion for a blood-typing laboratory was unsuitable. Measuring a pulse was possible. That choice—what these children could do with the available materials—was the teacher’s. The software supplied possibilities; Mahalakshmi decided which one became a lesson.[3]

A larger deployment subsequently offered Shiksha Copilot to 1,043 teachers in 757 Karnataka government schools between December 2024 and March 2025. Among 406 teachers with matched before-and-after surveys, self-reported lesson-planning time fell by 2.02 hours a week on average. The study combined surveys, interviews and usage logs; it had no untreated comparison group and did not measure pupil learning.[1]

The research reached the CSCW conference in October 2026. Developed by Microsoft Research India with Sikshana Foundation and evaluated with Cornell, the system generates curriculum-aligned English plans and translates them into Kannada. Curators review the content before classroom teachers adapt it. Regional examples are deliberate: a food illustration might feature a dosa, a familiar local food.[2]

The lesson on record

Teachers had to submit plans in Karnataka’s prescribed five-stage format: engage, explore, explain, elaborate and evaluate. Some authorities accepted printouts; others insisted on handwriting. Shiksha made compliant text easier to obtain. Experienced teachers told researchers that their submitted plans captured only a fraction of what they taught. They often adapted material mentally, without recording those changes in the software.[1]

A document can therefore be finished while preparation continues. A paragraph describing an activity still leaves decisions about grouping pupils, explaining the task and recognising when someone needs help. The valuable saving is time a teacher can use for those decisions.

The constraints were tangible. One teacher in the study described three teachers handling seven classes while a headteacher was away for training. Teachers reported shorter planning, but the software could not cover an understaffed timetable or remove unrelated administrative data entry.[1]

An idea that fits the room

Ravindra K. Nagaiah, another teacher in the early pilot, described requesting a different activity when a suggested one required materials he lacked. He could ask for a shorter video, too. The teacher supplied the classroom constraint, and the system returned another option.[3]

In the later Cornell account, teachers described collecting local leaves with different vein patterns. Such an activity needs no screen on every desk. The AI operates during preparation; the pupils can spend the lesson handling and comparing things. The Kannada materials required substantial language editing, adding another layer of local expertise before use.[2]

The programme is expanding. Sikshana says training for district resource people began in Telangana on 3 July 2026, using a cascading model in which those educators help train others. Cornell reported in October that about 30,000 teachers in Telangana were using the tool, in a larger deployment being studied.[4][2]

At the larger scale, faster paperwork is an achievement teachers can feel immediately. The harder question is whether the time and local knowledge released during preparation reach the children. Mahalakshmi’s pulse-counting lesson suggests what that second achievement could look like: an idea becomes something a pupil can try.

Report an error or suggest a correction