The situation
Early on, we built a feature that would read a student's written answer and say whether it was right. It is an obvious feature. Every parent asks for it.
The problem
When the AI marks an answer wrong and the AI is itself wrong, the child does not know which to believe. Our own test numbers, 74.5% in English and 58.3% in Marathi in one test of whether replies included a key fact, told us the AI is not yet right often enough to judge a ten-year-old's work.
What we tried
We tried softer wording, such as “this might need another look”. Students read it as a verdict anyway. We tried limiting it to simple definitions. The failures were fewer but no less confusing.
What we learned
A feature that is right most of the time is not good enough when being wrong undermines a child's confidence. Marking an answer has to be right more often than explaining one.
What it changed
Automatic answer checking and practice questions are turned off in the regular student app. They stay on our list of planned work, clearly marked as needing more testing. The privacy and care page says it in one sentence: VIDYA AI does not let its current AI decide that a child's answer is wrong.
Takeaway
If you see a study app that marks children's answers, ask how often the marking itself is wrong. If the answer is “we don't know”, that is the answer.
The numbers here come from our release reports and one reported phone test. They do not show any change in marks or exam results.
