E-Paper
AI in classes should generate questions, not answers: Anthropic’s Elizabeth Kelly
AI's success in classrooms hinges on enhancing learning, not just use. Anthropic's tools focus on generating questions, fostering deeper understanding among Indian students.
When AI enters a classroom, the measure of success is not whether students are using it but whether they are learning while they do, said Elizabeth Kelly, who leads beneficial deployment work at Anthropic. Her team’s approach in India, she told Hindustan Times at the AI Impact Summit, is to build tools that generate questions rather than answers — requiring students to do the cognitive work before any AI feedback begins.
The emphasis is deliberate, because Anthropic’s own data suggests the default runs the other way. The company’s India country brief, released days before the summit, found that Indian users delegate more decision-making autonomy to AI than the global average, that a fifth of Indian Claude use is coursework, and that students ask the model for direct answers nearly half the time — bypassing the reasoning that builds understanding.
The company announced a tool built with Pratham. The “anytime testing tool,” powered by Claude, generates curriculum-appropriate questions. Students handwrite answers, upload them, and receive feedback that adapts to their pace of learning. It is now being deployed across 1,500 students, with plans to reach 100 schools by year-end and 5,000 women learners through Pratham’s Second Chance programme.
The need, Kelly said, stems from a basic classroom reality. “There are often classes of 60 students at many different grade levels. Teachers are simply not able to give real-time feedback to all of those students at the same time.”
But designing one tool against the grain of user behaviour does not resolve the broader pattern. Anthropic’s country brief establishes that the default use of AI in Indian education already leans toward delegation rather than augmentation — and that pattern is scaling into a policy environment where the government plans to integrate AI into curricula from the third grade onward.
Asked whether Anthropic is developing evidence frameworks to distinguish engagement from impact — a challenge posed at the summit by economic researchers from J-PAL who warned of an “engagement trap” where high usage metrics can mask a failure to deliver real-world outcomes — Kelly said her team measures success by impact, not adoption. “That’s not just how many people are using our products. It’s where we’re seeing measurable improvements in educational outcomes and in public health,” she said, adding that partnerships are structured around organisations who are “actually testing and iterating to improve the product in order to increase the impact.”
Kelly’s team is also investing in making Claude work across Indian languages — a prerequisite for any claim to beneficial deployment at scale. Anthropic announced a collaboration with Karya, a Bengaluru-based data cooperative, to build training datasets and evaluations in health and agriculture for Indic languages. “Unless these languages are part of the data set, work well for the models, and reflect local context, we’re not actually going to achieve our goal as a Public Benefit Corporation,” Kelly said.
