# Building AI That Teaches: Lessons From Nine Years at Coursera

Educational AI requires a distinct engineering approach from consumer AI systems, according to insights from nearly a decade of work at Coursera, the major online learning platform that serves millions globally.

The difference centers on what AI must accomplish in education. Consumer AI optimizes for engagement and retention of individual users. Educational AI must balance learner outcomes, teacher effectiveness, institutional goals, and equity across diverse student populations. These competing demands reshape how systems get built.

Coursera has scaled learning systems to over 100 million learners worldwide. That scale reveals problems invisible at smaller levels. Personalization algorithms that work for affluent students may fail for those with limited internet access. Chatbots that answer questions perfectly in English struggle across 70+ languages Coursera supports. Assessment systems designed for traditional students miss learning patterns in working adults balancing jobs and family.

The engineering challenge extends beyond technical performance metrics. Educational AI must operate within institutional constraints. Universities want AI that works alongside existing curricula, not against them. K-12 schools need systems that comply with data privacy laws while providing actionable insights to teachers. These real-world requirements often conflict with what maximizes technical elegance.

Scaling learning systems also demands different data strategies than consumer platforms. Educational data reveals sensitive information about student struggle points, demographics, and learning disabilities. Systems must protect privacy while remaining useful. This tension between personalization and protection creates engineering constraints few consumer AI systems face.

Teacher involvement changes the engineering entirely. Consumer AI can ignore user feedback at scale. Educational AI fails without it. Teachers spot when recommendations miss cultural context or reinforce biases. They catch when algorithmic suggestions contradict pedagogical goals. Building systems that absorb and act on teacher feedback at scale requires different architecture than consumer products require.

Nine years at Coursera demonstrates that educational AI isn't simply consumer AI applied to