# Building AI that teaches: What nine years at Coursera taught me about scaling learning systems
A former Coursera executive argues that educational AI requires different engineering approaches than consumer-facing AI systems. The distinction matters because learning platforms operate under different constraints than social media or search engines.
Educational AI must prioritize pedagogical outcomes over engagement metrics. Consumer AI optimizes for clicks, time spent, and user retention. Learning systems optimize for knowledge acquisition and skill mastery, which demand different measurement frameworks and success criteria.
Coursera's nine-year track record reveals specific design principles for educational AI. These systems must adapt to diverse learner backgrounds, learning speeds, and prior knowledge levels. They must provide meaningful feedback that guides improvement rather than just flagging errors. They must scale affordably across global student populations without degrading quality.
The infrastructure challenges differ substantially. Consumer AI platforms prioritize real-time responsiveness and personalization at massive scale. Educational AI must maintain consistent rigor, ensure assessments remain valid as systems evolve, and produce transparent reasoning for grading decisions that students and instructors can understand and contest.
Data challenges also diverge. Consumer AI trains on billions of user interactions. Educational AI works with smaller datasets but requires higher-quality annotations. A single incorrect assessment in an educational system can mislead students about their actual competency levels, whereas consumer AI errors rarely carry such stakes.
Scaling learning systems requires different partnerships too. Building educational AI means collaborating with instructors, curriculum designers, and learning scientists, not just engineers and data scientists. Coursera's experience shows that AI tools succeeding in education gained institutional buy-in from educators first, technical implementation second.
This distinction explains why some AI companies struggle when entering education. Off-the-shelf large language models optimize for fluency and breadth. Educational AI must optimize for pedagogical alignment and measurable learning gains. The engineering discipline, the success metrics, and the