Duolingo

Angel Ortmann Lee argues that human-in-the-loop AI can fail when people defer to AI output instead of examining it. In a Duolingo English Test study, proctors accepted half of deliberately false cheating signals; clearer review guidance increased rejection of those signals. She shows how interface design can encourage independent judgment, produce richer feedback, and improve AI systems over time. Her design principles include matching review friction to the stakes, surfacing assumptions, structuring interactions, and treating each interaction as a potential label.

Cem Kansu explains how Duolingo makes learning engaging through lessons, streaks, leaderboards, character, and playful notifications, and how these features support retention and growth. He describes the company’s freemium model, its gradual experiments with ads, in-app purchases, and subscriptions, and its approach to expanding into new markets. He argues that translation tools do not replace the motivation and practice involved in learning a language, while generative AI helps Duolingo create course content faster and power speaking practice with its Video Call feature. He also discusses balancing A/B testing with quality standards and long-term thinking, then shares advice for product managers on design, business impact, and learning through practice.