High Signal: Data Science | Career | AI Episode 34: Duolingo and the Future of Personalized Education with AI
Feb 10, 2026
Bozena Pajak, VP of Learning at Duolingo and former learning scientist in linguistics and cognitive science, discusses AI’s evolution at Duolingo. She covers personalized difficulty models, generative AI characters for low-stakes speaking practice, agentic workflows to scale content, and a shift toward thematic lenses for deeper personalization.
AI Snips
Chapters
Transcript
Episode notes
Personalization Built On BirdBrain
- Duolingo's early AI focused on personalization of difficulty using a student model called BirdBrain.
- BirdBrain predicts per-user accuracy to adapt content and keep challenge levels appropriate.
Generative AI Scales Content And Feedback
- Generative AI unlocked scalable content and interactive features like conversational bots.
- AI also enabled personalized, flexible feedback instead of hard-coded mistake messages.
Learning Requires Longitudinal Measurement
- Measuring learning needs longer windows and independent assessments beyond short-term engagement metrics.
- Duolingo runs longitudinal pre/post studies and external tests to validate real proficiency gains.