Duolingo and the Future of Personalized Education with AI

Bozena Pajak, VP of Learning and Curriculum at Duolingo, discusses how AI has evolved from personalized difficulty models to generative AI characters enabling conversational language practice. The episode covers how AI addresses speaking anxiety—a primary obstacle in language learning—and explores agentic workflows for content scaling and the shift toward thematic personalization.

Conversational AI Addresses Emotional Barriers.

Duolingo discovered that "speaking anxiety" is a primary obstacle to language learning, not insufficient content. AI-powered video call characters provide a safe, low-pressure practice environment where learners feel comfortable making mistakes.

Domain Experts Must Evolve Into AI Orchestrators.

Rather than replacing linguists and learning scientists, Duolingo retrained them as prompt engineers and workflow designers. These experts remain essential for creating pedagogical rubrics and evaluators that ensure AI outputs meet rigorous educational standards.

Assessment Technology Drives Internal Learning.

The Duolingo English Test—initially dismissed but now accepted globally—provided assessment technology that feeds back into the learning app itself, closing the gap between testing and instruction.

Scientific Credibility Requires Patient Trust-Building.

When Pajak joined Duolingo as its first learning scientist, earning organizational trust took approximately two years of demonstrating that research insights directly improved engagement metrics and user retention.

Prescribed Paths Outperform User Choice.

Shifting from flexible tree structures to linear learning paths unexpectedly improved outcomes. Users prefer clear recommendations over autonomy, reducing cognitive load while ensuring pedagogically sound progression.

Long-Term Retention Trumps Immediate Accuracy.

Short-term interface improvements increasing accuracy can harm long-term retention. Effective learning measurement requires months-long studies rather than quick engagement signals.

Features Serve Dual Purposes: Engagement and Assessment.

AI video conversations function both as engaging practice and stealth assessment tools, providing rapid proficiency signals previously requiring expensive, slow research studies.

Personalization Shifts From Difficulty to Thematic Lenses.

Future personalization will tailor content themes rather than just difficulty levels—allowing learners to study identical grammar through sports, art, or other interest-based "lenses."

You can read the full transcript here.