Episode

Duolingo and the Future of Personalized Education with AI

Bozena Pajak, VP of Learning at Duolingo, joins High Signal to discuss the evolution of AI at Duolingo: from personalized difficulty models to the current generative frontier where AI characters provide low-stakes and high impact conversational practice. We discuss the role of AI in overcoming one of the biggest hurdles in language acquisition, speaking anxiety. We also talk about how Bozena's team leverages agentic workflows to scale content and why the next wave of personalization involves shifting from difficulty levels to "thematic lenses" tailored to specific user interests.

February 10, 2026

Listen on

Guest

Bozena Pajak

Duolingo

Bozena Pajak is the Vice President of Learning and Curriculum at Duolingo. She founded and leads the Learning and Curriculum team, a group of Learning Designers, Learning Scientists, and Creative Content Designers. Together, they work on ensuring that Duolingo delivers on the promise of effective education by bringing together insights from learning sciences and the Duolingo data from millions of users to understand how we learn, and then using those insights to build experiences that make learning more effective, efficient, and motivating.

‍

Guest

HOST

Hugo Bowne-Anderson

Delphina

Hugo Bowne-Anderson is an independent data and AI consultant with extensive experience in the tech industry. He is the host of the industry podcast Vanishing Gradients, a podcast exploring developments in data science and AI. Previously, Hugo served as Head of Developer Relations at Outerbounds and held roles at Coiled and DataCamp, where his work in data science education reached over 3 million learners. He has taught at Yale University, Cold Spring Harbor Laboratory, and conferences like SciPy and PyCon, and is a passionate advocate for democratizing data skills and open-source tools.

‍

Key Quotes

Key Takeaways

Conversational AI solves for anxiety, not just information.

Duolingo found that a key barrier to language acquisition isn't a lack of content, but "speaking anxiety." By using generative AI characters for video calls, they created a low-pressure environment where users feel safe making mistakes, proving that AI’s greatest value in education may be emotional cushioning rather than just data delivery.

‍

Domain experts must become AI orchestrators.

Rather than replacing learning scientists with engineers, Duolingo retrained their linguists to become prompt engineers and agentic workflow designers. Domain experts are the only ones capable of building the specific rubrics and evaluators needed to ensure generative AI outputs meet rigorous pedagogical standards.

‍

A controversial test became Duolingo's assessment backbone.

The Duolingo English Test (a high-stakes, AI-based proficiency exam for university admissions and job certification) was initially dismissed by the established assessment community. Over time, rigorous validation won acceptance from universities worldwide. Now, the assessment technology developed for that test is being applied back inside the learning app itself, closing the loop between testing and teaching.

‍

The "fourth leg" requires a multi-year trust cycle.

When Pajak joined as the first learning scientist, it took two years of proposing ideas and proving wins in engagement metrics before the engineering-heavy culture fully integrated the role. Scientific expertise in a tech company is only effective once leaders demonstrate that research insights translate directly into product performance and user retention.

‍

Deterministic paths outperform learner autonomy.

Transitioning from a flexible "tree" structure to a linear "path" unexpectedly improved learning outcomes. While users often think they want choice, providing a single, strong recommendation on what to do next reduces cognitive load and ensures they follow a scientifically optimized curriculum.

‍

Short-term accuracy is a vanity metric for learning.

Learning is non-linear; an interface change that increases a user’s immediate accuracy can actually lead to worse long-term retention. Because real learning outcomes often take months to manifest, AI leaders must balance fast engagement signals with longitudinal research to avoid optimizing for "shallow" success.

‍

Your engagement feature is also your best measurement instrument.

Duolingo is using Generative AI characters in video calls not just for conversational practice but as a stealth assessment tool, analyzing short conversations to estimate a user's speaking proficiency without a formal test. This dual-use approach, where a single feature serves both engagement and measurement, gives the team a fast learning signal that previously required slow, expensive research studies.

‍

Personalization is moving from "Level" to "Lens."

The next frontier of personalization isn't just adjusting the difficulty of an exercise, but changing the thematic content itself. AI will soon enable "thematic personalization," where two users learn the same grammar point through entirely different lenses (one through sports and another through art) based on their individual interests.

‍

You can read the full transcript here.

‍