Duolingo's AI Revolution
What 148 AI-Generated Courses Tell Us About the Future of Instructional Design & Human Learning
Hey folks!
Last week, Duolingo announced an unprecedented expansion: 148 new language courses created using generative AI, effectively doubling their content library in just one year. This represents a seismic shift in how learning content is created — a process that previously took the company 12 years for their first 100 courses.
As CEO Luis von Ahn stated in the announcement, "This is a great example of how generative AI can directly benefit our learners... allowing us to scale at unprecedented speed and quality."
But beneath the impressive numbers and corporate PR lies a fascinating story about the future of instructional design itself. Duolingo's Birdbrain AI system has fundamentally transformed each phase of the classic ADDIE framework, offering a glimpse into the evolving relationship between human expertise and artificial intelligence in learning design.
In this week's blog, I'll dissect exactly how Duolingo has reimagined instructional design through AI, what this means for the learner experience, and most importantly, what it tells us about the future of our profession.
TLDR:
Duolingo's AI-powered expansion reveals three critical shifts in instructional design:
Role transformation: Human IDs are evolving from content creators to AI orchestrators, prompt engineers, and quality guardians
Process acceleration: What once took years now takes months, with AI handling scale while humans ensure quality
Evaluation revolution: Continuous, real-time assessment is replacing traditional episodic evaluation
Let's dive in! 🚀
Analysis: From Intuition to Data-Driven Needs Assessment
Traditional Approach
In conventional instructional design, the Analysis phase typically involves subject matter expert interviews, literature reviews, and perhaps some focus groups to identify learning needs. It's largely qualitative, relies heavily on professional intuition, and often takes weeks or months to complete.
Duolingo's AI-Powered Approach
Duolingo's Birdbrain AI processes a staggering 1.25 billion daily exercises, using sophisticated data science to identify precise learning needs:
Statistical pattern recognition: Birdbrain detects when large percentages of learners struggle with particular concepts (e.g., Spanish preterite vs. imperfect tense) by analysing error rates across millions of users.
Predictive modelling: The system uses logistic regression to estimate the probability of a learner getting an exercise correct, creating a "difficulty score" for each concept.
Real-time pain point identification: Rather than waiting weeks for feedback, Birdbrain identifies learning obstacles as they emerge.
Example: When analysing Spanish language learners, Birdbrain can identify systematic error patterns with tense usage, flagging concepts that consistently trip up users and assigning difficulty scores to guide content development and sequencing.
Human ID Role in AI-Powered Analysis
While AI excels at identifying what learners struggle with, human instructional designers remain essential for understanding why:
Experimental design: IDs design A/B tests to investigate learning obstacles and determine effective interventions
Cultural context: IDs interpret whether confusion stems from linguistic features or cultural differences
Hypothesis formation: Based on data patterns, IDs formulate theories about learning challenges that guide intervention design
Design: From Linear Planning to Dynamic Curriculum Architecture
Traditional Approach
Conventional instructional design involves sequential planning, where learning objectives are mapped to content in a largely fixed structure. The process is manual, time-intensive, and typically produces a static learning path.
Duolingo's AI-Powered Approach
Duolingo has reimagined design as a dynamic, generative process:
Prompt-based content framework: Instead of manually designing each lesson, IDs create structured prompts for AI to generate appropriate content. As described in Duolingo's blog, these prompts function like "Mad Libs" for lesson generation, with specific parameters for language level, grammar focus, and more.
Shared-content pipeline: A base course framework is created once, then automatically localised to dozens of languages using large language models, dramatically accelerating the development of new courses.
Adaptive sequencing: Duolingo uses algorithms to balance introducing new concepts with reinforcing existing knowledge, targeting an optimal challenge level for learners.
Example: For new language courses, Duolingo's learning designers fill in key parameters (like language, CEFR level, and grammar focus), and the AI generates multiple exercises that fit those specifications in seconds.
Human ID Role in AI-Powered Design
Human instructional designers have shifted from content creation to:
Prompt engineering: Crafting precise instructions for AI to generate appropriate content
Cultural auditing: Reviewing AI outputs for dialectical appropriateness and cultural sensitivity
Constraint definition: Setting parameters to ensure AI-generated content aligns with learning levels and objectives
As Jessie Becker, Duolingo's Senior Director of Learning Design, noted: "Now, by using generative AI to create and validate content, we're able to focus our expertise where it's most impactful, ensuring every course meets Duolingo's rigorous quality standards."
Development: From Manual Creation to AI-Generated Content with Human QA
Traditional Approach
In conventional ID, development involves creating all learning materials from scratch—writing content, designing activities, and producing media. This phase is typically the most time-consuming.
Duolingo's AI-Powered Approach
Duolingo has transformed development into a human-AI collaboration:
Mass-generation: AI creates multiple exercise variants based on instructional prompts, dramatically accelerating content production.
Automated filtering: Birdbrain evaluates each exercise using difficulty scores and quality metrics, rejecting content that doesn't meet standards.
Multi-stage review: AI content passes through algorithmic gates before reaching human review.
Example: As demonstrated in Duolingo's own blog posts, their AI might generate ten different exercises focused on a specific grammar point like past tense usage. Learning designers then select the best options and make refinements for naturalness and learning value before implementing them in the app.
Human ID Role in AI-Powered Development
Human expertise becomes focused on edge cases and quality:
Selectivity and refinement: Choosing the best AI-generated options and making necessary adjustments
Quality assurance: Ensuring generated content meets linguistic and pedagogical standards
Naturalness checks: Making sure language sounds authentic rather than stilted or unnatural
The collaboration allows humans to focus on high-value judgment calls while AI handles volume and repetitive tasks.
Implementation: From Static Delivery to Real-Time Adaptive Learning
Traditional Approach
Traditional implementation often means simply publishing content and hoping learners engage with it as designed. Adaptations typically occur only in future iterations.
Duolingo's AI-Powered Approach
Implementation at Duolingo is a dynamic, continuously adapting process:
Real-time personalisation: Birdbrain adjusts exercise difficulty on the fly based on learner performance, ensuring users are consistently working at the appropriate challenge level.
Technical optimisation: Duolingo's rewrite of their Session Generator in Scala reduced exercise delivery time from 750ms to 14ms, enabling truly real-time personalisation.
Engagement optimisation: AI continuously tests and adjusts learning paths to maximise both learning outcomes and user engagement.
Example: Birdbrain can dynamically adjust the difficulty of exercises based on a learner's performance, providing easier exercises when they're struggling and more challenging content when they're succeeding.
Human ID Role in AI-Powered Implementation
Humans shift to systems-level oversight:
Monitoring engagement dashboards: Tracking metrics to identify problematic content
Adjusting gamification elements: Fine-tuning reward structures to maintain motivation without compromising learning
Ethical guardrails: Ensuring algorithms don't prioritise engagement at the expense of learning outcomes
Evaluation: From Episodic Testing to Continuous Improvement Loops
Traditional Approach
Evaluation in conventional ID often happens after implementation, with periodic assessments and course revisions based on limited data points.
Duolingo's AI-Powered Approach
Evaluation at Duolingo is continuous and deeply data-informed:
Comprehensive metrics: Birdbrain tracks numerous learning metrics across millions of users, allowing for unprecedented insight into performance patterns.
Scientifically-informed review: The system uses principles like spaced repetition to optimize review schedules for long-term retention.
Bias monitoring: Regular audits help identify and correct biases in content.
Example: By analysing patterns in user errors and retention across languages, Duolingo can identify concepts that require more frequent review in specific languages and automatically adjust the curriculum accordingly.
Human ID Role in AI-Powered Evaluation
Humans provide crucial interpretive and ethical dimensions:
Interpreting efficacy studies: Translating research findings into curriculum adjustments
Balancing metrics: Ensuring that optimization for one metric doesn't undermine others
Ethical oversight: Preventing over-optimization for engagement that might harm learning outcomes
Key Implications for Instructional Designers
What does Duolingo's AI-driven transformation tell us about the future of instructional design? Three clear implications emerge:
1. Role Specialisation is Accelerating
As I predicted in my April post on instructional design specialisation, Duolingo's approach confirms that the general "full stack" ID role is rapidly fracturing into specialised functions:
AI Orchestrators: Experts in crafting prompts and managing AI systems
Quality Guardians: Specialists who review and refine AI-generated content
Learning Scientists: Professionals who design experiments and interpret learning data
Cultural Context Specialists: Experts who ensure content is culturally appropriate and effective
The days of the instructional designer who handles everything from analysis to evaluation are numbered. Instead, we're seeing the emergence of specialized roles that reflect both the capabilities of AI and the areas where human expertise remains essential.
2. Speed & Scale Are the New Normal
Duolingo's achievement of creating 148 courses in one year (versus 12 years for their first 100) represents a fundamental shift in expectations. As Luis von Ahn noted, "We owe it to our learners to get them this content ASAP."
Instructional designers must adapt to this new reality by:
Developing skills in prompt engineering and AI management
Focusing on high-level strategy rather than content production
Building expertise in rapid experimentation and data interpretation
The organisations that thrive will be those that embrace AI for scale while deploying human expertise strategically.
3. Continuous Evaluation Replaces Episodic Assessment
Perhaps the most profound shift is from point-in-time assessment to continuous, real-time evaluation. Duolingo's billion-plus daily data points enable a level of insight previously impossible.
Future-focused instructional designers will need to:
Develop data literacy to interpret learning analytics
Design for continuous assessment rather than fixed checkpoints
Build feedback loops into every aspect of the learning experience
Conclusion: The Future of Instructional Design is Human-AI Collaboration
Duolingo's AI revolution doesn't signal the death of instructional design—it marks the beginning of a new, more specialised and powerful era for our profession.
As we've seen through each phase of ADDIE, AI excels at processing massive datasets, generating content at scale, and personalising experiences in real-time. Yet human expertise remains essential for understanding context, making ethical judgments, and providing the creativity and empathy that bring learning to life.
The most successful instructional designers of the future won't be those who resist AI, nor those who blindly embrace it, but those who learn to collaborate with it effectively—leveraging its strengths while applying uniquely human skills to areas where AI falls short.
In this new landscape, we have an unprecedented opportunity to impact learning at scale. By embracing specialised roles, developing AI literacy, and focusing on high-value human contributions, we can create learning experiences that are more personalised, effective, and accessible than ever before.
As von Ahn put it, "AI isn't just a productivity boost. It helps us get closer to our mission. To teach well, we need to create a massive amount of content, and doing that manually doesn't scale." This balance—scale through technology, quality through humanity—represents the future of our field.
Happy innovating!
Phil 👋
PS: Want to develop the AI skills needed for this new instructional design landscape? Apply for a place on my AI & Learning Design Bootcamp where we explore these concepts and build practical skills for the AI-powered future of learning.