Workflow · Production

Duolingo uses Large Language Models to generate lesson exercises faster at scale

The problem

Duolingo has fewer than 1,000 employees serving over 21 million daily users, and building, updating, and maintaining courses takes considerable time—most courses release new content only a few times a year—limiting how quickly the team can expand the depth and breadth of available learning content.

Workflow diagram · grounded in source
1
Curriculum design
trigger
“The Learning Designer plans out the theme, grammar, vocabulary, and exercise types for a given lesson.”
2
Prompt preparation
human_review
“The Learning Designer fills in a few more blanks, such as the exercise type and the grammar focus, to steer the AI in the right direction.”
3
AI exercise generation
ai_action
“In a matter of seconds, the AI model outputs ten exercises that fit the difficulty level, the grammar, and the theme.”
4
Human selection and editing
human_review
“The Learning Designer then chooses their three favorite sentences, but they can still make edits before they go into the app. In this case, the Learning Designer would probably make some tweaks for naturalness, learning value, and approp…”
Reported outcome

AI-assisted lesson generation allows Duolingo to produce quality content faster, enabling the team to reach more advanced CEFR levels, allocate resources to additional features, and expand to smaller language courses.

Reported metrics
Exercise generation speedin a matter of seconds
Reported stack
Large Language ModelBirdbrain
Source
https://blog.duolingo.com/large-language-model-duolingo-lessons/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

AI-assisted lesson generation allows Duolingo to produce quality content faster, enabling the team to reach more advanced CEFR levels, allocate resources to additional features, and expand to smaller language courses.

What tools did this team use?

Large Language Model, Birdbrain.

What results were reported?

Exercise generation speed: in a matter of seconds (source-reported, not independently verified).

How is this workflow AI workflow structured?

Curriculum design → Prompt preparation → AI exercise generation → Human selection and editing.