Back office ops · Production

Duolingo uses LLMs to speed up language lesson exercise creation

The problem

Creating, updating, and maintaining Duolingo courses takes considerable time, limiting most courses to releasing new content only a few times per year — a constraint for a lean team tasked with delivering world-class education at global scale.

Workflow diagram · grounded in source
1
Curriculum planning
trigger
“학습 디자이너가 해당 레슨에 대해 주제, 문법, 연습 문제 유형을 기획합니다”
2
Prompt configuration
trigger
“일부 규칙은 자동으로 채워져요(언어, CEFR 레벨, "향수 어린 추억"이라는 주제). 학습 디자이너는 AI가 올바른 방향으로 나갈 수 있게 연습 문제 유형과 핵심 문법 등 남은 빈칸을 더 채워 넣어요.”
3
LLM exercise generation
ai_action
“단 몇 초만에 AI 모델은 주어진 난이도, 문법, 주제에 맞는 연습 문제 10개를 생성합니다”
4
Human selection and editing
human_review
“학습 디자이너가 그중에서 가장 좋다고 판단한 문장 2개를 선택하게 되며, 실제로 앱에 들어가기 전에 편집을 할 수 있습니다”
5
Pedagogy expert final approval
validation
“듀오링고의 스페인어 교수법 전문가들이 최종 결정을 할 권한을 가지죠”
Reported outcome

LLMs now generate ten exercise candidates in seconds for a given difficulty, grammar, and topic; learning designers select two and edit them before app publication, with Duolingo's pedagogy experts holding final authority.

Reported metrics
Exercises generated per prompt10
Exercises selected by learning designer per run2
Exercise generation speed단 몇 초만에
Current content release frequency per course1년에 몇 번만
Reported stack
LLMBirdbrain
Source
https://blog.duolingo.com/ko/ai-speeds-up-duolingo-lessons/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

LLMs now generate ten exercise candidates in seconds for a given difficulty, grammar, and topic; learning designers select two and edit them before app publication, with Duolingo's pedagogy experts holding final autho…

What tools did this team use?

LLM, Birdbrain.

What results were reported?

Exercises generated per prompt: 10; Exercises selected by learning designer per run: 2; Exercise generation speed: 단 몇 초만에; Current content release frequency per course: 1년에 몇 번만 (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Curriculum planning → Prompt configuration → LLM exercise generation → Human selection and editing → Pedagogy expert final approval.

Related back office ops cases