Back office ops · Production

Duolingo uses large language models to accelerate language lesson creation

The problem

Duolingo has fewer than 1,000 employees but over 21 million daily active users, and building, updating, and maintaining course content takes so long that many courses can only release new content a few times per year.

First attempt

The previous Birdbrain AI could only select practice problems, not create them; human experts had to write each problem individually, then review, correct, and translate them — a slow per-exercise process.

Workflow diagram · grounded in source
1
Curriculum design
trigger
“カリキュラム設計者が、レッスンのテーマ、文法、語彙、練習問題の種類などを計画します”
2
Prompt preparation
integration
“プロンプトの一部はエンジニアの助けによって、プログラミングで自動的に書き出されます(言語、CEFRレベル、懐かしい思い出のテーマなど)。これに対しカリキュラム設計者は、練習問題の種類や文法の焦点などの必要項目をいくつか追加し、AIを正しい方向に導きます”
3
AI exercise generation
ai_action
“AIモデルが数秒のうちに、難易度、文法、テーマに合った10個の練習問題を出力します”
4
Curriculum designer review
human_review
“カリキュラム設計者はこの中から気に入った文章を3つ選びますが、アプリに入れる前に手作業で編集することもできます。その場合、カリキュラム設計者が自然さ、学習上の価値、語彙の適切さを考慮したうえで、調整を行います”
5
Expert final approval
validation
“最後にDuolingoの英語教育専門家が、最終的な判断を下します”
Reported outcome

AI tools allow Duolingo's education experts to generate large amounts of lesson content with a button click, enabling them to work more conveniently, faster, and more productively while maintaining content quality.

Reported metrics
Daily active users2,100万人を超えています
Employee count1,000人未満
exercises generated per AI run10個
Exercise generation time数秒のうちに
Show all 5 reported metrics
daily active users2,100万人を超えています
employee count1,000人未満
exercises generated per AI run10個
exercise generation time数秒のうちに
educator productivityより便利に、より速く、そしてより生産的に仕事ができるようになりました
Reported stack
大規模言語モデルバードブレイン
Source
https://blog.duolingo.com/ja/ai-lesson-creation/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

AI tools allow Duolingo's education experts to generate large amounts of lesson content with a button click, enabling them to work more conveniently, faster, and more productively while maintaining content quality.

What tools did this team use?

大規模言語モデル, バードブレイン.

What results were reported?

Daily active users: 2,100万人を超えています; Employee count: 1,000人未満; exercises generated per AI run: 10個; Exercise generation time: 数秒のうちに (source-reported, not independently verified).

What failed first in this deployment?

The previous Birdbrain AI could only select practice problems, not create them; human experts had to write each problem individually, then review, correct, and translate them — a slow per-exercise process.

How is this back office ops AI workflow structured?

Curriculum design → Prompt preparation → AI exercise generation → Curriculum designer review → Expert final approval.