AI Usage at a Glance
Timeline
Aug 1, 2016
RecommendationPractice documented: Duolingo uses an AI algorithm called Half-Life Regression to predict when a learner is about to forget a word or skill, then schedules a review exercise just before that happens. It works like a smart calendar for your memory — timing reminders to keep knowledge fresh without over-drilling what you already know. The system was published as a peer-reviewed research paper in 2016 and has been in production use since.
Practice DocumentedView practice →May 12, 2017
Creative GenPractice documented: Every voice learners hear in the Duolingo app — characters speaking sentences, words, and dialogues — is produced by AI voice synthesis, not live human recordings. Duolingo built custom synthetic voices for each of its animated characters by training AI on recordings made by human voice actors. Those voices now generate unlimited audio across 30+ languages automatically.
Practice DocumentedView practice →Apr 1, 2020
Data AnalysisPractice documented: The Duolingo English Test is the first major standardized language proficiency exam where AI handles every stage of the process — writing the test questions, calibrating their difficulty, choosing which questions each test-taker sees, grading answers, and producing final scores. The test costs $65, takes about an hour, and is accepted by more than 5,000 universities worldwide.
Practice DocumentedView practice →Apr 30, 2020
Data AnalysisNew evidence: The Duolingo English Test: AI-driven language assessment
Evidence AddedView practice →Oct 7, 2020
RecommendationPractice documented: Every Duolingo lesson is shaped by an AI system called Birdbrain, which constantly estimates how well a learner knows each skill and selects the 14 or so exercises most useful for that specific person at that moment. It works like a coach who watches you practice and adjusts every drill based on what you're struggling with — updating within minutes of each completed exercise.
Practice DocumentedView practice →Sep 14, 2022
RecommendationNew evidence: At Duolingo, humans and AI work together to create a high-quality learning experience
Evidence AddedView practice →Feb 23, 2023
Creative GenNew evidence: Duolingo makes learning language fun with help from AI — #LaunchWithAI
Evidence AddedView practice →Mar 14, 2023
Creative GenPractice documented: Duolingo uses a large language model from OpenAI (GPT-4) to power live conversation practice inside its app. Learners can text-chat with AI characters in simulated real-world scenarios, or video-call an animated AI character named Lily who speaks and responds in real time. These features launched in 2023 and are available to paid subscribers.
Practice DocumentedView practice →Jun 22, 2023
Creative GenPractice documented: Duolingo uses generative AI to write lesson exercises, dialogues, and entire language courses. Human experts set the rules — the target language, skill level, and grammar topic — and the AI produces the content for human review. In April 2025, Duolingo launched 148 new courses built primarily this way, roughly doubling its total course offerings in about one year.
Practice DocumentedView practice →Jan 8, 2024
ProductivityPractice documented: Duolingo cut approximately 10% of its contractor workforce at the end of 2023, with the company explicitly citing AI as a reason that fewer people were needed to do translation and content writing work. A second round of cuts followed in October 2024, again targeting writers. These were the first widely reported cases of a major consumer tech company publicly attributing contractor layoffs to AI adoption.
Practice DocumentedView practice →Jan 9, 2024
ProductivityNew evidence: Duolingo lays off staff as language learning app shifts toward AI
Evidence AddedView practice →Jan 9, 2024
ProductivityNew evidence: Duolingo cuts 10% of its contractor workforce as the company embraces AI
Evidence AddedView practice →Jan 31, 2024
Data AnalysisPractice documented: When someone takes the Duolingo English Test at home, an AI system monitors the session by analyzing video, audio, typing patterns, and eye movements — looking for signs of rule violations across more than 75 different behavioral and environmental signals. AI flags potential issues, and then human proctors make the final call on whether to certify the test.
Practice DocumentedView practice →Sep 24, 2024
Creative GenNew evidence: Video Call lets you have real life conversations with Lily
Evidence AddedView practice →Jan 16, 2025
Creative GenNew evidence: Duolingo Launches AI-Powered Video Call for Android
Evidence AddedView practice →Mar 18, 2025
Data AnalysisPractice documented: Duolingo uses a machine learning model to decide, in real time, whether to show each individual user an advertisement, a subscription offer, or neither — predicting which choice is most likely to generate revenue from that specific person at that moment. The system was credited with generating tens of millions of dollars in additional annual revenue.
Practice DocumentedView practice →Apr 28, 2025
ProductivityPractice documented: In April 2025, Duolingo's CEO Luis von Ahn publicly declared the company "AI-first," meaning AI capability would be considered when hiring people, evaluating employee performance, and deciding whether to grow teams. The announcement triggered significant public backlash, including the loss of more than 400,000 social media followers, and the CEO later walked back some of the framing.
Practice DocumentedView practice →Apr 30, 2025
Creative GenNew evidence: Duolingo launches 148 courses created with AI after sharing plans to replace contractors with AI
Evidence AddedView practice →Apr 30, 2025
Creative GenNew evidence: Duolingo Launches 148 New Language Courses
Evidence AddedView practice →May 4, 2025
ProductivityNew evidence: Is Duolingo the face of an AI jobs crisis?
Evidence AddedView practice →May 6, 2025
ProductivityNew evidence: Duolingo Embraces AI in Push for Scalable Learning
Evidence AddedView practice →Aug 18, 2025
ProductivityNew evidence: Duolingo CEO admits his controversial AI memo ‘did not give enough context’ and insists the company never laid off full-time employees
Evidence AddedView practice →Jan 1, 2026
OtherPractice documented: After getting a question wrong — or right — learners can tap a button called "Explain My Answer" to receive an AI-generated explanation of the grammar or vocabulary involved. They can ask follow-up questions as if chatting with a tutor. This feature launched in 2023 as a paid-only feature and became free for all users on January 1, 2026.
Practice DocumentedView practice →Recent activity
Duolingo uses an AI chatbot (powered by a company called Decagon) to handle the majority of customer support requests for the Duolingo English Test. The AI fully resolves about 80% of chat inquiries without any human involvement, answering questions about test registration, scores, and institutional acceptance.
Before adopting Decagon AI, Duolingo's English Test support team used a different AI vendor that deflected only about 30% of email tickets and failed to launch chat automation after a year of development. After switching to Decagon — with a reported go-live time of under one month — the 80% chat deflection rate was achieved immediately. The system automatically syncs with the Duolingo English Test FAQ database every hour to stay current. In early 2025, the DET team planned to expand the Decagon-powered system to email support. Note: No evidence was found of an equivalent AI chatbot deployment for the main Duolingo language app's customer support.
Duolingo uses a machine learning model to decide, in real time, whether to show each individual user an advertisement, a subscription offer, or neither — predicting which choice is most likely to generate revenue from that specific person at that moment. The system was credited with generating tens of millions of dollars in additional annual revenue.
Disclosed in a March 2025 blog post, the system was built on dbt and BigQuery ML infrastructure using an XGBoost model configured as a contextual bandit — an AI approach where the system learns over time which decisions generate the best outcomes. The key innovation was shifting from predicting baseline purchase probability to modeling causal uplift: estimating how showing a particular message actually changes a user's likelihood of subscribing, rather than simply identifying users who were likely to subscribe anyway. This distinction allowed Duolingo to target persuadable users rather than those who would have converted regardless. The system was cited as accounting for approximately 25% of year-over-year revenue growth.
Duolingo uses a machine learning algorithm to decide which push notification to send each user, at what time, and how often — optimizing for the messages most likely to get a lapsed learner back into the app. The system was published as research in 2020 and was shown to increase the number of daily active users by 0.5% and improve retention of new users by 2%.
Published at the ACM SIGKDD 2020 data science conference, Duolingo's "Sleeping Recovering Bandit" algorithm processes approximately 200 million notification examples to learn optimal notification strategies per user. A key technical challenge was accounting for "novelty effects" — users becoming desensitized to repetitive messages over time — which the algorithm handles through recovery periods. The system also manages the conditional eligibility problem: not every notification is appropriate for every user at every time (e.g., a user mid-lesson should not receive a reminder). Different notification messages were found to perform differently across languages and user segments. Duolingo released the dataset of 200 million notification examples publicly.
All practices by category
Duolingo uses a machine learning model to decide, in real time, whether to show each individual user an advertisement, a subscription offer, or neither — predicting which choice is most likely to generate revenue from that specific person at that moment. The system was credited with generating tens of millions of dollars in additional annual revenue.
Duolingo uses a machine learning algorithm to decide which push notification to send each user, at what time, and how often — optimizing for the messages most likely to get a lapsed learner back into the app. The system was published as research in 2020 and was shown to increase the number of daily active users by 0.5% and improve retention of new users by 2%.
When someone takes the Duolingo English Test at home, an AI system monitors the session by analyzing video, audio, typing patterns, and eye movements — looking for signs of rule violations across more than 75 different behavioral and environmental signals. AI flags potential issues, and then human proctors make the final call on whether to certify the test.
The Duolingo English Test is the first major standardized language proficiency exam where AI handles every stage of the process — writing the test questions, calibrating their difficulty, choosing which questions each test-taker sees, grading answers, and producing final scores. The test costs $65, takes about an hour, and is accepted by more than 5,000 universities worldwide.
Every voice learners hear in the Duolingo app — characters speaking sentences, words, and dialogues — is produced by AI voice synthesis, not live human recordings. Duolingo built custom synthetic voices for each of its animated characters by training AI on recordings made by human voice actors. Those voices now generate unlimited audio across 30+ languages automatically.
Duolingo uses generative AI to write lesson exercises, dialogues, and entire language courses. Human experts set the rules — the target language, skill level, and grammar topic — and the AI produces the content for human review. In April 2025, Duolingo launched 148 new courses built primarily this way, roughly doubling its total course offerings in about one year.
Duolingo uses a large language model from OpenAI (GPT-4) to power live conversation practice inside its app. Learners can text-chat with AI characters in simulated real-world scenarios, or video-call an animated AI character named Lily who speaks and responds in real time. These features launched in 2023 and are available to paid subscribers.
In April 2025, Duolingo's CEO Luis von Ahn publicly declared the company "AI-first," meaning AI capability would be considered when hiring people, evaluating employee performance, and deciding whether to grow teams. The announcement triggered significant public backlash, including the loss of more than 400,000 social media followers, and the CEO later walked back some of the framing.
Duolingo cut approximately 10% of its contractor workforce at the end of 2023, with the company explicitly citing AI as a reason that fewer people were needed to do translation and content writing work. A second round of cuts followed in October 2024, again targeting writers. These were the first widely reported cases of a major consumer tech company publicly attributing contractor layoffs to AI adoption.
Duolingo uses an AI algorithm called Half-Life Regression to predict when a learner is about to forget a word or skill, then schedules a review exercise just before that happens. It works like a smart calendar for your memory — timing reminders to keep knowledge fresh without over-drilling what you already know. The system was published as a peer-reviewed research paper in 2016 and has been in production use since.
Every Duolingo lesson is shaped by an AI system called Birdbrain, which constantly estimates how well a learner knows each skill and selects the 14 or so exercises most useful for that specific person at that moment. It works like a coach who watches you practice and adjusts every drill based on what you're struggling with — updating within minutes of each completed exercise.
Related Companies
Trace Foundation
Walmart
Stability AI
Topaz Labs
Pika
Have evidence about Duolingo's AI practices? Submit a report.
Submit a report →AI Trace is free and nonprofit. Support our work