How Duolingo went “AI-first” — and what the backlash taught it
Generative AI let Duolingo build 148 new courses in under a year — work that once took more than a decade. Then its CEO declared the company “AI-first,” the internet revolted, and Duolingo learned that how you roll AI out matters as much as the technology.
A mission bottlenecked by how slowly content gets made
Duolingo's reach depends on producing enormous volumes of teaching content — and for most of its history that content was hand-built, carefully and slowly. The bottleneck wasn't ambition; it was the sheer manual effort of creating courses.
Generative AI changed the math, letting the company produce learning material far faster and more cheaply than its hand-crafted process ever could.
Twelve years of courses, rebuilt in one
Using generative AI alongside an internal “shared content” system, Duolingo launched 148 new courses in under a year — roughly doubling its catalog. For perspective, its first 100 courses had taken about twelve years to build. The acceleration was an order-of-magnitude change in how fast the company could ship.
On the product side, Duolingo Max (launched March 2023 on GPT-4) brought learner features like “Explain My Answer” and “Roleplay,” putting the model directly in front of users.
A memo that lit a fire
In late April 2025, CEO Luis von Ahn sent an internal memo declaring Duolingo “AI-first.” It said the company would gradually stop using contractors for work AI could handle, would add headcount only when a team couldn't automate more of its work, and would weigh AI proficiency in hiring and even in performance reviews.
[Duolingo will] gradually stop using contractors to do work that AI can handle.
Luis von Ahn, CEO, Duolingo — “AI-first” internal memo, April 2025
How you say it matters as much as what you do
The memo set off a public backlash — frustrated users, unfollows on social media, and pointed criticism aimed at the contractor and “AI-in-reviews” framing. Von Ahn soon clarified that he hadn't given enough context and wasn't planning to replace employees with AI. By 2026, Duolingo had dropped the requirement to use AI as a performance-review metric.
None of this slowed Duolingo's actual AI usage. What changed was the framing: the company learned that mandating adoption — tying it to people's jobs and reviews — breeds resistance, even when the underlying tools are genuinely useful.
From mandate to outcomes
Duolingo is a bellwether for the “AI-first” operating model, where AI is the default tool and a real constraint on how teams grow. But its very public reversal on AI-in-reviews is just as instructive as its speed: the durable path to adoption is framing it around outcomes people want, not compliance they resent.
It's the same point Fautons makes about enablement: you can't mandate your way to fluency. Adoption sticks when people choose it because it makes their work better — and coaching, not coercion, is what gets them there.
The shift
- Course-building measured in years
- Content scaled by hiring and contractors
- AI treated as a product feature only
- “AI-first” pushed via mandates and reviews
- Adoption assumed rather than earned
- 148 courses shipped in under a year
- AI generating content at new speed
- AI woven into both product and operations
- AI-use-in-reviews mandate rolled back
- Adoption reframed around outcomes
I do not see AI as replacing what our employees do — we are in fact continuing to hire at the same speed as before.