Duolingo CEO on AI, hiring, and the future of language learning

Executive overview

Most companies talk about AI productivity gains they haven't actually seen. Luis von Ahn has run the experiment at scale inside Duolingo and reports a more nuanced picture: speed-ups in specific pockets, not company-wide multiples.

The real shift isn't replacing workers — it's raising the floor of what a single employee can ship. Two non-engineers built Duolingo's fastest-growing course using AI in six months.

Employees who use AI well will outcompete those who don't — but AI alone won't replace the employee.

How Duolingo actually uses AI internally

• Product managers now bring prototypes to reviews instead of written proposals — faster decisions, less ambiguity.
• Engineers have changed workflows via AI coding tools, but overall feature output is not 10x faster.
• Every employee — including HR and finance — was required to vibe code something during a company-wide day.
• Internal Slack channels ("Best AI Practices", "AI Fails") drive peer-to-peer learning more than top-down mandates.
• Tracking AI usage in performance reviews backfired; Duolingo removed it after employees felt pressured to use AI for its own sake.
• Personal dashboards vibe-coded by PMs — showing user data by country — are now common.

The chess course: a case study

• Two employees with no chess knowledge and no engineering background proposed adding a chess course.
• Von Ahn initially rejected it; changed his mind after Guatemala's Minister of Education cited chess for teaching logical thinking.
• With no engineering support, they vibe-coded the prototype in six months using Cursor.
• They trained a custom AI model on an existing chess puzzle database after finding off-the-shelf AI poor at generating puzzles.
• The finished course now has 7 million daily active users — Duolingo's fastest-growing course.
• Key steps: learn the domain, do market research, prototype fast, iterate on mobile builds, involve engineers only for production.

Where AI is falling short

• AI coding works well on the "happy path" but debugging AI-generated code when it fails is slower than writing it manually.
• AI narrative generation is inconsistent: demos look great, but at scale only ~30% of AI-generated stories are usable.
• Large existing codebases resist AI assistance more than greenfield projects.
• Productivity gains are real but uneven — most engineers don't spend eight hours coding; meetings and coordination can't be sped up.
• Von Ahn sees no Duolingo-wide 10x speed-up; gains are "in pockets."

Hiring, layoffs, and the AI scapegoat

• Duolingo has never done a layoff; von Ahn sees each hire as higher-ROI because individual productivity is up.
• His view: most AI-attributed layoffs at large companies are structural over-hiring from the COVID era.
• "AI is not going to take your job. Somebody using AI is going to take your job."
• Duolingo now screens for AI openness in interviews — not proficiency, but willingness to experiment.

The stock drop and founder mindset

• Duolingo's stock fell ~82% after a deliberate decision to prioritise user growth over near-term monetisation.
• Von Ahn and the executive team made the call knowing it would hit the price; the goal is a much larger long-term user base.
• He doesn't regret it: "If we continued the way we were operating, growth would have been capped."
• He tracks daily active users instead of the stock price — same psychological trap, but at least it's something he can influence.
• Mental framework for setbacks: "Will this matter in six months?" — the vast majority of things won't.

Language learning and the AI threat

• Half of Duolingo's 100M+ active users learn as a hobby — translation tools don't reduce that demand.
• The other half are learning English for practical necessity; no phone translator replaces fluency for university admission or immigration.
• Computers have beaten humans at chess since 1997; chess participation has grown since then — technology doesn't kill hobbies.
• Simultaneous translation was already excellent before LLMs (Google Translate); language learning demand has gone up, not down.
• ~1.8 billion of the 2 billion people learning a language globally are learning English — the market is enormous and growing.
• Von Ahn is unconcerned about vibe-coded competitor apps: "There are 2,000–3,000 language learning apps already. Now there'll be 20,000. We're not particularly worried."

Five professions: von Ahn's predictions

1. Social media manager — not going anywhere.
2. Translator — fewer needed; surviving as a premium service for high-stakes use cases.
3. Teacher — not going away; irreplaceable for motivation, context, and human connection.
4. Strategist — uncertain; AI handles known patterns well, but human ingenuity still needed.
5. Project manager — not going away; managing interpersonal dynamics between people is hard to automate.

Broader outlook

• Most professions will be transformed, not eliminated; some companies will do the same work with fewer people.
• Von Ahn finds predicting the future significantly harder now than five years ago — the pace of AI change makes extrapolation unreliable.
• He would not want to be choosing a college major in 2026.
• His advice for individuals: find the right AI tool for your specific job, automate parts of it yourself, and stay open.
• If starting over, he'd still choose languages — 2 billion learners, with no other subject close.