50M People. Still Can't Speak.

Luis Von Ahn is Guatemalan. Grew up watching people in his community unable to access economic opportunity partly because they couldn't afford to learn English. When he later invented CAPTCHA at Carnegie Mellon, the idea wasn't just security—every single time someone completed a human verification, they were also helping digitize books or train image recognition software. Millions of people doing useful computational work without realizing it, disguised as a security check. Von Ahn has always thought in terms of how do you get people to do something hard by making them not notice they're doing it.

That thinking is Duolingo's whole DNA. He launched the app in 2011 with co-founder Severin Hacker, which is genuinely his name, and 300K people signed up before it even launched. 10M downloads in the first year. The product wasn't just good—it was obsessively, almost pathologically optimized.

Here's one way to understand the level of that optimization. When Duolingo launched in 2012, 12% of new users came back the next day. One in eight. Through years of A/B testing basically everything—button shapes, notification timing, lesson length, reward animations—they got that number to over 40%. One in three. Just by changing product details. No new features. No marketing spend. Just running hundreds of experiments simultaneously and letting the data decide.

That's the kind of company Duolingo was.

The other thing they built that doesn't get enough credit: the streak. Every consecutive day you use the app your count goes up. Miss a day, it resets to zero. Duolingo's own data shows users with a seven day streak are two and a half times more likely to keep using the app than users without one. I've read posts from people who asked nurses to do a lesson for them during medical procedures so the streak wouldn't break. People on planes doing it at the gate. The streak isn't a feature. It's closer to a psychological contract. Something people feel genuinely attached to and protective of.

That level of user attachment matters a lot for understanding why what happened next went the way it did.

In 2018, Duolingo hit a wall. Daily active user growth was stagnating and the team trying to fix it was running out of ideas for tests that actually moved the needle.

Their answer was not to cut costs. Their answer was to build a model.

The data team broke down their entire user base into seven states—new users, current users, at-risk users, dormant users, resurrected users—and then ran simulations on every possible lever to find out which one, if you moved it just 2% per quarter, would have the biggest downstream impact on DAU growth.

The answer was Current User Retention Rate. Not acquisition. Not reactivation. Keeping the people who were already showing up.

They spun up a dedicated team around that one metric. Ran experiments specifically aimed at it. And between 2019 and the pandemic, they 4x'd daily active users. That is their actual playbook when things slow down. Find the lever through data, build a team around it, test relentlessly, let the numbers decide.

Keep that in mind.

Content creation at Duolingo has always been a crowdsourced problem. That's not new.

In 2013 they launched the Duolingo Incubator—a platform where bilingual volunteers from around the world could build new language courses within Duolingo's framework. No salary. Just the mission and the community. Over eight years, volunteers built over 90 courses and brought free language learning to 300M people. That was version one of the content model.

Version two was contractors. Translators and writers who replaced the incubator model post-IPO when they needed more consistency and quality control. Still relatively lean, still mission-aligned, but professional.

Version three is AI. And that's what the April 2025 memo announced.

Von Ahn posted it himself on LinkedIn. Not leaked—he put it up himself. The memo said Duolingo was going AI-first. Contractors doing work AI could do would be phased out. Teams wanting new headcount would have to prove the work couldn't be automated first. AI use would factor into hiring and promotion decisions.

He framed it like their mobile-first pivot in 2012, which had worked well. He said they couldn't wait for AI to be perfect. Move fast, take some quality hits, don't miss the moment.

The internet completely lost it.

TikTok videos about quitting piled up millions of views in days. Reddit went hard against them. A guy with a 1,613-day streak—over four years of showing up every single day—posted "Now AI. Goodbye." Not angry. Just done. There's a difference.

Duolingo's response was a case study in how to make this kind of thing worse. They went silent. Wiped their social accounts and left cryptic placeholder messages. Posted a fake video of a masked "disgruntled employee" that landed completely wrong. Then Von Ahn went on a podcast and said schools might mostly exist for childcare in the future — while his users were already upset about workers being replaced by AI. Genuinely bad timing.

By Q2 earnings, user growth had slowed. Von Ahn addressed it on the call: "The reason we came towards the lower end was because I said some stuff about AI and I didn't give enough context." He wasn't saying the strategy was wrong. He was saying how he communicated it was wrong. The plan didn't change.

Now. Here is where I think most people got the analysis wrong.

The memo was not wrong about what AI can do for the content problem. Von Ahn said it took Duolingo 12 years to build their first hundred language courses. Using AI they built 148 new courses in about a year. That's real. For a company whose mission is making language learning free and accessible to people who can't pay for tutors, spinning up courses for more languages faster connects directly to what they were supposed to be doing.

And look at where the content model came from. Volunteers. Then contractors. Now AI. There's a continuity there that the outrage kind of glossed over. The critics acted like cutting contractors was a betrayal of the company's soul, but the soul of the company was always about figuring out how to make high quality content at low cost. That was the original CAPTCHA insight applied to language learning. Version three of the same model isn't a betrayal.

What did take a hit was the feeling of the product. Former contractors described AI output that was flat and boring where Duolingo used to be quirky. In smaller courses like Irish, voice synthesis was reportedly mangling pronunciation badly enough that listening exercises became actual guesswork. The wit, the slightly too funny example sentences, the sense that a person was in on the joke with you—that was made by the humans who got cut, and it turned out to be harder to replicate than anyone expected.

So the technology helped with volume and hurt the texture of the product. That tradeoff is real.

But here is what almost nobody covered. The entire AI-first announcement was about making more of what Duolingo already does. Faster, cheaper, at bigger scale. What Von Ahn never mentioned was using AI to fix the thing Duolingo has never been able to do. The thing it's been failing at since 2011.

To be genuinely fluent in a language you need four things. Reading. Writing. Listening. Speaking.

Duolingo is great at one of those and pretty good at two others. Speaking is the one it never fixed. What passes for speaking practice on Duolingo—saying a phrase into a microphone and getting a pronunciation score back—is not speaking practice. That's something else.

Real speaking practice is what happens when you're halfway through a sentence and the word you need isn't there. It's when someone responds in a way you didn't predict and you have to process it and reply in real time. The cognitive skill of forming what you want to say in another language before you open your mouth—the mental preparation alone—takes hundreds of hours of actual back-and-forth to develop. You cannot get that from multiple choice exercises, and Duolingo has known this since 2011 and never really fixed it.

I went to a Mandarin conversation group a while back. There was a guy there who'd been on Duolingo for over a year. Solid streak, every day. Could barely put a sentence together in real conversation. Not because he wasn't trying. Because the app doesn't train that skill. I gave up on Mandarin myself so I kept my mouth shut—turns out that was the one thing Duolingo had taught both of us to do.

For years the alternatives were either expensive (private tutors on Italki, Cambly) or unreliable. Language exchange apps tried to fix the cost problem by connecting learners with native speakers but had their own issue. First conversation was usually great. Second one less so. Third one neither person showed up for.

Then AI voice got actually good.

Right now today ChatGPT in voice mode will have a real conversation with you in whatever language you're learning. It adjusts to your level. It corrects you gently. It slows down when you need it. It never runs out of patience. It costs less than a Duolingo subscription. There are apps built specifically around this—MakesYouFluent, JumpSpeak, others—doing AI conversation practice as their entire product.

The gap Duolingo left open for fourteen years is being filled right now, by people who are not Duolingo. While Von Ahn was figuring out how to cut content production costs, someone else was building the product Duolingo should have been.

And remember: when Duolingo hit a growth wall in 2018, they ran simulations, found the right lever, and built a team around it. They had the exact same analytical toolkit available in 2025. The question is what problem they pointed it at.

The MySpace comparison fits here, but maybe not the way you'd expect.

MySpace didn't get beaten by better technology. Facebook used the same technology. What MySpace lost was the product. It had the users, the cultural moment, the emotional attachment. Then it started optimizing for ad revenue instead of experience. The platform got cluttered, started extracting from users instead of serving them, and Facebook walked in with something cleaner. The migration happened slowly at first and then very fast.

Duolingo has genuine user attachment. The streaks are real. 50M people showing up daily is not a small thing. But the product teaches you to read a language and tells you it taught you to speak one. At some point enough people discover that gap, tell their friends, and the narrative changes.

There is a larger question hanging over this whole category that I don't think anyone wants to sit with yet.

Real-time AI translation is already functional. You can have a conversation today between two people speaking different languages and AI will translate for both with minimal lag. It makes mistakes. Some of them are genuinely funny. But it works well enough for most practical situations, and it is going to get significantly better over the next few years.

If that's true, the reasons most people start learning a language start to erode. The trip coming up. The person you want to communicate with. Getting by in a country where you don't speak the language. All of that becomes solvable without learning anything.

That doesn't make language learning worthless. There are things you only get from actually knowing a language. The way something reads in the original. A joke that doesn't survive translation. A conversation with no technology mediating between you and another person. Those things matter to a lot of people.

But they matter to a different kind of person than the one downloading a free app before a vacation. The casual learner market is most of Duolingo's users. And something is coming for that market that has nothing to do with whether their content was made by humans or AI.

The most recent earnings, end of February 2026, were actually fine by the numbers. Beat revenue estimates. First year over a billion. $400M stock buyback. Then Von Ahn said they're deliberately slowing revenue growth in 2026 to focus on getting to 100M daily active users by 2028. Expanding into chess and math and music.

Expanding into other subjects might be the smartest thing on the table. If language learning as a category shrinks, having a gamified learning platform that 50M people are already attached to—where the habit is already built, the streak psychology is already working—that's a genuinely valuable asset for teaching almost anything. The streaks and the owl might matter more as a general learning platform than as a specifically language one.

But that's a different company than the one Von Ahn built. And it requires admitting the original product had a ceiling that wasn't going to get fixed by going AI-first in the direction he chose.

They had the playbook. They used it once already and 4x'd their users. The window to point that same analytical rigor at the speaking problem, and actually make language learning work all the way through, was open. It's smaller now. Someone else is standing in it.