When Duolingo launched in 2012, the language-learning app became the poster child of gamification. The app is shameless and magnificent in its efforts to get users hooked on lessons with streaks, leaderboards, and timed challenges.
Many users — including me and my 1300-plus day streak — fell for Duolingo’s cartoon mascots and bizarre social media posts. The company has never been afraid to be belligerent in tone; Duo the owl is cute, but Duolingo has adopted a successful strategy of not coddling its users. The app regularly sends me push notifications from my own boyfriend begging me not to let us “break up” (our friend streak). Look, it doesn’t not work.
But this spring, Duolingo had a huge messaging misfire over AI adoption, and brought a lot more users close to ending things.
In April, Duolingo CEO Luis von Ahn announced to employees that Duolingo would be going all-in on AI. The company would look for AI expertise in future hires, and AI usage would be evaluated in employee performance reviews. It would also move to replace contract workers with AI where possible. It was this statement that stuck in the craw of many users, and honestly surprised me when I read it. I’m cynically certain that plenty of companies would love to replace expensive human workers with machines. Admitting it is another thing.
The memo was a called shot: Duolingo’s leadership sees AI as a paradigm shift, similar to the adoption of mobile phones in the 2010s. At the time, common wisdom would have dictated that a language-learning program should prioritize widely adopted platforms like PCs. Instead, the company went “mobile-first.” That bet certainly paid off. Duolingo saw 103 million users a month in 2024. Now, it wants to go “AI-first.”
It’s too soon to tell what the long-term effects of the decision will be. But in the short term, fallout has been loud and angry across social media. Longtime users are deleting the app, destroying 1000+ day streaks. The announcement has been painted as a failure in multiple publications. The Duolingo subreddit melted down so thoroughly that mods placed a moratorium on posts about AI.
Meanwhile, Duolingo stock prices have soared to over $500 (as of June 2, 2025), indicating that whatever users may feel about AI, the big boys who shovel money around think it’s here to stay.
Von Ahn later made a second statement, not walking back the “AI-first” shift, but couching it in gentler language.
“I do not see AI as replacing what our employees do,” he wrote. “I see it as a tool to accelerate what we do, at the same or better level of quality. And the sooner we learn how to use it, and use it responsibly, the better off we will be in the long run.”
Duolingo isn’t the only company doing this. Across the tech industry, workers are being evaluated on their AI usage, encouraged to experiment with AI tools in the service of supposed future productivity, and asked to train their own replacements.
It’s not being framed this way. Rather, executives are speaking about their AI initiatives like Luis von Ahn did: as tools to help people, rather than replace them.
Duolingo’s AI shift has been brewing for years
Of course, Duolingo has been inching towards “AI-first” for years. In 2023, it used OpenAI’s GPT-4 to create AI features that are only available in Duolingo Max, the highest subscription tier on the app.
One of these is “Explain My Answer,” which ostensibly tells users why their response to an exercise is wrong. In general, Duolingo will give you the correct answer if you get something wrong, but it won’t explain why you were wrong. When it comes to typos or misspellings, the error can be obvious. But it doesn’t help users if they’re fundamentally misunderstanding, say, a grammatical concept.
Previously, Duolingo hosted a forum where users could see explanations from other users and native speakers directly in the app. This was removed in 2022.
Now, the Duolingo subreddit is awash with users looking for answers to their questions. And many of them are blaming AI for their confusion. It’s is the scapegoat for nonsensical conversations, translation errors, and just plain awkward exercises.
Without confirmation from Duolingo, it’s impossible to say which of these issues is actually caused by Duolingo’s implementation of AI. In some cases, users are genuinely encountering software bugs rather than AI-created lessons.
But elsewhere, AI has genuinely changed Duolingo’s lessons for the worse. In Aftermath, Riley MacCleod writes that the Irish course he was pursuing has been ruined by AI voices that don’t pronounce Irish words correctly — a dire situation for a language that is literally endangered.
I spoke to Callie R., a former Duolingo user who is learning Japanese. They noticed that there was a mismatch between how words were pronounced by the robotic voiceover in word banks, versus how those same words were pronounced in exercises.
“This is just an aspect of how Japanese is written, that it isn’t possible in general to tell how a kanji is supposed to be pronounced when you see it in isolation,” Callie said. “It makes sense that an automated content generation process would make this kind of mistake, but a human team actively developing the course with learning outcomes in mind would not do this.”
They also pointed to observations from other users that Duolingo’s robotic voice isn’t capable of correctly speaking a Japanese pitch-accent, a crucial aspect of the language, and one that a native English speaker can’t easily pick up on.
“It wasn’t worth literally learning the language wrong on purpose,” they said.
After two years, Callie R. deleted the app and nuked their 700+ day streak.
AI should be good at this
The thing is, language learning is a field where AI large language models can actually be useful. These LLMs aren’t reliable truth-tellers, but they can be functional conversation partners.
Duolingo has long been criticized for not effectively teaching users how to speak — the app naturally focuses more on reading and listening, and the “speaking” lessons are more about pronunciation than they are about actively recalling words from memory. The latter is critical for genuine fluency in another language.
Duolingo is trying to address that flaw with two more Max-exclusive AI features that let users have conversations with Duolingo’s cartoon mascots. The most impactful of these is Video Call, where users can have a brief “phone call” with Lily, Duolingo’s resident depressed goth girl.
I had some conversations with Lily during a Duolingo Max free trial earlier this year. In each, she would ask me a question, repeat back to me what she had understood from my response, and then ask a simple follow-up. We talked about things like what animals or fruits I liked, or how my vacation was going. It forced me to recall Italian vocabulary on the fly, without a word bank to help me out.
This is an area where LLMs excel: generating human language based on speech patterns.
Unfortunately, LLMs fail in exactly the areas Duolingo is trying to disrupt. In his Blood in the Machine newsletter, journalist Brian Merchant spoke with a former Duolingo employee whose job had gone from writing lessons, to training AI how to write lessons, to non-existent.
“We had been working with their AI tool for a while, and it was absolutely not at the point of being capable of writing lessons without humans,” this employee told Merchant.
For Duolingo’s leadership, the flaws in the system are the cost of what they see as the cutting edge. Duolingo’s lessons are not supposed to be good.
“We can’t wait until the technology is 100% perfect,” von Ahn wrote in his email to Duolingo employees. “We’d rather move with urgency and take occasional small hits on quality than move slowly and miss the moment.”
The users who remain tapped into these conversations are suffering no small amount of confusion. A recent study showed that admitting to AI usage can cause people to trust you less. This is the situation that seems to be playing out on the Duolingo subreddit, where users are in a constant battle to figure out what is AI and what isn’t.
Some are deleting the app like Callie R. did. But there is a bubbling fear that a silent majority may simply not care or even be aware of any of these issues. The Duolingo subreddit has over 508,000 members — that’s less than 5% of Duolingo’s reported 116 million monthly users. And the subreddit itself isn’t entirely anti-AI. Plenty of users accept it, or simply don’t think there’s any point in fighting the tide.
Duolingo’s AI policy calls the app’s mission into question
My own Duolingo usage has always been predicated on one assumption: it won’t hurt your language-learning. Plenty of ink has been spilled over the fact that Duolingo most likely can’t make you fluent in another language. Sure, I’ve always reasoned, I know that. But doing a 5-minute Italian exercise every day when I’m too lazy or cheap or unmotivated to seek out a tutor is better than nothing. I am still learning, even if I’m not exactly leaping and bounding towards fluency.
But the influx of AI content puts this justification at risk. After all, language students don’t know what they don’t know.
“I don’t really care that it’s AI as long as there’s oversight and someone willing to pull the plug if it’s not producing real Japanese,” Callie R. said. Instead of pulling the plug, the people in charge at Duolingo are actively enabling users to learn bad Japanese, in the hopes that someday the AI will teach good Japanese instead.
There’s no obvious road map is to get there. LLMs can be taught to speak a language — it’s not clear that they can be taught to teach.
Duolingo is facing a problem of scale: it wants to offer lots of language courses, and creating those courses takes time and money. It has turned to AI to fill the desperate gaps where humans might be right, but can never be fast enough.
What makes Duolingo’s AI creep even more nefarious is that it’s most likely to affect languages with smaller userbases — like Irish or Navajo, both endangered languages. The vast majority of the app’s users are studying English, French, or Spanish. These are the courses that see a real investment of resources.
Duolingo gets great press for creating lessons that purport to familiarize users with Navajo. But what will happen if AI is used to “scale up” the Navajo program, with seemingly few human guardrails to ensure that the exercises are correct?
“Without AI, it would take us decades to scale our content to more learners,” von Ahn wrote in his first statement. “We owe it to our learners to get them this content ASAP.”
My question is… why? Why do we need more content for users immediately, when that content might be wrong or of low quality? It’s here that Duolingo’s mission of making language accessible crashes headlong into its role as a publicly traded company. Lessons need to scale so that users stay on the app, so that the app can make money.
Actually learning a language — or even simply treading water in one — doesn’t have a part to play.
From Polygon via this RSS feed
