When Duolingo launched in 2012, the language-learning app turned the poster little one of gamification. The app is shameless and luxurious in its efforts to get customers hooked on classes with streaks, leaderboards, and timed challenges.
Many customers — together with me and my 1300-plus day streak — fell for Duolingo’s cartoon mascots and weird social media posts. The corporate has by no means been afraid to be belligerent in tone; Duo the owl is cute, however Duolingo has adopted a profitable technique of not coddling its customers. The app frequently sends me push notifications from my very own boyfriend begging me to not allow us to “break up” (our buddy streak). Look, it doesn’t not work.
However this spring, Duolingo had an enormous messaging misfire over AI adoption, and introduced much more customers near ending issues.
In April, Duolingo CEO Luis von Ahn introduced to workers that Duolingo could be going all-in on AI. The corporate would search for AI experience in future hires, and AI utilization could be evaluated in worker efficiency opinions. It could additionally transfer to switch contract employees with AI the place attainable. It was this assertion that caught within the craw of many customers, and actually shocked me after I learn it. I’m cynically sure that loads of corporations would love to switch costly human employees with machines. Admitting it is one other factor.
The memo was a known as shot: Duolingo’s management sees AI as a paradigm shift, just like the adoption of cell phones within the 2010s. On the time, widespread knowledge would have dictated {that a} language-learning program ought to prioritize broadly adopted platforms like PCs. As an alternative, the corporate went “mobile-first.” That wager actually paid off. Duolingo noticed 103 million customers a month in 2024. Now, it needs to go “AI-first.”
It’s too quickly to inform what the long-term results of the choice shall be. However within the brief time period, fallout has been loud and offended throughout social media. Longtime customers are deleting the app, destroying 1000+ day streaks. The announcement has been painted as a failure in a number of publications. The Duolingo subreddit melted down so totally that mods positioned a moratorium on posts about AI.
In the meantime, Duolingo inventory costs have soared to over $500 (as of June 2, 2025), indicating that no matter customers might really feel about AI, the massive boys who shovel cash round assume it’s right here to remain.
Von Ahn later made a second assertion, not strolling again the “AI-first” shift, however couching it in gentler language.
“I don’t see AI as changing what our workers do,” he wrote. “I see it as a software to speed up what we do, on the similar or higher stage of high quality. And the earlier we discover ways to use it, and use it responsibly, the higher off we shall be in the long term.”
Duolingo isn’t the one firm doing this. Throughout the tech business, employees are being evaluated on their AI utilization, inspired to experiment with AI instruments within the service of supposed future productiveness, and requested to coach their very own replacements.
It’s not being framed this manner. Relatively, executives are talking about their AI initiatives like Luis von Ahn did: as instruments to assist individuals, relatively than exchange them.
Duolingo’s AI shift has been brewing for years
After all, Duolingo has been inching in direction of “AI-first” for years. In 2023, it used OpenAI’s GPT-4 to create AI options which might be solely out there in Duolingo Max, the very best subscription tier on the app.
Considered one of these is “Clarify My Reply,” which ostensibly tells customers why their response to an train is improper. Usually, Duolingo gives you the proper reply in the event you get one thing improper, however it gained’t clarify why you have been improper. In the case of typos or misspellings, the error may be apparent. However it doesn’t assist customers in the event that they’re essentially misunderstanding, say, a grammatical idea.
Beforehand, Duolingo hosted a discussion board the place customers may see explanations from different customers and native audio system instantly within the app. This was eliminated in 2022.
Now, the Duolingo subreddit is awash with customers in search of solutions to their questions. And lots of of them are blaming AI for his or her confusion. It’s is the scapegoat for nonsensical conversations, translation errors, and simply plain awkward workouts.
With out affirmation from Duolingo, it’s inconceivable to say which of those points is definitely brought on by Duolingo’s implementation of AI. In some circumstances, customers are genuinely encountering software program bugs relatively than AI-created classes.
However elsewhere, AI has genuinely modified Duolingo’s classes for the more severe. In Aftermath, Riley MacCleod writes that the Irish course he was pursuing has been ruined by AI voices that don’t pronounce Irish phrases accurately — a dire scenario for a language that’s actually endangered.
I spoke to Callie R., a former Duolingo consumer who’s studying Japanese. They observed that there was a mismatch between how phrases have been pronounced by the robotic voiceover in phrase banks, versus how those self same phrases have been pronounced in workouts.
“That is simply a facet of how Japanese is written, that it isn’t attainable basically to inform how a kanji is meant to be pronounced while you see it in isolation,” Callie stated. “It is smart that an automatic content material technology course of would make this type of mistake, however a human workforce actively creating the course with studying outcomes in thoughts wouldn’t do that.”
Additionally they pointed to observations from different customers that Duolingo’s robotic voice isn’t able to accurately talking a Japanese pitch-accent, a vital side of the language, and one {that a} native English speaker can’t simply decide up on.
“It wasn’t value actually studying the language improper on function,” they stated.
After two years, Callie R. deleted the app and nuked their 700+ day streak.
AI ought to be good at this
The factor is, language studying is a discipline the place AI giant language fashions can really be helpful. These LLMs aren’t dependable truth-tellers, however they are often purposeful dialog companions.
Duolingo has lengthy been criticized for not successfully instructing customers tips on how to communicate — the app naturally focuses extra on studying and listening, and the “talking” classes are extra about pronunciation than they’re about actively recalling phrases from reminiscence. The latter is crucial for real fluency in one other language.
Duolingo is making an attempt to deal with that flaw with two extra Max-exclusive AI options that permit customers have conversations with Duolingo’s cartoon mascots. Probably the most impactful of those is Video Name, the place customers can have a quick “cellphone name” with Lily, Duolingo’s resident depressed goth woman.
I had some conversations with Lily throughout a Duolingo Max free trial earlier this yr. In every, she would ask me a query, repeat again to me what she had understood from my response, after which ask a easy follow-up. We talked about issues like what animals or fruits I preferred, or how my trip was going. It pressured me to recall Italian vocabulary on the fly, with no phrase financial institution to assist me out.
That is an space the place LLMs excel: producing human language primarily based on speech patterns.
Sadly, LLMs fail in precisely the areas Duolingo is making an attempt to disrupt. In his Blood within the Machine e-newsletter, journalist Brian Service provider spoke with a former Duolingo worker whose job had gone from writing classes, to coaching AI tips on how to write classes, to non-existent.
“We had been working with their AI software for some time, and it was completely not on the level of being able to writing classes with out people,” this worker instructed Service provider.
For Duolingo’s management, the issues within the system are the price of what they see because the innovative. Duolingo’s classes will not be speculated to be good.
“We will’t wait till the expertise is 100% excellent,” von Ahn wrote in his e mail to Duolingo workers. “We’d relatively transfer with urgency and take occasional small hits on high quality than transfer slowly and miss the second.”
The customers who stay tapped into these conversations are struggling no small quantity of confusion. A current research confirmed that admitting to AI utilization could cause individuals to belief you much less. That is the scenario that appears to be enjoying out on the Duolingo subreddit, the place customers are in a continuing battle to determine what’s AI and what isn’t.
Some are deleting the app like Callie R. did. However there’s a effervescent concern {that a} silent majority might merely not care and even pay attention to any of those points. The Duolingo subreddit has over 508,000 members — that’s lower than 5% of Duolingo’s reported 116 million month-to-month customers. And the subreddit itself isn’t totally anti-AI. Loads of customers settle for it, or just don’t assume there’s any level in combating the tide.
Duolingo’s AI coverage calls the app’s mission into query
My very own Duolingo utilization has at all times been predicated on one assumption: it gained’t damage your language-learning. Loads of ink has been spilled over the truth that Duolingo almost certainly can’t make you fluent in one other language. Positive, I’ve at all times reasoned, I do know that. However doing a 5-minute Italian train each day after I’m too lazy or low-cost or unmotivated to hunt out a tutor is higher than nothing. I’m nonetheless studying, even when I’m not precisely leaping and bounding in direction of fluency.
However the inflow of AI content material places this justification in danger. In spite of everything, language college students don’t know what they don’t know.
“I don’t actually care that it’s AI so long as there’s oversight and somebody prepared to tug the plug if it’s not producing actual Japanese,” Callie R. stated. As an alternative of pulling the plug, the individuals in cost at Duolingo are actively enabling customers to study unhealthy Japanese, within the hopes that sometime the AI will educate good Japanese as an alternative.
There’s no apparent highway map is to get there. LLMs may be taught to talk a language — it’s not clear that they are often taught to educate.
Duolingo is going through an issue of scale: it needs to supply plenty of language programs, and creating these programs takes money and time. It has turned to AI to fill the determined gaps the place people is likely to be proper, however can by no means be quick sufficient.
What makes Duolingo’s AI creep much more nefarious is that it’s almost certainly to have an effect on languages with smaller userbases — like Irish or Navajo, each endangered languages. The overwhelming majority of the app’s customers are finding out English, French, or Spanish. These are the programs that see an actual funding of assets.
Duolingo will get nice press for creating classes that purport to familiarize customers with Navajo. However what’s going to occur if AI is used to “scale up” the Navajo program, with seemingly few human guardrails to make sure that the workouts are right?
“With out AI, it could take us many years to scale our content material to extra learners,” von Ahn wrote in his first assertion. “We owe it to our learners to get them this content material ASAP.”
My query is… why? Why do we want extra content material for customers instantly, when that content material is likely to be improper or of low high quality? It’s right here that Duolingo’s mission of creating language accessible crashes headlong into its function as a publicly traded firm. Classes must scale in order that customers keep on the app, in order that the app can make cash.
Really studying a language — and even merely treading water in a single — doesn’t have a component to play.