How AI killed the soul of the world's favorite language-learning app

Last week, I hit day 850 on Duolingo, and I realized I’m mostly just protecting a number at this point.

Puedo comprender español un poco, pero no en el nivel que quiero.

I’ve lived with that sentence in my head for over two years now. It’s the honest version. I can catch the gist of a conversation, but I can’t think in Spanish, I can’t argue in it, and I certainly can’t write with any soul.

For a long time, I viewed my streak as proof of character — discipline, consistency, showing up. And yes, maybe a flex I can gloat about in nerd circles. But as someone entrenched in UX and GenAI, I can see the gears grinding, and “guess” the points where the system is breaking.

There is a visceral difference between a product that is “sticky” because it’s useful and one that has become a psychological gridlock. Lately, my Duolingo streak feels less like progress and more like leverage — not leverage for me, but leverage on me.

I’m paying a retention tax to a unicorn that has traded its pedagogical soul for a bandit algorithm designed to keep me clicking until I’ve reached a state of total loss aversion.

§The diabolical science: the recovering bandit

When I started feeling that Duolingo was nudging me in strangely calculated ways, I went looking for the why. I found the blueprint in their own research: a 2020 paper titled A Sleeping, Recovering Bandit Algorithm for Optimizing Recurring Notifications. It confirms that the app treats your psychology like a multi-armed bandit game.

The app optimizes reminders using what it calls a “Recovering Difference Softmax Algorithm.” It’s a dry, academic name for a system that views your psychology as a series of probabilities to be solved. In this framework, the reward the machine seeks isn’t your fluency or a breakthrough in a difficult Spanish tense. It is a binary data point: did you open the app within two hours of the notification?

It’s an optimization loop for addiction rather than education.

There is even a recency penalty baked into the code. The system doesn’t actively spam you — it waits. It models your memory decay using exponential functions, precisely calculating the moment the “novelty effect” of a specific nudge has faded enough to work again. It lets the image of the sad owl, or the guilt-tripping “we’ll stop sending these” reminders, breathe just long enough for your emotional guard to drop. These are timed, probabilistic “arms” in a multi-armed bandit setup, specifically designed to trigger a Pavlovian response.

But the part that feels truly diabolical is the sleeping nature of the algorithm. The most aggressive engagement tactics stay dormant until you hit specific thresholds — like a simple 3-day streak.

At day 850, I’ve probably moved past being a “learner” in the eyes of the machine. I am now a high-value data point with maximum loss aversion. The system knows exactly how much I have to lose. It weaponizes that 850-day anchor, using it to force an engagement reward that has absolutely nothing to do with my ability to speak Spanish.

§The unicorn stagnation and the exclusivity failure

When apps become unicorns, or when foundational startups become corporate companies, they often lose the spark that made them successful in the first place. They move from solving a problem to protecting a metric.

Duolingo tries hard to sell an exclusivity feel with tiers like Duolingo Max, but it feels hollow because it’s a premium price for a mediocre experience. At $168/year, you’re paying for 30-second AI “video calls” and “roleplays” that lack depth and barely help you learn.

They’ve replaced the human soul of the app with an “AI-first” mandate. In January 2024, they fired 10% of their contract workforce — over 100 curriculum experts and translators — to let AI “accelerate” the work. CEO Luis von Ahn was blunt about the trade-off: “We’d rather move with urgency and take occasional small hits on quality than move slowly and miss that moment.”

As a builder, I find this heartbreaking. Human-in-the-loop (HITL) isn’t a luxury or a nice-to-have. It’s literally the cornerstone of quality, especially when we’re talking about education. Linguists aren’t just there to correct typos or philosophize conjugations. They preserve the irregular, cultural, and frustratingly human edges of a language. When you replace them with “stochastic parrots” optimized for scale, you get linguistic smoothing — content that is grammatically correct but emotionally and culturally flat.

§High haptics, low heart

Objectively speaking, the app has never looked better. The animations are crisp, powered by a budget-friendly “system of states” in Rive. The haptics are tuned to the millisecond to provide a dopamine hit for every correct answer. But this is the polish tax: the interface improves while the depth is thinning.

Duolingo’s website displays a graph showing improved test scores in 2024 compared to previous years. But power users don’t live in averages. In 2022, they shut down the user forums — the one place where a context layer actually existed for users to discuss linguistic nuance. Now, you’re left in a solitary cell with an AI that doesn’t want to teach you so much as it wants to keep you clicking.

Instead of adding meaningful features — like a social-media-style platform for real conversation, a word-of-the-day widget, or reminders that actually offer value instead of gaslighting — they’ve doubled down on animations and character personalities. They are perfecting the vibration of the “correct” button while the house of pedagogy burns down.

§The user-retention tax

Entrepreneurship teaches us about sunk costs. My 850 days are gone regardless of what I do tomorrow. I’ve been convincing myself to stay because I might “recommit,” but the app’s architectural drift makes that recommitment impossible. The gridlock is so complete that newcomers are bombarded with “shitty” ads and a pushy journey, while loyalists are alienated by the lack of depth.

There is a cognitive load of maintaining a number that no longer maps to growth. The business side reflects this tension, too. In November 2025, Duolingo’s stock dipped 31%, wiping out $4 billion in valuation as the market realized that a learning app without soul can’t sustain its price point.

Discipline inside a misaligned system isn’t learning — it’s closer to being just a habit. And if the system is so optimized for keeping me that it’s no longer optimized for growing me, then the tax is simply too high to pay. For my Spanish, and my sanity, that distinction matters more than the owl’s feelings.