The owl is tapping at the window again
Duolingo is the world's most popular language learning app. It’s also a disruptive hotbed for experimentation with innovative business models. But now, it too faces disruption.
A phone buzzes in the dark.
Sleepy voice: “Who’s messaging you at this time of night?”
A sigh. “Lily.”
“Lily? Who’s Lily?”
“One of Duo’s friends.”
Awake now, bristling for a fight: “The owl?”
“Look, I just—”
“Can’t you just leave it?”
“But my streak—”
“It’s late. I’ve got meetings tomorrow.”
“I’ll be, like, five minutes.”
“This is becoming a problem.”
Don’t say you can quit anytime. “I’m doing this for us.” Is that better? “Remember that place on the plaza, in the sunshine, with the locals and the nice food?” Phew, nice save.
“I got ill, remember? I’m lactose intolerant.”
“And if I could speak the language, then that wouldn’t happen again… Look, I promise, I’ll be quick.”
Heavy silence. “Fine. But keep it down.”
A ruffle of blankets. Floorboards creek. A door opens and closes softly.
Muffled, but slowly enunciated: “Sí, la mujer bebe leche.”
We’ve all been there, right? Lured into Duolingo’s compulsion loop, completing a few exercises while the kettle boils, squeezing in a bit of practice last thing in the day, whispering phrases into a phone in the bathroom at work…
I’ve had an on-off relationship with Duolingo since it launched. I was drawn by its appealing promise – the free, fun, and effective way to learn a language! – but also intrigued by its business model. After several years of the bird, I relapsed while on a break in Barcelona and a sudden need for some Spanish.
It turns out that Duolingo is at a crossroad. It’s both a model for AI use cases, but also a warning.
Not all heroes wear capes…
… but as the last 32 films in the Marvel Cinematic Universe would have it, every hero needs an origin story.
Luis von Ahn grew up in a wealthy family in Guatemala City. He set his sights on going to college in the United States to pursue an academic career in computer science. Despite attending a private English-language school, he needed to take the TOEFL as part of his college application. And the closest TOEFL centre was a $1200 flight away in El Salvador.
By luck of birth he had the money, but he knew he was in a privileged position. He left the experience feeling the system was profoundly unfair and extractive.
Several years later, Luis von Ahn completed a PhD on human computation – or games with a purpose. This is where a game is designed for humans to support computers with tasks that computers can’t successfully or accurately complete independently.
In 2004 he came up with ESP Game. The gameplay used a simple mechanic based on guessing. Players are paired randomly and presented the same image; they suggest labels for the image until they both enter the same label. They compete against the clock and sometimes obvious labels are taboo, forcing them to reach for more obscure labels. The game was licensed by Google and provided both metadata for Google Images and a training dataset for its AI-powered labelling programme.
In 2007, aided by a MacArthur Fellowship, von Ahn created ReCAPTCHA. CAPTCHAs require users to either identify details in pictures in order to prove that they’re human. ReCAPTCHA presents users with scrambled-looking letters to transcribe. These illegible-to-machine letters were taken from a text-scanning programme, and so users were helping with transcription.
The 2000s were a period of mass digitisation. It shouldn’t come as any surprise that Google bought ReCAPTCHA for an undisclosed amount and put it to work on its Google Books programme.
Around the same time, von Ahn met Severin Hacker. Hacker was deeply interested in video games and had come from Zurich to study at Carnegie Melon University.
They had a shared interest in language learning and equity in education. And so they began working on Duolingo, a game-like environment that would weave in elements of human computation. It was launched in private beta in late 2011 before being made public in June 2012.
Mi dinero
Right from the start, Duolingo was free to users. As Hacker told Mashable in 2015:
The single biggest thing that we did differently was to make Duolingo 100% free… We wanted even the most disadvantaged person with an internet connection to have access to high quality language education.
Seed funding came early because von Ahn had built a reputation from his previous two businesses. In 2011, Duolingo raised $3.3m in Series A funding from Union Square Ventures. As Brad Burnham, a Duolingo board member and partner at Union Square Ventures told Forbes:
We were very impressed with what [von Ahn] had done in the past and with him personally. We didn’t really try to model the business.
To scale development, money had to come from somewhere. But von Ahn and Hacker had a cunning plan to generate revenue from this entirely free platform, using von Ahn’s experience of human computation.
Duolingo presented users learning English with paragraphs of English text to translate into their mother tongue. The paragraphs would come from whole articles. Multiple users would translate the same section of text, so there then followed a sorting process to arrive at the best single translation. At scale, articles could be quickly translated with the final filtered-and-stitched translation being of much higher quality than was available by machine translation at the time.
International news websites, including CNN and Buzzfeed, paid for this service. But the effort required to collate and stitch the translations was taking valuable time away from developing the product.
So, true to the start-up philosophy of the time (scale first, worry about money second), they abandoned plans for monetisation and simply focused on the product. They leaned heavily on venture capital, which came easy because the prevailing philosophy of the time was that you invest in the person not the product.
Von Ahn described the journey to monetisation to Tim Ferriss in 2022:
Around 2015, 2016, we had raised a bunch of venture capital – around $100 million total – and we weren’t monetising at all but we had grown to a pretty big amount of active users. We raised a round of funding from – around the time it was called Google Capital, now it’s called Capital G – and a partner that was on our board, Leila Sturdy, took me out to a bar and said:
“Listen, you’ve been raising a lot of capital and your valuation keeps going up and up [at the time I think we were valued at something like half a billion dollars], right now you’ve just raised money from Google, and let me tell you […] you’re not going to find another fool to raise bigger money at a higher valuation with no revenue. You’ve got to figure out how to make money.”
And so she gave me enough drinks to convince me to promise her that next time we spoke I would have figured out how to make money. And so I went back and I said: “Alright everybody, we’ve got to figure out how to make money.”
And we did, but it was very important to us we really cared that we— you know, the easiest way to make money when you’re teaching is just to charge them to learn. It was very important to us that we didn’t do that because the mission of our company was really giving access to education to everyone for free. So, and we were a little constrained here, in the end we ended up finding a monetisation model that worked out really well, which is this freemium model […] you don’t have to pay, but you may see some ads, we make money from the ads and if you don’t like the ads you can pay us to subscribe to the premium version of Duolingo and we turn off the ads and you have a few extra perks. And that combo worked out really well – enough that that’s what we used to IPO and it’s grown to hundreds of millions of dollars of revenue. […] We’re very proud that you can still use Duolingo for free […] something like 97% of our active users use Duolingo for free, yet we still make more money than other education apps, so we’re very proud. The monetisation was a big shift in our company.
So, below is a question for you – and it’s a question we’ll come back to later. This is a long post, so if your email (looking at you Gmail) cuts it off, you’ll need to visit the post online to read the full thing. Miraculously, responding to this poll will take you there. So, all the more reason to play along.
Fiscal mechanics
The freemium model comes straight from the world of gaming. So it’s worth taking a moment to look at the way Duolingo’s game mechanics are optimised around revenue.
First, the basic mechanics:
Each language learning path consists of themed units, broken down into quizzes, stories, occasional videos, and treasure boxes. You generally progress through in a linear sequence, but can jump ahead if you can pass a test.
Each quiz or story consists of objective-answer questions, testing one or more of reading, writing, listening and speaking skills.
For each quiz or story successfully completed, you earn Lingots – Duolingo’s in-app currency. You can also earn Lingots by opening treasure boxes, completing challenges or side-quests, or by converting real-world cash into Lingots.
Players begin with five hearts. For each wrong answer, you lose a heart. If you lose your last heart during a quiz, you’ll exit the quiz and will not be able to re-try until you have more hearts.
You can earn extra hearts by waiting until they regenerate, practising, watching an advert, or exchanging Lingots for one or more hearts.
Players can change some of the mechanics by upgrading to premium. For a monthly fee of $7.99 (or a discounted annual subscription) for Super Duolingo, users can banish the ads and have unlimited hearts.
But if you’re super-flush, you can shell out for Duolingo Max, which has all the features of Super Duolingo combined with ChatGPT 4.0-powered practice conversations. Duolingo Max is priced at an eye-watering $30 a month (or a discounted annual subscription) and currently only available for English, French and Spanish.
So, if you did the quiz above (you did do the quiz, right?), would you be amongst the 97% who take the free ride or are you reaching for your wallet?
For the year ending 31 December 2022, Duolingo had 4.2m paying subscribers but 60.7m monthly active users. In 2022, Duolingo earned (GAAP) revenue of $369,495,000. Subscriptions account for 74% of revenue, advertising 12%, and in-app purchases 5%. (We’ll get to the remaining 9% later.)
This small base of paying users is common enough in the freemium model, where these users are known as “whales”. There are relatively few of them, but they’re power-consumers who effectively subsidise the free users.
But is Duolingo profitable? In 2022 it posted a loss of $59.6m, a little better than the previous year’s loss of $60.1m. It’s no wonder that the app pushes its subscription layers so heavily…
Learning design
Duolingo’s monetisation is not the only thing that has changed. Its approach to learning design has professionalised.
Back in the distant beginnings of the app, the approach was a true home-brew proof of concept:
To create the initial courses, Hacker and Von Ahn read a pile of books on language instruction, including Spanish for Dummies, and built crude basic courses for English speakers to learn Spanish and German. They googled the 3,000 most commonly used words in each language, translated them into English and used the words to compose simple sentences. Then they wrote an algorithm that would spit out lessons that included prompts for sentence translation, listening, writing and speaking.
Picture a green owl vomiting up a pellet containing the undigested bones of Margarita Madrigal.
As Duolingo received funding, so it expanded its team of experts and worked on building scale. True to the spirit of the 2010s, the company harnessed user-generated content. More than 1000 volunteers contributed to the design of language courses, supporting in-house experts.
When I first heard about the volunteer communities, I had mixed feelings. A repeated theme of monetising unpaid labour cuts through von Ahn’s businesses. These volunteers were helping to expand and refine a business aggressively raising VC cash. But on the other hand, without passionate, knowledgeable and self-motivated experts, would Duolingo have expanded into Welsh, Irish, Scottish Gaelic, Haitian Creole, Navajo, and, erm, Klingon or High Valyrian?
User-generated content came with its own challenges. With no underlying philosophy of learning or principles to structure the courses, quality between languages was variable. Tools to support data-informed decisions were developed and rolled out to volunteers, which supported working towards a greater alignment to CEFR language levels.
In 2021, as Duolingo raised $521m at IPO (giving it a $6.5bn valuation1), the company ended the volunteer programme. A fund of $4 million was shared amongst the contributors who had helped develop the courses – and any further work would be paid. But this shift was more radical than it sounds. It ushered in greater standardisation, the introduction of linear learning paths, and formal development schedules and templates.
With professionalisation came an increased use of AI in the production process. A 2022 blog provides a high-level overview of the way Duolingo uses AI in content generation. The process begins with human experts creating the curriculum design and raw content, before increasingly handing over to AI tools to generate and scale exercises within the parameters set in the earlier stages.
We should avoid thinking about AI as being one amorphous black box. Various AIs are used in the product to various specific ends. For example: character voices are now AI generated for at least some courses; Birdbrain analyses user data and serves up activities that target weaknesses or refresh vocabulary that the user hasn’t tested for a while; a dedicated AI checks the user’s pronunciation during spoken exercises.
And as Duolingo developed the GPT-4-powered Max super-subscription features, it spoke about applying the principles of building a human-developed foundation before increasingly handing over to the AI.
Duo tiene un problema
But, arguably, Duolingo has handed over more of the work to ChatGPT earlier in its development process.
Humans write the scenarios that learners see in Roleplay—they make sure the initial prompt (Talk about a vacation! Ask for directions!) is aligned with where the learner is in their course. Our experts also write the initial message in the chat and tell the model where to take the conversation.
[…]
If you notice a mistake or error in an AI-generated response, you can report it by holding down the inaccurate message—a menu will pop up and allow you to select your reason for reporting. Our team will watch for these reports and use them to train the model to be more accurate.
Is vetting and providing feedback to improve the AI is being left to paying users? The question is blunt, but it’s fair to ask given what we know about human computation models.
But here’s the rub. The extra functionality is a close parallel with ChatGPT’s native environment. Consequently, it’s clear that there’s an alternative chatbot that can help identify and correct your language mistakes outside of Duolingo. And if you don’t mind using v3, it’s free.
The BBC recently carried an article under the headline “Students switch to AI to learn languages”, outlining various developments on this front:
To stay ahead of the trend, well-established language-learning apps have been integrating AI into their own platforms. Duolingo began collaborating with OpenAI in September 2022, using that company's GPT-4.
It seems unlikely that AI chatbots will replace Duolingo altogether. Joy Ehonwa, an editor and writer in Lagos, mainly uses Duolingo to learn French.
But lately she's also been using an AI chatbot developed in Nigeria to help with her French.
She interacts with this chatbot, Kainene vos Savant, the way she would with “an all-knowing human friend”.
Ms Ehonwa explains, “My questions revolve around ‘why’? When I fail an exercise on Duolingo, the app is not great at helping me understand why the thing that was wrong, was wrong. So, I ask Kainene. Also, when I think a sentence should be a certain way and it isn't, I ask Kainene why and she helps me understand why it cannot be the way I expect it to be, as an English speaker.”
As I speculated in my last post, easy-to-substitute content is at risk from generative AI. There’s a potential direct replacement to challenge Duolingo’s premium subscription layer.
And for a highly leveraged VC-funded business, that’s a problem. Let’s go back to this FT article from 2 May: Education companies’ shares fall after warning over ChatGPT.
On Tuesday, Chegg’s shares plunged by half and the warning rattled other companies, with shares in London-listed Pearson falling 15 per cent, language-learning platform Duolingo down by 10 per cent and US-listed education company Udemy dropping by more than 5 per cent.
Duolingo has been able to build and expand its user base for such a long time without turning a profit because it’s backed by investors with their eyes set on a long-term horizon. Capital raised has been a bet that the company will turn profitable in the future.
Compare this to Uber, where venture capital subsidised rides in order to undercut traditional competition – with the goal of leaving Uber the only game in town. It’s the same strategy Starbucks applied by opening coffee shops so close to each other they crowded out the competition. Spending to aggressively achieve market dominance before then worrying about profit is a well-established strategy for well-funded rapid-scaling start-ups.
But OpenAI is also a VC-backed aggressively expansive start-up. And OpenAI is currently dominating the weird “OMG my product is going to destroy the world” discourse that’s consuming so much of the oxygen in the generative AI discussion. This is Godzilla versus Mothra stuff. Could ChatGPT start to lure away some of Duo’s whales?
If so, Duolingo has a secret weapon. And it’s one which may be more concerning for the incumbent education sector.
Yo tengo un examen en ingles
The 2016 launch of the Duolingo English Test (DET) was foreshadowed in the origin story of Duolingo. An affordable, recognised English proficiency test that could be taken from any location with an internet connection. This is, ultimately, the solution to the problem von Ahn identified on the plane to El Salvador.
Worldwide, more than 4500 schools, colleges and universities recognise DET. Of those, 1816 higher education institutes accept DET as part of their entry requirements for undergraduate programmes. In 2022, DET earned $32,718,000, making it the business’s third largest source of income – accounting for just shy of 9% of total revenues.
At the time of writing, Duolingo is on the hunt for “[a]n experienced professional to build the strategy and lead DET’s institutional engagement teams outside the US! You will lead a team based in our international priority markets of Australia, Canada, and the UK.”
A handful of universities in the priority markets already accept DET. If Duolingo’s core business starts to give way to ChatGPT (or similar generative AIs), then growing this share is vital for the company’s ongoing success. And once you’ve got penetration in these core English-speaking markets, other markets will follow based on trust and reputation. And as quite a few readers will know firsthand, English language testing is a well-established market.
So, let’s take a closer look at Anglia Ruskin University for no better reason than it appears near the top of DET’s alphabetical list… ARU lists three preferred tests for international students applying for an undergraduate or post-graduate programme: IELTS (Academic), TOEFL iBT, and Pearson PTE (Academic).
The website goes on to say:
We are aware that from time to time there may be difficulties in undertaking in-person tests at recognised test centres, due to closures, reduced testing availability, location, etc. and where this is the case we may be able to accept one of the following alternative tests
Which are: TOEFL iBT Special Home Edition, IELTS Indicator (taken from 1 July onwards), IELCA Home Edition.
But no mention of Duolingo. After some judicious googling I arrived at the English requirements for international students at Anglia Ruskin University London. ARU London’s preferred English test is its own computer-based test, developed with Password. But, if students aren’t able to take the ARU London English Test:
We will also consider the Duolingo English Test, with scores of 100 for undergraduate entry and 110 for postgraduate entry, and no less than 90 in all subscores (Literacy, Conversation, Comprehension, Production).
Is it just me or does “consider” sound kind of ambivalent? Perhaps it’s just because there’s an own-label equivalent exam available? Or perhaps there are concerns about quality or security.
Acceptance and more widespread recognition is, at least in part, an engineering problem. How to maintain a high level of security in an entirely online on-demand exam? How to test each skill domain accurately, effectively and fairly? How to create and manage a sufficiently large bank of high-quality questions to reduce predictability and repetition? How and where to use human quality control and validation?
If nothing else, Duolingo’s origins show that engineering problems are a particular strength. While established language exam providers are swimming upstream to transfer their well-established quality processes and procedures to the digital realm, Duolingo is swimming downstream. This too is the stuff of kaiju battles.
For context, as of late June 2023 Duolingo’s current market cap is $6.07bn whereas Pearson’s current market cap is $7.41bn. So, the owl is definitively a unicorn.