Inside language app Duolingo’s controversial new ‘artificial intelligence first’ approach to teaching

Duolingo has been riding a years-long wave of popularity since cooped-up citizens spent their pandemic lockdowns using the popular app to learn a second language, but its shift to an “AI-first approach” has sparked fears about its quality — and worried linguistics experts. The owl-mascoted app’s bosses revealed last year it would cut 10 per cent of its contractor roles and this week outlined it would “gradually stop using contractors to do work that AI can handle” in an effort rapidly expand its course offering. Some experts suggest the decision will have broader ramifications on how second languages are taught beyond Duolingo’s own app. Duolingo chief executive Luis von Ahn said Duolingo’s use of AI to accelerate content creation was “one of the best decisions we made recently,” even if it meant taking “occasional small hits on quality”. He told employees: “AI isn’t just a productivity boost. It helps us get closer to our mission.” Mr von Ahn said Duolingo would shift to become “AI-first despite the technology not being ‘100 per cent perfect’ yet.” “To teach well, we need to create a massive amount of content, and doing that manually doesn’t scale,” he said. “Without AI, it would take us decades to scale our content to more learners. We owe it to our learners to get them this content ASAP.” Underscoring its aggressive push to infuse artificial intelligence into all corners of its business, Duolingo recently released 148 new AI-developed courses, which Mr von Ahn said would once have taken more than a decade to achieve. “Developing our first 100 courses took about 12 years, and now, in about a year, we’re able to create and launch nearly 150 new courses,” he said. “This launch reflects the incredible impact of our AI and automation investments, which have allowed us to scale at unprecedented speed and quality. Despite the rapid rise in course offerings, some experts fear the shift to AI will have a hit on the quality of learning, pointing out that artificial intelligence “doesn’t care if it’s right or wrong.” “AI is a predictability machine, not a thinking machine,” Murdoch University associate professor of communication Tauel Harper said. “It’s doing its best, but it doesn’t care if it’s right or wrong.” He pointed out that as a user, “your capacity to check if AI is correct or not is diminished with it being in another language.” “So the user won’t know if it’s right or wrong either,” Dr Harper said, encouraging users to regard the learning app sceptically and to check things with trained translators or native speakers. He said the migration to AI on learning apps was just the beginning, however, and that it was “very likely we’ll see a shift in how languages are taught in classrooms.” “As we move further and further down this path, it’s really important that we maintain craftsmanship and our own skills and abilities, because AI does make mistakes,” he said. “And if there are no experts around to spot the mistakes when they happen, then we will find ourselves in a very different world very quickly.” Perth mum Patricia Cuozzo was only a few weeks away from a 1000-day learning streak when she went into labour with her first-born. “The thing I never liked though about Duolingo was that it’s hard to understand grammatical rules,” she told The Sunday Times. But she questioned whether Duolingo would run the risk of “hitting and missing a fair bit”, if it didn’t use the correct prompts when getting AI to generate lessons in certain languages. Monash University associate Professor Abhinav Dhall said that for some lesser spoken languages and dialects, generative AI tools may be more likely to produce inaccurate results “due to a lesser amount of training data”. “Generative AI tools such as ChatGPT are trained on massive amounts of data, which allows them to answer accurately to a large range of queries ... particularly in widely spoken languages,” he said. “However . . . for low-resource languages, generative AI tools are required to be fine-tuned with more human knowledge. The same challenges may arise from a cultural sensitivity perspective too. “So in terms of structured learning, these tools currently fall short of replicating the systematic progression found in traditional language-learning applications. The AI tools need to be explicitly trained or fine-tuned for this purpose.” Dr Dhall, of Monash’s Department of Data Science and Artificial Intelligence, said Duolingo’s move toward generative AI was not isolated and was mirrored across the education technology industry. “This shift clearly reflects a broader and likely lasting transformation in the way languages are learned,” he said, adding the learning experience would become more dynamic, personalised and accessible with generative AI learning tools embedded in language-learning platforms. “A balance between AI generated automated assistance and human language expert’s judgement will be important in ensuring that language education remains both high-quality and contextually relevant,” he said. Perth local Tess O’Brien is on a 520-day streak learning French on Duolingo and said she’d recently noticed changes in the app. “But it feels the same in terms of the content and lessons. I’m not noticing it’s more tailored to me or anything,” she said. Twenty-one-year-old Annika Rasmusses has been using Duolingo for nine years to learn Danish, and praised some of the platform’s newest updates. “They’ve been super helpful,” she said. “There’s these new lessons now where they’ve got this AI-created video call with little avatars.” University of Melbourne Lecturer of Applied Linguistics, Haeun Kim, said the shift to AI would mean more courses and activities for learners given traditional class development was quite time and resource-intensive. “It’ll save a lot of time. But it all depends on quality. It doesn’t matter if we have a thousand courses if they’re poorly written,” she said. “There needs to be a reviewing process.” Dr Kim also said the reliance on AI models risked linguistic “fossilisation” and the reinforcement of mistakes. “For beginners, where they’re learning simple sentences, it will likely be OK, but for more advanced classes where sentence construction is longer and more complex, there’s a much higher chance of errors,” she said. A spokesperson for Duolingo said contractors were still involved in content creation work, particularly in language specialist roles. “We rigorously test all content, AI or human-generated, and there’s no evidence that AI leads to more mistakes,” they said. “Most learners can’t tell whether a sentence was written by a human or an AI and we’re not seeing any notable activity around subscription cancellations. “Our fun, distinctive style isn’t going anywhere, and it’s even part of how we prompt AI for content generation.” Duolingo employees have been using models like OpenAI’s GPT to create curriculums since 2023, but investments into “shared content systems” were also key to the ramp-up in course materials, Head of AI Klinton Bicknell said in an interview. That means that the content in say, an English class, will be the same regardless of what language you decide to learn it in. Previously, the English course for French speakers was developed separately from that for Spanish speakers, he said.