A Tale of Two Philosophies: Duolingo vs Google Translate

Earlier this year, language-learning app Duolingo faced significant backlash over a botched “AI first” initiative. Using generative AI to create its lessons, users felt that this hurt the quality and lacked the human connection they hoped to see in a language app.

In contrast, to much less fanfare, Google Translate is testing a new feature, “Learn with AI”. It also relies on generative AI, but instead of using AI once to cheaply generate pre-made lessons, it uses it to create lessons dynamically to match the user’s skill level and needs. I tried out the Spanish feature and had conversations with the AI, booking a room in a hotel, asking where and when breakfast would be served, and even ordering a Margarita. It’s currently free, and if it’s available in your language, I encourage you to see for yourself how it works.

While it remains to be seen whether Google’s approach will revolutionize language learning, it already highlights an interesting philosophical difference:

• We can try and use AI to do more cheaply what we’re already doing

• We can try and use AI to radically do better what we’ve been doing so far

I doubt that AI can completely replace other, more formal, ways of language instruction (or the best way, which is total immersion). Still, a large language model’s ability to tailor responses right in the moment to its inputs has great promise: If the learning tool is well built, it can constantly keep the learner in the zone of optimal difficulty: Not too easy, not too hard. It can provide tailored feedback and, at scale and at cost, offer tailored grading. No more “fill in the gaps with one of these pre-selected words”.

Just in time for Canada’s Thanksgiving Weekend, I’m thankful for the potential that AI, when it’s well-done and in service of higher goals, can offer (and I’ll be back Tuesday)