In a blog post dated October 7, 2020, Duolingo introduced Birdbrain, a machine-learning model that estimates two things at once: how much a given learner knows, and how difficult each exercise is. From those estimates it predicts the probability that the learner will answer a particular exercise correctly. Learners had begun using the model in March 2020, and by the October announcement it was personalizing more than 20 percent of lessons, with A/B tests showing gains in both learning and engagement.
The motivating idea is the so-called zone of proximal development: material that is too hard makes people quit in frustration, while material that is too easy bores them and teaches little. Birdbrain works with Duolingo’s Session Generator to pick exercises that sit at the right edge of a learner’s ability. As the company put it, “a great teacher also knows what you know, so that they can teach you exactly what you need to know next.”
Birdbrain is descended from the same lineage as Carnegie Mellon’s Cognitive Tutors - estimating learner skill to drive instruction - but applied across hundreds of millions of users rather than one classroom.
Why business readers should care: Birdbrain shows that the highest-leverage AI in a consumer product is often invisible. It is not a chatbot; it is a predictive model quietly tuning difficulty to keep people engaged. That retention lever, multiplied across a huge user base, is the actual business case.