In Applied AI, or 'machine learning', methods such as neural networks are used to train computers to perform tasks without human intervention. In this article, we question the applicability of these methods to education. In particular, we consider a case of recent attempts from data scientists to add AI elements to a handful of online learning environments, such as Khan Academy and the ASSISTments intelligent tutoring system. Drawing on Science and Technology Studies (STS), we provide a detailed examination of the scholarly work carried out by several data scientists around the use of 'deep learning' to predict aspects of educational performance. This approach draws attention to relations between various (problematic) units of analysis: flawed data, partially incomprehensible computational methods, narrow forms of 'educational' knowledge baked into the online environments, and a reductionist discourse of data science with evident economic ramifications. These relations can be framed ethnographically as a 'controversy' that casts doubts on AI as an objective scientific endeavour, whilst illuminating the confusions, the disagreements and the economic interests that surround its implementations.