Abstract: Some philosophers and computer scientists, such as Anderson and Anderson (2007), Sullins (2019), and Railton (2020), have envisaged the possibility that AI can learn to be moral. Indeed, that it can be trained to be virtuous and develop artificial phronesis via machine learning from the bottom-up. In this paper, I examine this prospect from the perspective of Aristotelian virtue ethics and sketch some of the promise and challenges it involves. First, I outline the key commitments of the Aristotelian theory of moral learning and of a would-be moral machine learning framework. Second, I compare the key commitments of Aristotelian moral learning and moral machine learning and underline similarities and differences in cognitive modus operandi. Third, I note how these differences pose problems for AI developing human-like virtue and artificial phronesis (what Ι call Strong Moral AI).