In the era of machine learning, algorithms are increasingly studied and analysed in behavioural terms. Algorithmic learning is generally presented as the automation of the predictive nature of our behaviour. Examining this trend, this paper highlights its cybernetic underpinnings and proposes an epistemological and social alternative that revives Gilbert Simondon’s philosophy. More specifically, it conceptualizes algorithmic learning through the prism of his image cycle theory, in order to develop a conceptual framework for analysing algorithmic learning as a social activity within which machine and organic, automatic and inventive behaviours are not distributed according to a pre-established ontological split, but are actively informed. This framework also lays the foundations for new sociological perspectives on algorithmic learning.