Archives are assembled for humans to keep and explore while datasets are intended for machines to process and produce. Arguably, archives are created under a historical impulse; they are organised according to the record- keeping needs of the cultures that build them. This historical impulse requires a high degree of structuring, usually in the form of hierarchical ontologies that facilitate cataloguing and retrieval, and that aspire to a certain degree of historical accuracy, integrity, and permanence. These are some of the characteristics that we tend to expect from archives. Datasets also respond to the sense-making needs of the cultures that build them, however they come together under a different impulse; they tend to be contingent assemblages, made in response to specific a priori needs and questions relevant to engineering and the sciences. In data science and machine learning engineering, for example, datasets tend to be granular, flattened to matrix- like structures whose individual items are not meant to be publicly accessible or even individually meaningful to human observers. Instead, data in these systems is much more mercurial, f lowing from one repository to another through interconnected networks, and constantly changing shape depending on the specific applications for which it is intended. And because contemporary AI applications rely on vast amounts of data, datasets made for machine processing can be significantly larger and are usually not limited to canonical artefacts, or a small collection of pieces deemed significant enough to go ‘on the record’.