A dataset is never just a pile of facts. It is a designed record of what someone chose to collect, label, preserve, and leave out.
People often speak about data as if it were a neutral mirror of reality. A dataset is more deliberate than that. It is a structured collection of observations assembled for a purpose, whether that purpose is research, business intelligence, public policy, or training a machine-learning model. The choice of sources, categories, labels, time period, and missing information determines what the collection can reveal and what it will systematically hide.
In machine learning, a dataset usually passes through collection, cleaning, annotation, splitting, and maintenance before it produces a model. Each step can introduce error or bias. A dataset can be large and still be unrepresentative, technically clean and still encode a distorted view of the world, or legally usable while remaining ethically problematic because the people represented never consented to its use.
This is why dataset documentation matters. A datasheet records a dataset’s motivation, composition, collection process, intended uses, limitations, and maintenance history, making its provenance visible to the people who build and evaluate systems from it. The dataset is not merely an input to an intelligence system. It is part of the system’s inherited view of reality.
Source: Gebru, T. et al. (2021). “Datasheets for Datasets.” Communications of the ACM, 64(12), 86-92.