Data models are foundational to information processing, and in the digital world they stand in for the real world. When machines are used to make algorithmically-informed decisions, their algorithms are informed by the data models they use. And the data structured by data models is numerical of necessity, since machines must perform logical operations, and not creative interpretations. It follows that data used in machine operations are machine-language translations of real-world phenomena, expressed in a data model designed for efficient processing. It should not be surprising then that as information systems increasingly make decisions that affect people and communities, their operations are in a very direct sense an extension of the messy human world. This has resulted in information systems that reflect human racism, sexism, and many otherisms, with real-world harm to individuals and communities. But given the black-box nature of “machine learning” algorithms, how do we know what happens inside the black box? How can we document machine bias so as to design algorithms that don’t perpetuate social harms?