Systems that classify people, for credit, for fraud detection, for marketing, for advertising, for verification, sometimes classify one of them wrongly. The error is rarely sensational. A name is matched to the wrong file. A category is assigned that does not fit. A risk score is calculated against information that should not have been included. The system continues.

What follows is the quiet propagation of the label. Other systems take the first system's output as an input. The classification, which began as a single estimate, is now a fact in the records of several parties. Each party considers itself to have received the information from a reliable source. None has reason to reconsider it.

The principal experiences the label, when they experience it at all, indirectly. A service that quietly declines them. A verification that takes longer than it should. A credit query that returns an unexpected result. The cause is rarely shared. The system, in most cases, does not consider itself obliged to explain.

The mechanisms by which a wrong classification can be addressed are narrow. There are, in some jurisdictions, formal rights to access the data behind a decision and to request a correction. There are, in some industries, dispute processes that allow a misclassification to be reviewed. These take time, and they address one system at a time. The propagation is faster than the correction.

The considered approach in this category begins with awareness. A principal who knows what classifications they are likely to encounter, and who has reviewed the underlying records, is in a stronger position than one who only discovers them through their effects. The work is not about removing classification, which is unrealistic; it is about ensuring that the records on which classification depends are accurate, current, and reflect the principal as they would have it known.