NIST’s March 2022 publication, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, treats AI bias as a sociotechnical issue. Its analysis extends beyond the accuracy of a model to the institutions, people and practices involved in developing and using it. NIST Special Publication 1270
The publication discusses systemic, statistical and computational, and human biases. Systemic bias can arise from institutional or historical conditions. Statistical and computational bias can involve problems such as unrepresentative samples. Human biases concern patterns in human judgment. These categories show why improving one technical metric does not, by itself, resolve every source of unfairness.
The OECD’s accountability principle complements that analysis. It calls for accountability based on the actor’s role and context and for traceability of relevant datasets, processes and decisions throughout an AI system’s lifecycle. Traceability helps make it possible to examine how a system was built and used, rather than relying only on a final output. OECD principle
Together, the documents provide a vocabulary for examining bias and responsibility. They are reference materials, not results from a named university fellowship or evidence that a particular enterprise study demonstrated a quantified improvement.
Archive note: the earlier fellowship, authorship, sample and study-result claims were not substantiated.