Building on existing literature, our objective is to recommend definitions for key synthetic data terminology regarding its use cases in privacy preservation. By unpacking the words we use, we expose unsettled terminology and propose definitions which offer clarity and consistency, thus affording consensus to support increased adoption of synthetic data.

Taking into account the differing interests of creators, guardians and users of synthetic data, we applied a principled approach to selecting literature that could reveal insight on:

We examined a range of academic and grey literature and, on finding a lack of definitions and consistency, dug deeper to explore the differences in terminology depending on type and usefulness for different use cases.

Understanding the use case is key to creating accurate definitions. The terms ‘utility’ and ‘fidelity’ are separate but related concepts and feature commonly in literature, often interchangeably. Yet whilst ‘utility’ is dependent on context (is it right for this use case?), ‘fidelity’ is an inherent quality of the data. Different levels of fidelity can be important for some use cases, but increasing fidelity will not always increase utility. Furthermore, structural fidelity is different to statistical fidelity, yet this differentiation is rarely acknowledged. As such, terms related to fidelity remain ambiguous, while others, such as ‘microdata’ are simply assumed by the author to be understood by the reader with no definitions provided.

Inconsistent and undefined use of terminology leads to miscommunication, and consequently a lack of reliability and efficiency. It is important to solve this through wide stakeholder engagement for purposes of good governance and public trust. We recommend a non-static glossary where contested terms are recognised and clarity is formed.