Traceability for Trustworthy AI: A Review of Models and Tools

摘要

Traceability is considered a key requirement for trustworthy artificial intelligence (AI), related to the need to maintain a complete account of the provenance of data, processes, and artifacts involved in the production of an AI model. Traceability in AI shares part of its scope with general purpose recommendations for provenance as W3C PROV, and it is also supported to different extents by specific tools used by practitioners as part of their efforts in making data analytic processes reproducible or repeatable. Here, we review relevant tools, practices, and data models for traceability in their connection to building AI models and systems. We also propose some minimal requirements to consider a model traceable according to the assessment list of the High-Level Expert Group on AI. Our review shows how, although a good number of reproducibility tools are available, a common approach is currently lacking, together with the need for shared semantics. Besides, we have detected that some tools have either not achieved full maturity, or are already falling into obsolescence or in a state of near abandonment by its developers, which might compromise the reproducibility of the research trusted to them.