Tracing the Data Trail: A Survey of Data Provenance, Transparency and Traceability in LLMs

TL;DR

This survey addresses the opacity of training data in large language models by presenting a taxonomy of data provenance, transparency, and traceability, along with six supporting pillars: bias, uncertainty, data privacy, and provenance tools & techniques. Analyzing 95 publications, it identifies key methodologies across data generation, attribution, watermarking, data curation, and privacy, and discusses the trade-offs between openness and opacity. A central contribution is the taxonomy of domains and corresponding artifacts that structure end-to-end data flows in LLMs, enabling grounded discussion of open-weight versus closed-weight models, interpretability, and explainability. The work provides a roadmap for future governance-oriented research and practical tools to improve transparency, reproducibility, and accountability in data-driven AI systems.

Abstract

Large language models (LLMs) are deployed at scale, yet their training data life cycle remains opaque. This survey synthesizes research from the past ten years on three tightly coupled axes: (1) data provenance, (2) transparency, and (3) traceability, and three supporting pillars: (4) bias \& uncertainty, (5) data privacy, and (6) tools and techniques that operationalize them. A central contribution is a proposed taxonomy defining the field's domains and listing corresponding artifacts. Through analysis of 95 publications, this work identifies key methodologies concerning data generation, watermarking, bias measurement, data curation, data privacy, and the inherent trade-off between transparency and opacity.

Figures (5)

• Figure 1: Illustration of data provenance (Chapter \ref{['chap_dp']}), traceability (Chapter \ref{['chap_trace']}), and transparency (Chapter \ref{['chap_trans']}) linked to the data flow in LLMs. The supporting pillars i.e., bias, uncertainty (Chapter \ref{['chap_bias']}), and privacy (Chapter \ref{['chap_privacy']}) are directly linked to the training data sources, and provenance tools and techniques (Chapter \ref{['chap_tools']}) enhance the data cycle at various intersections.
• Figure 2: Taxonomy of our systematic literature review, outlining the three main axes - Data Provenance, Transparency, and Traceability - together with their supporting pillars. Bias & Uncertainty, Data Privacy, and Provenance Tools & Techniques provide additional context for characterizing data origins of LLMs.
• Figure 3: Illustration of the Scopus search, query construction, and manual reviewing process. The timeframe 2016 - 2025 spans over our three filters: Data, Language Model, and Attributes, leading to 129 search results.
• Figure 4: Overview of the considered articles and their domains.
• Figure 5: Differentiation between the key characteristics of open- and closed-weight LLMs, highlighting the difference in visibility, traceability, and data provenance.

Theorems & Definitions (2)

• definition 1: Interpretability
• definition 2: Explainability