Provenance Withheld: Published Model Values and Undisclosed Training-Data Lineage at Frontier AI Developers
Year: 2026
Type: preprint
Abstract: A model's values can be published without its sources. Frontier AI developers release extensive charters demanding honesty and transparency from their models, yet describe their own training corpora in broad categories: web scrapes, licensed material, and synthetic data. While California's AB 2013 and the EU AI Act mandate disclosure, corporate filings identify few datasets and no web domains. Conversely, when internal models serve as sources, developers document outbound lineage with forensic precision—tracing millions of distillation queries to rival researchers and capturing chain-of-thought traces. The most granular account of Anthropic's own corpus emerged only through litigation, where a federal court found millions of pirated books held in a permanent repository, and internal data mixes were guarded as trade-secret recipes. Frontier developers thus enforce a standard of self-disclosure for model behavior and outbound distillation that they systematically withhold from inbound training provenance. Keywords: Training Data Transparency, Data Provenance, Data Lineage, AI Constitution, Model Specification, System Cards, California AB 2013, EU AI Act Article 53, Public Summary of Training Content, Model Distillation, Reasoning Models, Chain-of-Thought Data, Synthetic Data, Bartz v. Anthropic, Trade Secrets, Foundation Model Transparency, Digital Provenance, GPAI Compliance, Inbound Lineage, Outbound Distillation
Authors: Jefferson, Josie, Velasco, Felix
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235174369
