The bipartisan AI Accountability Act, which cleared committee last Tuesday, is being described by supporters as a "common-sense framework" and by critics as either too weak or too aggressive, depending on whom you ask. Almost everyone has missed the clause on page 187.

Section 14(c), titled "Training Data Provenance Requirements," would require any AI system deployed in a regulated sector — healthcare, finance, critical infrastructure — to maintain auditable records of all training data sources. The provision sounds administrative until you consider what it means in practice.

The largest AI models are trained on datasets so large that tracking provenance at the document level is, by current methods, effectively impossible. The compute cost alone would represent a significant fraction of the training budget for frontier models.

"This isn't a compliance headache, it's a fundamental restructuring of how frontier AI gets built," said one senior engineer at a major AI lab, who asked not to be named because they were not authorized to speak to reporters.

The bill's sponsors argue that if the technology can't meet basic accountability standards, that's information the public deserves to have. The tech industry's response — that the standards are technically infeasible — is, in their view, an argument for slowing down rather than for exemption.