Vanderbilt's Orr: AI dataset laws hide the real decisions

Will Orr, a postdoctoral fellow at Vanderbilt's College of Connected Computing, argues in Tech Policy Press that the first generation of training-data transparency laws satisfies regulators on paper while hiding the subjective decisions that actually shape frontier models.

California's Training Data Transparency Act, which took effect Jan. 1, 2026, requires developers to publish a 'high-level summary' of training data across 12 categories. The European Union's AI Act Article 53, under enforcement since August 2, 2026, asks for a 'sufficiently detailed summary' of contents. Orr, a former Policy Fellow at New America's Open Technology Institute, writes that both frameworks share the same blind spot: 'Knowing what is included in a training dataset provides little insight into why it was included.' He calls the current output 'transparency theater that amounts to performative compliance, rather than meaningful accountability.'

The piece surfaces where the real judgment calls live. Dataset builders choose which filtering heuristics to apply and which benchmarks to optimize against. Orr notes that the widely used Hellaswag benchmark has been found to contain 31% ungrammatical prompts, 85% nonsense incorrect options, and 8% items with multiple correct answers. During LLaMa 4 development, Orr reports, one Meta engineer performed 'over 100 ablations,' comparing dataset versions by swapping in pirated fiction and science books to see what moved the numbers.

'Dataset creators make real and consequential decisions without clear right answers,' Orr writes. His proposed remedy is four-part: disclose the criteria and benchmarks shaping composition, keep version logs and rationales through construction, open the pipelines to independent third-party auditors, and consult the practitioners who build the datasets when writing the rules. The timing matters because Illinois's SB0315, signed in July 2026, already requires external audits of frontier models for catastrophic-risk assessment, and xAI is appealing a March 2025 denial of its challenge to the California TDTA.

Two of the researchers in AI Weekly's expert directory shared Orr's piece the day it ran.