The Frontier AI Data Accountability and Training Transparency Act

By Achuthan Panikath

Mon Aug 03 2026

The question has changed For most of the internet's history, the legal question about web data was narrow: may this page be technically accessed and copied? Artificial intelligence has made that question obsolete. Frontier AI developers now train models on corpora assembled from billions of documents: websites, books, code repositories, images, transcripts, and licensed and unlicensed collections of every kind. The scale converts an old question about access into a new question about accountability. When information is extracted at industrial scale and transformed into commercial systems worth billions of dollars, should the extraction carry obligations of transparency, provenance, and respect for the legal restrictions attached to the source material?

Today, largely, it does not. The public record of what the major models were trained on has grown thinner as the commercial stakes have grown larger. Early research models published detailed dataset documentation; current frontier models frequently disclose almost nothing. Rights holders who believe their work was used cannot verify it. Courts adjudicating the wave of copyright litigation against AI developers must reconstruct training practices through discovery. Regulators evaluating claims about data practices have no records to examine. The market operates on an asymmetry of information so complete that accountability is structurally impossible.

This Act would not resolve the copyright question. It would make the copyright question, and every adjacent question, answerable.

Deliberately not a copyright verdict The temptation in this field is to legislate the conclusion: either that all AI training on unlicensed material is infringement, or that all of it is fair use under Section 107 of the Copyright Act. Both conclusions would be wrong as blanket rules, and the courts are actively working the boundary case by case. Training on licensed data differs from training on pirated collections; transformative research differs from market substitution; and early decisions have already begun distinguishing lawful acquisition from unlawful acquisition even within a single case.

The Act takes the position that the legal system can only draw those distinctions if the facts exist somewhere. Its purpose is to guarantee that they do: a transparency and accountability framework allowing courts, regulators, and rights holders to distinguish lawful training from unlawful acquisition, without prejudging which is which.

The Training Data Provenance Ledger Covered frontier developers, defined by compute thresholds and commercial scale so that the obligations reach the largest actors and not academic laboratories or startups, would maintain an internal Training Data Provenance Ledger documenting every significant training-data acquisition. For each material source, the ledger would record the category of source; the method of acquisition, whether licensed, scraped, purchased, or generated; the collection period; licensing status and terms where applicable; whether the source had expressed machine-readable opt-out signals and whether they were honored; the identity and behavior of any crawler used; the provenance of any third-party dataset incorporated; material transformations applied such as filtering and deduplication; and any known legal restrictions attached to the source.

From the ledger, developers would publish a standardized public summary: detailed enough that the public and rights holders can understand, at the level of domains and dataset families, what a model was trained on, without exposing trade secrets, personal information, or the proprietary recipes that distinguish one developer's data pipeline from another's.

The Act would also require covered developers to honor legally valid data-use restrictions, including contractual terms and standardized machine-readable opt-outs, and to operate a mechanism through which rights holders can inquire whether identified works were used and challenge unauthorized use.

The European precedent This architecture is not speculative. The European Union's Artificial Intelligence Act already requires providers of general-purpose AI models to maintain copyright-compliance policies and to publish sufficiently detailed summaries of training content, and the European Commission's implementing template requests information about datasets and scraped online sources, including crawler identities and domain-level information. Every major American developer serving the European market is already building the capability this Act would require. The choice before Congress is not whether frontier training transparency will exist; it is whether American law will shape it, or whether American companies will practice in Europe an accountability they are never asked to practice at home.

Enforcement The Federal Trade Commission (FTC) would hold primary enforcement authority under its existing unfair-and-deceptive-practices jurisdiction, extended and specified by the Act. The Commission could investigate and penalize deceptive public claims about training data; failure to honor contractual restrictions or valid opt-outs; fraudulent or knowingly false provenance records; and data acquisition the developer knew to be unlawful, such as knowing ingestion of pirated collections. The Copyright Office would administer a rights-holder registry and, with the National Institute of Standards and Technology, maintain the technical standards for machine-readable reservations of rights, giving opt-outs a stable technical meaning.

Disclosure operates in tiers, which is the answer to the inevitable trade-secret objection: a public summary for everyone; detailed ledger records available to regulators under investigation authority; and the most sensitive material producible under protective order in litigation. Developers keep their weights, their architectures, and their pipeline engineering. What they surrender is the ability to say, to courts and rights holders alike, that no one can know what the model was trained on.

The compliance-burden objection Developers will argue that provenance documentation at corpus scale is burdensome. Two answers. First, the thresholds confine the Act to organizations spending hundreds of millions of dollars on training runs, for whom documentation is a rounding error and, under European law, an existing obligation. Second, and more fundamentally: an organization that cannot say where its training data came from is not describing a burden, it is describing the problem. Provenance ignorance at that scale is a governance failure that the market's largest actors should not be permitted to convert into a legal shield.

Summary of the proposal Require covered frontier developers to maintain a Training Data Provenance Ledger documenting acquisition method, licensing status, opt-out handling, crawler identity, dataset provenance, and known restrictions; publish standardized public summaries; honor valid machine-readable and contractual data restrictions; and operate a rights-holder inquiry mechanism. Enforce through the FTC with tiered disclosure protecting trade secrets, supported by a Copyright Office rights-holder registry and NIST technical standards.

Key references 17 U.S.C. § 107 (fair use). Federal Trade Commission Act, 15 U.S.C. § 45. European Union Artificial Intelligence Act, general-purpose AI provisions, and the European Commission's training-content transparency template. Ongoing federal copyright litigation concerning AI training data, and dataset-documentation scholarship including datasheets and data-statement frameworks.