Published disclosures and the datasheet discipline

Two documents do work that an internal data catalogue cannot. California AB 2013 forces training-data disclosure into public view. Datasheets for Datasets gives you a schema that engineers will actually complete. One is law with a deadline; the other is a convention with no enforcement and more practical value per hour spent.

California AB 2013: twelve disclosures, published

AB 2013 is operative from 1 January 2026 and reaches back: it applies to generative AI systems released on or after 1 January 2022. The developer publishes, on its own website, documentation about the data used to train the system. The enumerated disclosures cover the sources or owners of the datasets; whether datasets were purchased or licensed; whether the datasets include data protected by copyright, trademark or patent, or are in the public domain; whether the datasets include personal information; whether they include aggregate consumer information; the number and types of data points; the labels used; cleaning, processing or modification applied; the time period over which data was collected, including whether collection is ongoing; the dates of first use; and whether synthetic data generation was used in development.

Three practical consequences for a regulated firm. First, if you fine-tune and release a generative system, you may be a developer for these purposes and the disclosure is public, which is a very different review standard from an internal record. Second, retroactivity to 2022 means the record has to be assembled for systems built before anyone was keeping one. Third, and most useful: your vendors are now publishing this material, so a third-party assessment that does not read the supplier's published disclosure is leaving free evidence on the table.

Datasheets for Datasets: seven sections

The Gebru et al. schema (arXiv:1803.09010) organises a dataset record into seven sections: Motivation, Composition, Collection Process, Preprocessing, Uses, Distribution and Maintenance.

Adopt it rather than inventing a template, for two reasons. It is open and citable, so an auditor can check your structure against a published source. And it is organised around the questions someone building a dataset actually confronts, which is why it gets completed. Home-grown templates are organised around the regulation, which is why they get filled in with the word "N/A".

Map it once against Article 10 and AB 2013 and keep the mapping. Motivation and Uses carry the Article 10(2)(a) design choices and the intended purpose. Composition and Collection Process carry 10(2)(b), the AB 2013 source, licensing, copyright and personal information disclosures, and the collection period. Preprocessing carries 10(2)(c) and the AB 2013 cleaning disclosure. Distribution and Maintenance carry the questions nobody asks until the dataset is three years old: who may receive it, under what terms, who maintains it, and when it is retired.

The section everyone skips

Maintenance. A dataset is treated as a fixed artefact and then quietly appended to for years. Without a maintenance record you cannot answer the question that decides an incident investigation: which version of the data was behind the model version that produced the decision being complained about. That link between a deployed model version and the dataset versions behind it is the single most valuable thing in the whole record, and it is trivial to maintain prospectively and impossible retrospectively.

Worked example: a bank publishes and regrets the order of operations

A bank fine-tunes an open-weight model on its own correspondence to draft customer replies, and makes the assistant available through a subsidiary in California. Legal concludes the AB 2013 disclosure applies. The team assembles the twelve items and finds three problems, each of which should have been a design decision rather than a discovery.

The correspondence corpus contains personal information, which must be disclosed as such. Nobody had decided whether personal data belonged in a training set, because the question was never framed. The base model's training data is undisclosed by its provider, so the bank cannot answer the copyright and public domain questions for anything below its own fine-tuning layer. And a synthetic augmentation step was used to balance a rare complaint category, which is disclosable and which nobody outside the modelling team knew about.

The bank records "the provider does not disclose" for the base model rather than leaving those fields blank. That is the correct treatment: a refusal by an upstream supplier is a finding to record and to price into the vendor relationship, not a gap in your own document. It also adds the base model's undisclosed provenance to its third-party assessment as a named open risk, with a review date.

The order of operations lesson stands on its own. Had the datasheet been written at dataset creation, all three items would have been decisions taken by the right people in advance. Written at disclosure time, they were discoveries taken by whoever happened to be in the room, under a publication deadline.

As of 24 September 2026.