Overview
California AB 2013 requires covered developers to post website documentation about data used to train generative-AI systems or services made publicly available to Californians, including after covered substantial modifications. This series first tests scope, then reads the required high-level summary as a connected schema covering provenance, purpose, scale, types, rights, acquisition, privacy, processing, synthetic data, and timing. OpenAI and xAI show how provider statements make development data more legible without creating a shared measurement system. The practical method preserves units, dates, qualifiers, and attribution, and treats disclosure as a public representation—not independent proof of legality, completeness, privacy, safety, quality, or performance.
Learning outcomes
- AB 2013 creates structured website visibility into covered development data, not a raw-corpus publication duty.
- Scope comes first: developer role, generative function, release timing, California public availability, and narrow exclusions all matter.
- The twelve categories connect provenance, purpose, scale, composition, rights, privacy, processing, synthetic data, and time.
- OpenAI and xAI show that provider disclosures can inform scrutiny without independently proving legality, completeness, safety, quality, or performance.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does California AB 2013 require AI developers to disclose?
It requires a covered developer to post website documentation about data used to train a covered generative-AI system or service, including a high-level summary across twelve specified areas. The duty applies at the statutory timing points for covered releases and substantial modifications made publicly available to Californians, subject to narrow exclusions.
Does AB 2013 require companies to publish their entire training corpus?
No. The cited provisions require a high-level summary of development datasets, not every underlying record, URL, file, model weight, or raw corpus. The summary still must address the categories named by the statute.
How should readers compare OpenAI and xAI training-data disclosures?
Compare the questions each provider addresses, but keep their units, ranges, dates, stages, and qualifiers separate. OpenAI reports broad source classes and modalities; xAI adds dynamic datasets, recipe stages, and approximate dates, without a common denominator for ranking.
What can a training-data disclosure prove?
It can establish what a provider publicly reports about its development data and make uncertainty more specific. It does not independently prove that the data is lawful, complete, private, safe, high-quality, or effective.