Overview

AB 2013 gives training-data disclosure several distinct clocks: collection period, ongoing collection, first use in development, model training, and public availability to Californians, with later substantial modifications creating another release event. OpenAI reports collection beginning approximately in 2018, continuing, and first use in 2021. xAI reports ongoing collection since its founding in March 2023 and approximate training dates for named Grok versions. These dates anchor particular events; they do not identify full dataset composition, prove every collected item was used, or convert training into release. Preserve each date’s subject, verb, qualifier, and version context.

Learning outcomes

Collection, first use, model training, and public release are different events; one date cannot stand in for all four.
OpenAI reports an approximate 2018 collection start, ongoing collection, and first use in 2021.
xAI’s named Grok dates are approximate training anchors, not exact release dates or full dataset-composition records.

Keywords

AB 2013timelinecollectiontrainingreleasedates

Training Data Has a Timeline

California AB 2013 gives training-data transparency a temporal dimension. A developer must describe when data was collected, whether collection is ongoing, and when datasets were first used in development; public availability and later substantial modifications create a separate release timeline. This chapter keeps those clocks apart so readers can tell what a date actually anchors, how OpenAI and xAI describe their histories, and why a model’s training date does not reveal the full composition of its dataset.

Dates matter because training data is not necessarily a closed object. A provider may continue collecting data after a dataset was first used. A model may be trained at one point, fine-tuned later, and made publicly available at another. A substantial update can create another public event even when the provider describes it as a new version or feature release. If those events are compressed into one vague “training date,” the public account becomes easier to misunderstand.

Four clocks, not one

Section 3111 asks for two direct time facts about the datasets: the period during which the data was collected, including notice if collection is ongoing, and the dates the datasets were first used during development. Those are not the same event.

Collection period describes the period during which the data was collected. It can span months or years. Ongoing collection tells the reader that the period has not simply ended at the date of publication. First use in development identifies when the dataset began to participate in developing the system or service. A dataset can be collected in one period and used for development later, or be used while collection continues.

A fourth clock comes from the provider’s own technical description: when a named model was trained. xAI, for example, lists approximate training dates for named Grok versions. Those dates are useful evidence about the model timeline, but they are not automatically the same as the first-use date for every dataset, the end of data collection, or the date of public availability.

A fifth clock matters for the scope of the law: when the system or service is made publicly available to Californians for use. That public-availability event is the trigger described in §3111. It is related to development history, but it is not a synonym for collection or training. The statute’s recurring rule requires the documentation to be posted before each later covered system, service, or substantial modification is made available under the stated conditions.

The distinctions can be summarized without treating them as a rigid universal pipeline:

stateDiagram-v2 [*] --> Collection Collection --> OngoingCollection: Collection continues Collection --> FirstUse: Dataset first used in development OngoingCollection --> FirstUse FirstUse --> ModelTraining: Named model is trained ModelTraining --> PublicRelease: System or service is released PublicRelease --> SubstantialModification: Material functionality or performance change SubstantialModification --> FirstUse: New development cycle

The diagram’s value is causal, not merely chronological. Collection supplies the historical window. First use connects a dataset to development. Training ties development to a named model when the provider reports that fact. Public release creates the external event relevant to the statute. A material modification can send the system through another cycle.

Published by DeepMox

FAQ

Are a model’s training date and public release date the same?

No. A training date anchors when a named model began training, while public release is a separate event tied to availability under the statute; neither date automatically identifies collection or first use.

What does ongoing data collection mean for a training-data disclosure?

It means the collection process continues, so a published scale or history may describe a moving dataset rather than a fixed future-complete corpus. It does not prove that every newly collected item is used for training.

What dates does OpenAI report for its development data?

OpenAI reports that collection began approximately in 2018, continues, and that its datasets were first used for systems development in 2021. These statements do not provide a dataset-by-dataset first-use table or public-release date for each system.

Previous10 minAB 2013: Rights, Privacy, and Processing

Next12 minAB 2013: What Disclosure Cannot Prove

AB 2013: Training-Data Timeline

Training Data Has a Timeline

Four clocks, not one

FAQ