Overview

OpenAI and xAI answer the first four AB 2013 questions in different shapes. OpenAI reports broad source classes, intended purposes, and trillions of tokens across textual, image, audio, and audiovisual content. xAI describes Grok’s recipe with public Internet, third-party, user or contractor, and internal data, a range of trillions to tens of trillions of tokens that includes dynamic datasets, multiple development stages, and approximate dates for named versions. The statements support attributed comparison of disclosure shape, not a shared measurement, provider ranking, compliance finding, or proof that each purpose produces an outcome. A sound reading preserves each unit, timeframe, exception, and qualifier, then states what remains unreported.

Learning outcomes

OpenAI’s account emphasizes broad source classes, purposes, token scale, and modalities.
xAI’s account adds a Grok recipe, dynamic datasets, development stages, and approximate model-linked training dates.
The disclosures support comparison of questions and qualifiers, not a common denominator, provider ranking, or compliance finding.

Keywords

AB 2013OpenAIxAIGroktokenscomparison

Reading the Schema in Public: OpenAI and xAI as Test Cases

OpenAI and xAI offer two useful test cases for reading California AB 2013 in practice. Their public statements can be placed against the first four statutory questions—source, purpose, scale, and type—but they do not describe their data in identical units or categories. Comparing them carefully gives readers a concrete judgment method: identify what each provider actually says, preserve the provider’s qualifiers, and stop before a high-level statement is mistaken for a complete inventory or a compliance finding.

The comparison is valuable precisely because the documents are not interchangeable. OpenAI presents a compact account organized around broad source classes, purposes, scale, and modalities. xAI presents a recipe for Grok that adds development stages, dynamic datasets, and dates tied to named versions. Both make development more visible. Neither supplies a common measurement system that would support a simple ranking.

OpenAI: source classes joined to an intended purpose

OpenAI says it offers publicly available generative-AI systems in California and develops those systems using several kinds of data. The summary names publicly available data, data accessed through third-party partnerships, information that users or human trainers and researchers provide or generate, and synthetic data.

That is a source account. It tells the reader how OpenAI groups the origins of its development material, but it does not identify every dataset or every owner. The wording is broad enough to cover different collections and development activities. It is also more informative than the generic phrase “data from the internet,” because it distinguishes public availability from third-party access, user-related material, human-provided or generated information, and synthetic material.

OpenAI then supplies a purpose account. It says data helps its systems better understand human language and the world. It connects that understanding to enhancing human creativity, advancing scientific discovery and medical research, and enabling people to improve their daily lives. These are statements about intended use and desired function. They answer the statutory question of how the datasets are meant to further the system’s purpose, but they do not independently establish that a particular dataset produces any listed benefit.

This distinction is easy to lose when a provider’s purpose language is ambitious. A reader should record two separate observations:

OpenAI reports several broad source classes.
OpenAI describes those data as serving particular system-development purposes.

The second observation does not transform the first into proof that every source is relevant to every purpose. A summary can state the developer’s proposed connection without providing an outcome study for each dataset. AB 2013’s source-and-purpose pair is useful because it asks the developer to make that connection visible, not because it settles the connection independently.

OpenAI also reports datasets containing trillions of tokens of textual, image, audio, and audiovisual content. This statement does two jobs in the schema. “Trillions of tokens” communicates scale. The list of textual, image, audio, and audiovisual content communicates broad characteristics or modalities.

Again, the level of detail matters. The statement tells a reader that OpenAI describes development at very large scale and across multiple forms of content.

Published by DeepMox

FAQ

What does OpenAI report about its data sources and content modalities?

OpenAI reports publicly available data, data accessed through third-party partnerships, user or human-trainer and researcher material, and synthetic data, plus trillions of tokens across textual, image, audio, and audiovisual content.

What extra lifecycle detail does xAI provide for Grok?

xAI provides a multi-stage Grok recipe, mentions dynamic datasets with trillions to tens of trillions of tokens, and gives approximate training dates for named versions.

What does xAI’s “on or about” wording mean for Grok dates?

It means the dates are approximate provider-reported training dates, not exact release dates or a universal calendar for collection, first use, or public availability.

Previous9 minAB 2013: High-Level Data Schema

Next10 minAB 2013: Rights, Privacy, and Processing

OpenAI and xAI: AB 2013 Disclosures

Reading the Schema in Public: OpenAI and xAI as Test Cases

OpenAI: source classes joined to an intended purpose

FAQ