How does xAI source training data?

xAI's stated training data mix#

xAI's public training content summary for Grok 4.7 states that the model was pre-trained using a data recipe incorporating publicly available information, including scientific text, legal and official documents, X social media posts and source code, plus content from third parties, users, contractors, and internally generated data.

The same disclosure notes that training data was curated and filtered for quality and safety, and that image training data drew on publicly available sources including visual art and X social media images.

Use of X platform data#

xAI's Grok 4.7 and Grok 4.5 training content summaries both explicitly list X social media posts and images as part of the public data used for pretraining, reflecting xAI's relationship with the X platform.

Grok-1 open release#

xAI publicly released the weights and architecture of Grok-1, a 314 billion parameter model, in March 2024, describing it as a base model trained from scratch and concluding pretraining in October 2023, without a detailed breakdown of training sources in that announcement.

Corporate structure note#

xAI's EU AI Act filings identify the provider as SpaceXAI LLC, formerly known as xAI LLC, and xAI's consumer FAQ states that the company is legally separate from X Corp, even though it draws on X platform content.

What this means for a business#

A business posting content on X that wants to understand how it might be used for AI training should review xAI's published EU AI Act summaries directly, since those documents are the most specific public source on what categories of X content are included.

SourceX and xAI#

SourceX is not affiliated with xAI. This page summarizes public reports; it does not mean xAI buys data through SourceX.