How does Anthropic source training data?

Anthropic's stated sources#

Anthropic's privacy center lists four categories of training data that may contain personal data: publicly available internet information, datasets obtained under commercial agreements with third-party businesses, data provided by users or crowd workers, and data from users who allow their chats or coding sessions to be used to improve Claude.

Anthropic's Claude 4 System Card, cited in Stanford's Foundation Model Transparency Index report on Anthropic, states that Claude Opus 4 and Claude Sonnet 4 were trained on a proprietary mix of publicly available internet information, non-public third-party data, data from paid data-labeling contractors, opted-in user data, and internally generated data.

The ClaudeBot web crawler#

Anthropic operates ClaudeBot, a general-purpose web crawler for gathering public web content for model training, alongside separate bots for live search and user-directed retrieval. Anthropic says the crawler follows robots.txt instructions and does not access password-protected or paywalled pages.

Commercial licensing and contractors#

Stanford's Foundation Model Transparency Index research describes Anthropic as acquiring data through web crawling, licensing, human data-labeling and annotation contractors, and synthetic data generation, based on Anthropic's own system card disclosures.

Consumer data and training opt-in#

Anthropic states that for its Free, Pro and Max consumer plans, users can choose whether their conversations are used to train future models, while commercial plans such as Team, Enterprise and the API are not used for training unless a customer opts in through programs like the Development Partner Program.

What this means for a business#

A U.S. company weighing whether to supply Anthropic or any lab with historical records should treat it like any vendor contract: confirm in writing whether data trains models or stays isolated, strip personal and confidential details before sharing, and designate who internally approves what leaves the business.

SourceX and Anthropic#

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