Anthropic is an AI safety and research company, building reliable, interpretable, and steerable AI systems.
We’ve prepared this policy (“Policy”) to explain how our large language models are ‘trained’ and how personal data may be used as part of the training process. We’ve also included information about the ways in which personal data of individuals who are not registered users of our services may incidentally be processed as part of our services, and the privacy rights such individuals may enjoy with respect to their personal data.
To understand how we collect and use information from registered users of our services, please see our Privacy Policy.
1. What Claude Is
Claude is a content-based generative AI system used by many people around the world to assist with a wide range of tasks. These include improving writing, providing coding assistance, offering strategic advice, conducting business analysis, summarizing documents, and helping with various daily activities. Claude is available to individuals, developers, and businesses around the world.
2. How Our Models are Trained
Large language models such as Claude are ‘trained’ on a variety of content such as text, images and multimedia so that they can learn the patterns and connections between words and/or content. This training is essential so that the model performs effectively and safely.
Training our models involves a number of separate stages, but broadly these fall into two steps: (1) pre-training and (2) post-training.
“Pre-training” involves the analysis of content in order for the model to understand the fundamentals of language – that is, syntax, the patterns and connections between words – as well as basic facts about the world. To understand the full diversity of language, the model is trained on large amounts of content (including text) that is converted into “tokens” (short chunks of text often just a few characters long, like common words, parts of longer words, or even punctuation marks). The model processes this content and builds a complex network of relationships between tokens. At the end of pre-training, the model can predict the most likely next word in a sequence based on the words that come before it, much like a sophisticated autocomplete feature.
Models do not store text like a database, nor do they simply “mash-up” or “collage” existing content. Models identify general patterns in text in order to help people create new content, and they do not have access to or pull from the original training data once the models have been trained.
“Post-training” helps make the model more useful, effective, and safe, by integrating human feedback. The model generates multiple responses to a given prompt, which are then evaluated by human reviewers based on criteria such as quality, helpfulness, safety, and accuracy. This reinforcement learning feedback loop helps align the model's outputs with human expectations and values, gradually improving a model’s ability to provide relevant and appropriate responses. In addition, further “fine-tuning” that provides the model with the examples of appropriate outputs (e.g., specific to a particular domain or use case) helps improve the model.
We’ve described the stages of model training in more detail in Section 4 below.
3. Collection of Personal Data
Anthropic may obtain personal data from the following sources in order to train our models:
- Publicly available information via the Internet. Our automated tool, ClaudeBot, collects content from web pages that are freely available to anyone browsing the internet. Online material is broad and varied - for example, websites, publicly shared documents, programming languages, and content such as image or video. We take steps to minimize the privacy impact on individuals as part of our data collection practices. For example, ClaudeBot does not access password protected pages or bypass CAPTCHA controls, and it respects ‘Do Not Crawl’ signals (robots.txt). For more information on how we collect publicly available data, see https://anthropic.com/crawl.