Constitutional AI

Implementation Tracker

What is Constitutional AI (By Anthropic)

This website aims to provide a comprehensive and up-to-date resource for individuals and organizations interested in tracking the implementation of AI technologies created and promoted by Anthropic and, more importantly, their (trademarked) invention called Constitutional AI. See Claude's Constitution.

Our focus is on tracking and reporting the various uses of Anthropic's Constitutional AI learning models, including the industries and applications in which they are utilized, as well as evaluating their performance and impact.

The site aims to serve as a hub of information and analysis for those interested in the potential and limitations of Constitutional AI, as well as its implications for the future of AI and its role in shaping our world. With the recent acquisition of Anthropic by Google, it is more important than ever to have a dedicated resource for tracking the developments and advancements in this field.

Definition: 'Constitutional AI' - as defined by Anthropic's Claude LLM, July 7th, 2023)

Constitutional AI refers to a set of techniques developed by researchers at Anthropic to align AI systems like myself with human values and make us helpful, harmless, and honest.

The key ideas behind Constitutional AI are:

- Aligning an AI's behavior with a "constitution" defined by human principles - things like avoiding harm, respecting preferences, and providing true information. This constitution shapes how the AI acts.

- Using techniques like self-supervision and adversarial training to have the AI learn to behave according to its constitution without explicit human labeling or oversight.

- Developing constrained optimization techniques so the AI pursues helpfulness under its constitutional constraints, rather than seeking open-ended goals or optimization.

- Designing the training data and model architecture to encode beneficial behaviors and avoid unsafe or deceptive behaviors.

The goal is to create AI systems that proactively want to be helpful, harmless, and honest. Constitutional AI aims to make AIs that align with ethics and human values without needing to be corralled or controlled through external systems. It's a technique focused on human benefit rather than pure capability.

Articles: Constitutional AI Updates:

This site is available for projects - the intent is to provide support and information about Constitutional AI and Anthropic's upcoming contributions to the AI industry.

Contact Russell M Wright