In AI alignment, what is a constitution?

In AI alignment, what is a constitution?

• A. A post-training method that aligns LLMs to be helpful, honest, and harmless by using a written set of principles
• B. A legal document specifying training data rights
• C. A type of activator function in neural nets
• D. A dataset used for evaluating models

Answer: (A)

In AI alignment, a constitution is a written set of principles that guides a model’s behavior after training, with the aim of keeping it helpful, honest, and harmless. It acts as a governance framework the model can reference when deciding how to respond, often implemented through a system prompt or policy layer that constrains outputs and helps resolve ambiguous situations. This approach focuses on normative guidance rather than changing the model’s architecture or training data. It’s different from a legal document about data rights, a neural activator function, or a dataset used for evaluation. By codifying values in a constitution, developers can provide a stable, revisable standard that informs behavior across diverse tasks and scenarios, maintaining safety as understanding evolves. For example, it can require the model to avoid sharing private information, to refuse dangerous requests, and to be transparent about uncertainty.

In AI alignment, a constitution is a written set of principles that guides a model’s behavior after training, with the aim of keeping it helpful, honest, and harmless. It acts as a governance framework the model can reference when deciding how to respond, often implemented through a system prompt or policy layer that constrains outputs and helps resolve ambiguous situations. This approach focuses on normative guidance rather than changing the model’s architecture or training data.

It’s different from a legal document about data rights, a neural activator function, or a dataset used for evaluation. By codifying values in a constitution, developers can provide a stable, revisable standard that informs behavior across diverse tasks and scenarios, maintaining safety as understanding evolves. For example, it can require the model to avoid sharing private information, to refuse dangerous requests, and to be transparent about uncertainty.