A written document in which an AI developer sets out how it wants its models to behave: what they should help with, what they should decline, and how to weigh goals such as helpfulness, honesty and safety when they conflict. It serves as a target for training and a public statement people can hold the developer to, though it describes intended behaviour and does not guarantee the model follows it.
Technical view
OpenAI published a first Model Spec in May 2024, organised into objectives, rules and default behaviours, with a chain of command for when developer and user instructions conflict; Anthropic publishes a constitution for its Claude models that plays a similar role. Specifications guide human feedback guidelines and synthetic training data, and some developers train models to reason over the specification directly. They also give evaluators a reference for deciding whether an output is a bug or intended. Weaknesses are that natural language leaves ambiguity, conflicting principles need case-by-case judgement, and conformance has to be measured empirically rather than assumed from the document.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.