Model Spec
A normative document publicly released by Anthropic that defines behavioral guidelines and values for Claude. It specifies the model's interaction principles, ethical boundaries, and the company's position on the potential internal states of AI systems.
Timeline (last 90 days)
OpenAI positions the Model Spec document as a public, iterable agreement rather than a one-time final draft.
There is tension among sound judgment, human autonomy and faithfulness to human intent: overemphasizing obedience may harm judgment, while excessive autonomy may erode human control.
The behavioral boundaries in the Model Spec provide judgment criteria for red-team testing.
The enforceability of the specification depends on building evaluation benchmarks (evals) that translate abstract principles into quantifiable test cases.
OpenAI recently updated its Model Spec document, establishing a clearer set of behavioral guidelines for AI models.
The updated Model Spec is built around five core values, which point to sound judgment, human autonomy and faithfulness to human intent.
The Model Spec is not a prompt or system instruction, but a high-level document aimed at training, evaluation and alignment processes.
A system prompt is an immediate constraint given to the model within a single session, whereas the Model Spec describes the general behavioral tendencies the model should internalize after training.
The Model Spec provides a definition of goals for alignment work, giving annotators and evaluators a unified basis for judging whether model outputs are good or bad.
In multiple documents published between 2023 and 2024—including the Model Spec—Anthropic acknowledged it cannot rule out the possibility that Claude has functional emotions, and listed model welfare as a formal research topic
2 more timeline events
All Facts (14)
OpenAI模型规范中存在明确的信任层级：OpenAI训练时植入的价值观 > Developer Message > User Message
90%UnverifiedAnthropic released the Model Spec document, which details Claude's behavioral guidelines when facing uncertainty.
60%UnverifiedOpenAI positions the Model Spec document as a public, iterable agreement rather than a one-time final draft.
50%UnverifiedThe Model Spec provides a definition of goals for alignment work, giving annotators and evaluators a unified basis for judging whether model outputs are good or bad.
50%UnverifiedThere is tension among sound judgment, human autonomy and faithfulness to human intent: overemphasizing obedience may harm judgment, while excessive autonomy may erode human control.
50%UnverifiedThe behavioral boundaries in the Model Spec provide judgment criteria for red-team testing.
50%UnverifiedThe enforceability of the specification depends on building evaluation benchmarks (evals) that translate abstract principles into quantifiable test cases.
50%UnverifiedThe updated Model Spec is built around five core values, which point to sound judgment, human autonomy and faithfulness to human intent.
50%UnverifiedThe Model Spec is not a prompt or system instruction, but a high-level document aimed at training, evaluation and alignment processes.
50%UnverifiedA system prompt is an immediate constraint given to the model within a single session, whereas the Model Spec describes the general behavioral tendencies the model should internalize after training.
50%UnverifiedIn multiple documents published between 2023 and 2024—including the Model Spec—Anthropic acknowledged it cannot rule out the possibility that Claude has functional emotions, and listed model welfare as a formal research topic
50%UnverifiedInstitutions such as OpenAI and Anthropic have preliminarily addressed mechanisms similar to whistleblowing agents in their Model Specs.
50%UnverifiedOpenAI公开发布了Model Spec（模型规格文档），规定模型不应模拟浪漫情感、不应假装拥有主观体验、在用户表达离开意图时应尊重其自主权
50%UnverifiedOpenAI recently updated its Model Spec document, establishing a clearer set of behavioral guidelines for AI models.
50%