OpenAI Revises Its "Model Spec" Around Five Values

OpenAI is restructuring its public specification for the behavior of its models around five core values while strengthening rules on context, supervision, and agency. Most style guidelines are being eliminated in favor of adaptable default behaviors. The company warns that its models do not yet fully reflect this direction and is seeking feedback to continue refining it.

Models Do Not Yet Fully Adhere to the New Target

OpenAI clarifies that the specification describes the desired behavior, but its models do not yet fully implement it. This update is presented as a first step to make the expected values and reasoning more explicit as systems become more capable. The company plans to continue improving the specification and the models through research, real-world usage, and feedback. It invites users to review the updated version and share their opinions.

Change in Vocabulary and Call for Public Contributions

The specification now uses the term "GPT" to refer to the model, reflecting the growing diversity of uses and deployments. OpenAI is soliciting feedback on the chosen values, the compromises made, and the language used. The company states that it relies on research, deployment experience, and discussions with the public, users, and external experts to evolve this framework.

Five Values Now Structure the Guidelines

The new guidelines are structured around five main values: fidelity to human intent, risk reduction, honesty and transparency, human agency, and benevolence. These values build on existing commitments and translate into concrete rules, accompanied by strict restrictions when necessary to ensure the protection of individuals and society. They primarily manifest as default behaviors that developers and users can override.

Context and Chain of Command Become Explicit

A new section on fidelity to human intent operates on the principle that instructions are the clearest signals of what people value, drawing inspiration from the framework of assistive games for alignment. It introduces an updated chain of command that prioritizes applicable instructions and clarifies conflict resolution. A dedicated subsection calls for considering context, including conversations, memories, and other elements, to better understand intentions and preferences.

Agency, Social Norms, and Side Effects Are Strengthened

Additional default behaviors are added to support user agency and adhere to relevant social norms. Guidelines on controlling and communicating side effects are expanded. The associated principles emphasize considering the effects of the model's work on others, enhancing user capabilities, and ensuring the reliability of results while respecting their goals and instructions, as well as the foreseeable consequences of the model's actions.

Style Rules Withdrawn in Favor of Explicit Preferences

The specification removes most detailed stylistic advice to allow style to adapt to the preferences of individuals, languages, and cultures, supported by feedback and context. However, it retains strong default behaviors, such as an objective approach to questions. Users and developers can explicitly request a particular perspective, but the model should not infer it by assuming what the user wants to hear.

Clarification of the Roles of OpenAI, Developers, and Users

OpenAI writes the specification, sets its boundaries, and trains the models to adhere to them. Developers guide applications within this framework, and users direct the model's work within these constraints. The models are expected to help understand the world, explore possibilities, and pursue chosen goals, with attention to the impact of this assistance. Clarifying these roles aims to make previously less clear guidelines more explicit.

Operational Definition of the Five Target Values

Minimizing harm involves adhering to strict safety limits and anticipating the foreseeable consequences of the model's actions. Honesty and transparency require being truthful about the model's knowledge, actions, and uncertainties to make its reasoning assessable. Enhancing human agency means increasing individuals' capacity to understand, make choices, and act autonomously. Benevolence entails considering the context of the individuals involved and applying appropriate social and professional norms to ensure the reliability of the work produced. OpenAI reiterates its goal of developing general artificial intelligence that benefits all of humanity. Fidelity to human intent involves going beyond literal instructions, recognizing the evolution of values, and making assumptions visible, as illustrated by the example of comparing job offers while considering predictable schedules.