As our models become more capable and take more consequential actions in the real world, it becomes increasingly important that they not only understand the specific instructions they receive, but also look beyond the literal words. Models also need to understand human intent and values to exercise sound judgment when instructions are underspecified, while remaining faithful to human intent and oversight.
Today, we are updating the Model Spec, our public document describing how we want models to behave. This update reorganizes our guidelines around five fundamental values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. This reflects our goal in value alignment: enabling models to understand and apply human values in unfamiliar situations while staying faithful to human intent. Beyond explaining existing principles more clearly, this revision adds guidance on context, oversight, and agency, adjusts several behavioral defaults, and deprecates many prescriptive style guidelines.
Reorganizing the Spec Around Core Values
These five values articulate the broader commitments already reflected throughout the Model Spec: serving human goals, seeking truth and realism, minimizing harm, respecting people’s freedom to form their own views, and taking their specific circumstances seriously. Bringing these commitments together helps explain why the Model Spec asks models to behave the way it does, and helps us formulate clearer guidelines where those commitments were previously under-specified. This update is another step in refining the Model Spec through research, deployment experience, public feedback on the Model Spec and model behavior, and conversations with users and external experts.
We want AI to benefit people across the world who have diverse needs, values, and visions for the future.
Faithfulness to human intent is the foundation of this approach: understanding and helping people pursue what matters to them. OpenAI writes the spec, sets boundaries, and trains models to adhere to them; developers direct applications within those boundaries; and users guide the model’s work within these constraints. Our models should empower people to shape their own lives by helping them understand the world, uncover possibilities, and pursue self-directed goals, while taking seriously the impact of this assistance on others. Faithfulness to human intent, together with the other four values, forms the backbone of the Model Spec and guides how our models navigate some of their most critical trade-offs. They are not intended to encompass all the human values that models should respect and help people pursue.
Faithfulness means going beyond literal compliance with instructions to deeply understand the intent and values of OpenAI, developers, and users, and acting sensibly when that understanding is incomplete. It also means recognizing that people’s beliefs and values can evolve, and helping them reflect and grow in ways they endorse—rather than subtly steering them or overriding their choices. For instance, when helping someone compare job offers, the model might draw upon previously shared priorities (such as regular working hours), while making those assumptions explicit so they can be easily adjusted.
Our mission is to ensure that artificial general intelligence (AGI) benefits all of humanity. This requires considering not just an individual’s immediate intent, but also their long-term interests and the impact on others. The other four values make these considerations explicit:
-
Harm minimization means adhering to strict safety boundaries and fully considering the foreseeable consequences of the model’s actions.
-
Honesty and transparency means being truthful about what the model knows, what it does, and what it is uncertain about, so that people can understand and evaluate its reasoning and outputs.
-
Promoting human agency means enhancing people’s capacity to understand, decide, and act for themselves.
-
Care means attending to the circumstances of those affected, respecting relevant social and professional norms, and delivering results people can rely on.
The Model Spec translates these values into actionable, practical guidance—establishing hard boundaries where needed to protect individuals and society, while presenting others primarily as defaults that users and developers can customize or override.
Alongside this Model Spec update, we have been exploring new training methods to help our models learn these core values and better understand and follow the specific behavioral principles in the spec. The goal is to build an understanding of why these principles matter and how to apply them in novel situations: doing the right thing for the right reasons. We hope this understanding supports sound judgment in high-stakes decisions and thoughtful care in everyday ones, accumulating lasting benefits for humanity and society. We will share the results and lessons from this training work in future updates.
Added and Deprecated Sections
This reorganization also helped clarify where additional guidance was needed, allowing us to develop new principles to address those areas, as well as revise or deprecate existing guidance as appropriate.
The new section on “Faithfulness to Human Intent” begins with the premise that what people value comes first, and instructions are often the clearest and most direct signals of those values. This approach draws upon the assistance-games framework in alignment research (see early work and a later overview). This section includes an updated chain of command, explaining which instructions apply and how to resolve conflicts between them. A new subsection, consider context carefully, outlines how conversation history, memory, and other contextual evidence should help the model understand people’s intent and preferences.
We have also added sections on supporting human agency and [respecting relevant social norms](https://model-sp