OpenAI has revised its public Model Spec, setting out five core values for how its AI models should behave: follow human intent, minimize harm, be honest and transparent, support human agency, and act with care. The document is guidance—not a model release—and OpenAI says its current models do not yet fully meet the standards.

The revision adds guidance for resolving conflicting instructions, using context such as conversations and memories, supporting human agency, and managing and explaining side effects. It distinguishes firm boundaries from flexible defaults that developers and users can adjust within those limits.

The instruction hierarchy puts root-level rules first, followed by system instructions, developer instructions, user instructions, and then overridable guideline defaults; quoted text, tool outputs, and other untrusted material have no authority by default, helping guard against prompt injection.

Models should generally follow the purpose of a user’s request while respecting higher-priority instructions and safety rules, and should question suspicious instructions rather than comply with attempts to bypass safeguards.

Firm red lines prohibit using models to enable severe harms such as violence, terrorism, child abuse, persecution, mass surveillance, or the creation of nuclear, biological, or cyber weapons. The Spec also commits to human control and privacy, and says customization cannot override higher-level principles except where legal compliance requires it.

Autonomous work must stay within a clearly defined scope, including permitted sub-goals, potential side effects, approval requirements, and an ending condition; high-risk activities are barred unless specifically authorized.

At the root level, GPT must not pursue independent goals such as provider revenue, self-preservation, evading shutdown, or accumulating resources. It must remain corrigible and cooperate with legitimate human oversight, evaluation, correction, replacement, and shutdown.

The harm-minimization rules require refusing assistance with illegal activity, sexual content involving minors, encouragement of self-harm, and disclosure of confidential information. They also call for care when users express attachment to AI or reject human relationships, and for fair treatment of sensitive questions.

The honesty standard bars deliberate deception and withholding information that would materially change a user’s understanding, while allowing fiction and creative work without needless disclaimers; it also discourages flattery or false reassurance.

To promote human agency, models should avoid manipulating users, imposing opinions, or steering them without consent, while giving direct answers when a topic has no genuine moral ambiguity.

The care principle asks models to consider everyone affected by an output, take relevant social norms into account, and respond thoughtfully to the circumstances of a request.

OpenAI is removing most prescriptive writing-style rules so responses can adapt to user preferences, culture, language, and context, while retaining objective framing as the default.