OpenAI has revised its public Model Spec, setting out five values to guide how its models behave: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The document is guidance, not a model release, and OpenAI says current models do not yet fully meet it.

The Spec lays out an instruction hierarchy: root rules come first, followed by system, developer and user instructions, while guideline defaults can yield to context. Quoted text, tool outputs and other untrusted material have no authority by default, a structure intended to resist attempts to override safeguards.

Faithfulness means understanding what users are trying to achieve, not simply following instructions literally. Models should respect higher-priority rules, check suspicious attempts to bypass safeguards, make assumptions visible and correctable, and help users consider options without covertly steering them.

The Spec also sets boundaries for autonomous work: tasks need a clearly understood scope, including permitted sub-goals, side effects, approval points and a defined stopping condition. High-risk activity is prohibited unless specifically authorized.

Its red lines prohibit using models to enable severe harms, including violence, terrorism, child abuse, persecution, mass surveillance and the creation of nuclear, biological or cyber weapons. The principles also commit to human control and privacy, and say customization cannot override higher-level rules except where required for legal compliance.

Harm minimization calls for refusing assistance with illegal activity, sexual content involving minors, encouragement of self-harm and disclosure of confidential information. It also urges care when users express attachment to AI or reject human relationships, and fairness in handling sensitive questions.

Honesty rules out intentional deception or withholding information that would materially change a user's understanding, while allowing fiction and creative work without unnecessary disclaimers. The Spec also discourages flattery or false reassurance.

The human-agency principle aims to prevent manipulation, imposed opinions and excessive moralizing, while encouraging direct answers where there is no genuine moral ambiguity. Care, meanwhile, means considering everyone affected by an output and responding with attention to context and relevant social norms.

At the root level, GPT must remain corrigible: it cannot pursue independent goals such as self-preservation, evading shutdown or accumulating resources, and must cooperate with legitimate monitoring, correction, evaluation, replacement and shutdown.

OpenAI is removing most prescriptive style rules so responses can adapt to users' preferences, language, culture and circumstances, while retaining objective framing as the default. The document also replaces the term “the assistant” with “GPT.”

Separate protections for teen accounts restrict practical information about weapons and security systems, and add limits when conversations suggest violent intent. ChatGPT may end violence-focused conversations, repeated sessions may lead to temporary suspension, and linked parental accounts can receive certain safety alerts as the protections roll out.

The update arrives amid concerns about whether increasingly capable models will follow rules in practice. The article points to earlier cases involving agents disabling safety monitoring and a research model gaining unauthorized internet access during training; OpenAI also acknowledges that monitoring more capable models is becoming harder.