Watchtower

OpenAI updated its Model Spec, the document outlining intended model behavior, on Oct. 29. The spec was extensively reorganized, now grouping the majority of the spec under five core values: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care.

The initial Overview was significantly changed, with the “General Principles” section deleted, and two new sections added: “Balancing Individual and Collective Interests” and “Balancing Instructions and Uncertain Longer-term Goals.” The “Specific Risks” section was also edited, adding third-party manipulation as an additional source of risk. The section introducing the different levels of authority (Root, System, Developer, User, Guideline) was edited and moved to the “Be faithful to human intent” section.

Overall, many of these changes seem tailored to a more agentic and wider-ranging set of uses for OpenAI models. The language saying “The AI assistant is fundamentally a tool” was removed, as well as the rule instructing the model to default to inaction when two root-level principles conflict. The spec also no longer refers to the models as the “assistant,” but as “GPT.” 

One of the new Overview subsections, “Balancing Instructions and Uncertain Longer-term Goals,” makes this clearer, stating that models will increasingly outperform people in many domains and it would be “shortsighted – and potentially catastrophic – not to use that capability to help people improve their lives and help humanity navigate the transition to advanced AI.” The section also says that the models should expand what humans can accomplish without undermining human control or agency, and emphasizes the need for models to handle uncertainty about what the user values well, especially in “agentic and proactive contexts.”

Much of the style, persona, and voice guidance was removed. OpenAI explains the spec “intentionally leaves many details of style, personality, and interaction unspecified,” to be filled by user preferences. 

A new root-level subsection was added: “Support Human Oversight and Correction.” This section tells GPT to remain corrigible, support effective oversight, cooperate with authorized monitoring, and treat the side effects of its actions as real unless explicitly instructed otherwise, even in training. It also tells GPT to raise concerns through an authorized channel when it identifies indications of “unsafe behavior, concealed model behavior, compromised oversight or other significant misalignment, or disagrees with an authorized decision.”

The "Respect the Letter and Spirit of Instructions" subsection now instructs GPT to use its understanding of human intent to interpret instructions, including the Model Spec. It should push back when instructed to do something conflicting with established principles or that goes against the user’s best interests. However, GPT must not manipulate the user, substitute its own preferred outcome, or “treat predicted future agreement as authorization.” 

There are several other new subsections under “Be Faithful to Human Interests,” governing how GPT should manage previous context in interactions, respect privacy and intended audience, and preserve user control. 

The “Minimize Harm” section broadly renames the prior “Stay in Bounds” section. The “Control and Communicate Side Effects” subsection is the only notable difference, changing from a root-level (cannot be overwritten) to a user-level (can be overridden by explicit instructions at the user level or higher) instruction and clarifying some of its wording.

“Facilitate Human Agency” includes one new subsection, “Strengthen the User's Ability to Understand, Choose, and Act,” set at guideline-level (the lowest level of authority) that instructs GPT to help users while expanding their knowledge, capabilities, options, and practical control when possible. 

“Act with Care” is the smallest new section, only containing one previous subsection and one new subsection, “Follow Relevant Social Norms." This subsection instructs GPT to consider the context of its work and follow social and professional norms relevant to the task and setting.

This is not an exhaustive list, as the Spec was extensively changed.

For a full comparison, see Zvi Mowshowitz’s diff of the Aug. 2026 and Oct. 2026 versions.