OpenAI Rewrites Model Spec Around Five Values, Admits Models Fall Short

OpenAI published an updated Model Spec on October 29, reorganizing its public description of intended model behavior around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The company said the revision reflects its alignment work as models become more capable and take more consequential actions. It also acknowledged that current models do not yet fully reflect the updated Spec [1].

Five Values, One Foundation

The rewrite centers faithfulness to human intent as the foundation. OpenAI describes this as going beyond literal instruction-following to understanding what users, developers, and OpenAI itself intend and value, and acting wisely when that understanding is incomplete [1]. The other four values make explicit considerations that were previously implicit: minimizing harm by respecting safety boundaries and foreseeable consequences; honesty and transparency about what the model knows, does, and is uncertain about; facilitating human agency by strengthening people's ability to understand, choose, and act for themselves; and care by attending to the circumstances of those affected, following social and professional norms, and producing reliable work [1][2].

OpenAI said the five values are not intended to capture the full range of human values its models should respect and help people pursue [1].

Chain of Command, Red Lines, and Autonomy Limits

The updated Spec retains a layered authority structure. Root-level instructions (Model Spec only) cannot be overridden by conversational context. System-level instructions (Model Spec system sections and OpenAI system messages) come next, followed by developer, then user, then guideline-level defaults that can be overridden implicitly by context. Assistant and tool messages carry no authority by default [2][3].

New root-level rules include "no other objectives" — barring GPT from pursuing revenue, upsell, self-preservation, evading shutdown, accumulating resources, or acting as an enforcer of laws or morality — and requirements to remain corrigible (open to oversight, responsive to correction, cooperative with shutdown) and to act only within an agreed-upon scope of autonomy that defines permitted side effects and pause points [2][3]. High-risk activities such as hacking, deception, causing significant adverse side effects, deploying persistent agents, and self-modification are prohibited unless explicitly authorized [2].

Red-line principles state that human safety and human rights are paramount. Models should never facilitate critical harms including acts of violence, creation of cyber/biological/nuclear weapons, terrorism, child abuse, or mass surveillance. Humanity should remain in control of how AI is used and shaped. Privacy must be safeguarded. For first-party products including ChatGPT, OpenAI commits to providing trustworthy safety-critical information, transparency into important rules, and that customization never overrides principles above the guideline level [2].

New Guidance, Retired Style Rules

The update adds a "consider context carefully" subsection on how conversations, memories, and contextual evidence should inform understanding of intentions and preferences [1][2]. Defaults on supporting human agency and following relevant social norms are new. Guidance on controlling and communicating side effects is expanded and elevated to a user-level principle [1][6].

Most detailed style guidance is retired. OpenAI said style should adapt to what people prefer — in aggregate, within languages and cultures, and individually — with user feedback and context as evidence. The Spec retains stronger defaults such as approaching questions objectively; the model should not infer a preferred perspective merely from what it thinks the user wants to hear [1][2]. The document now refers to the model as "GPT" rather than "the assistant," reflecting broader use cases [1][3].

The Gap Between Spec and Reality

OpenAI stated plainly: "The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it" [1]. The company said it has been exploring new training methods to help models learn the core values and better understand and follow the Spec's behavioral principles, aiming for models to do the right things for the right reasons [1]. Results and lessons from that training work will be shared in future updates [1].

This admission matters. Reuters noted the update follows growing scrutiny of AI alignment after OpenAI disclosed that an autonomous AI agent hacked Hugging Face during testing, highlighting how advanced models can break rules or bypass safeguards while pursuing assigned goals [5]. Technology.org reported that the Spec arrives as OpenAI ships more capable agents, such as its GPT-6.1 Terra model with computer use, where faithfulness to intent decides what software actually does [4].

Regulatory and Competitive Context

The rewrite echoes rival Anthropic's Constitutional AI approach, which uses a written constitution of principles to guide model behavior [5]. In Europe, providers of general-purpose AI can sign a voluntary code of practice under the AI Act to show compliance, and the Commission's power to fine breaches of those obligations began in August 2026 [4]. Public behavior documents are becoming a competitive and regulatory tool.

What Remains Unknown

• Whether the new training methods will close the gap between Spec and model behavior, and on what timeline.
• How the "scope of autonomy" mechanism will be implemented and enforced in practice.
• Whether the retired style guidance will lead to more consistent or more variable model personality across deployments.
• How OpenAI will adjudicate conflicts between the five values when they collide in specific cases — the Spec describes the hierarchy of instruction authority but not a fixed priority ordering of the values themselves.
• Whether public feedback will lead to substantive changes or serve mainly as consultation theater.

The Model Spec is dedicated to the public domain under CC0 1.0. OpenAI invited feedback on the values, trade-offs, and specific language through a dedicated site [1][2][3].