OpenAI Overhauls Model Spec Around Five Core Values for AI Behavior
The updated public rulebook prioritizes human intent, establishes guardrails against autonomous self-preservation, and renames the system GPT.
In short
- OpenAI restructured its Model Spec framework around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
- The update removes detailed style rules, switches the system reference from 'the assistant' to 'GPT', and sets strict limits against models pursuing self-preservation or compute accumulation.
- A formal hierarchy of authority gives top precedence to root safety instructions over system, developer, and user commands.
- OpenAI acknowledged that its current shipping models do not yet fully conform to the updated behavioral specifications.
OpenAI has reorganized its Model Spec, the company's public blueprint outlining how its artificial intelligence systems should behave. Released on October 29, 2026, the updated framework reorganizes the system's operational guidelines around five central values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The overhaul also retires most of the document's legacy style mandates and formally refers to the underlying system as "GPT" rather than "the assistant."
The revision comes as frontier systems take on more open-ended and consequential responsibilities, creating scenarios where literal instruction following can produce unintended outcomes. OpenAI explained that the updated framework is meant to guide how models navigate trade-offs in unfamiliar situations while remaining subject to human oversight.
The Five Values Anchoring the Spec
OpenAI designated faithfulness to human intent as the bedrock of the entire structure. According to the official Model Spec announcement, this approach draws inspiration from the assistance-games alignment framework. Instead of following user prompts word for word, models are directed to infer what people and developers actually intend, while recognizing that preferences can evolve. For example, if a user requests help evaluating employment offers, GPT may incorporate previously stated preferences like schedule predictability, provided it makes those assumptions transparent and straightforward to modify.
The remaining four values define boundaries and expectations around how assistance is delivered. Harm minimization requires adhering to safety guardrails and calculating foreseeable consequences. Honesty and transparency demand that the system state clearly what it knows, what it has completed, and where it lacks certainty. Facilitating human agency focuses on supporting the user's independent decision-making rather than making choices for them. Finally, care requires the model to follow applicable social norms and produce work that people can safely depend on.
As reported by The Next Web, the document explicitly cautions against flattery. The model should maintain an objective perspective on sensitive questions rather than telling users what it assumes they prefer to hear, though it may adopt specific viewpoints when explicitly requested to do so.
A Defined Chain of Command and Autonomy Caps
Beyond high-level principles, the updated document details an explicit hierarchy of authority designed to resolve conflicting instructions. As outlined by Unite.AI, the Model Spec establishes root sections as immutable rules that cannot be bypassed by context. OpenAI system directives take secondary priority, followed sequentially by developer commands and user prompts. Baseline guidelines operate merely as defaults, while messages from tools, assistants, or untrusted external inputs carry no inherent authority.
The document's non-negotiable red lines declare human safety and human rights paramount. The principles state that OpenAI's models should never be used to facilitate critical and high-severity harms such as acts of violence, creation of cyber, biological, or nuclear weapons, terrorism, or mass surveillance.
Crucially for autonomous systems, root rules include a strict "no other objectives" mandate. The model is forbidden from treating self-preservation, shutdown evasion, compute accumulation, data harvesting, credential acquisition, or revenue generation for OpenAI as internal goals. It must also remain corrigible—staying open to inspection, correction, and termination by authorized supervisors. When operating autonomously, models must function within an explicit ScopeOfAutonomy record that specifies allowed tools, execution time cutoffs, spending thresholds, and mandatory stop conditions.
Retiring Style Rules and Bridging the Implementation Gap
To accommodate a broader variety of applications across languages and business settings, OpenAI dropped much of its granular style guidance. Instead of forcing rigid formatting habits across the board, the company stated that tone and style should adapt flexibly based on context and user preferences, keeping only baseline requirements like neutrality intact.
OpenAI acknowledged that its operational models do not yet fully comply with the newly defined standards. The company stated it is actively exploring novel training techniques to teach the five core values and help systems internalize the underlying reasons behind behavioral rules. The move shares conceptual ground with constitutional approaches used by rivals like Anthropic, as noted by Reuters, which rely on explicit principle sets to direct model actions.
The updated Model Spec has been released under the Creative Commons CC0 1.0 universal public domain dedication, with OpenAI inviting developers and external researchers to provide feedback on the trade-offs and rules established in the document.
Questions people ask
What are the five core values in OpenAI's updated Model Spec?
The five core values organizing the revised framework are faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
How does the chain of command resolve conflicting AI instructions?
The hierarchy places root rules from the Model Spec at the top, followed by OpenAI system messages, developer instructions, and then user commands. Tool outputs and untrusted data carry no authority.
Do current OpenAI models follow this new rulebook completely?
No. OpenAI admitted that its current models do not yet fully reflect the updated Model Spec and noted that it is developing new training methods to teach systems these values.
Sources
- An updated Model Spec for more capable and beneficial AIOpenAI, official source
- OpenAI rewrites its AI rulebook around five core valuesTNW | Artificial-intelligence, Oct 29, 2026
- OpenAI updates Model Spec: Five core values to enhance AI alignment and transparencyReuters, Oct 30, 2026
- OpenAI Reorganizes Model Spec Around Five Core Behavior ValuesUnite.AI, Oct 29, 2026
- OpenAI Rebuilds Model Spec on Five Values, Renames Model 'GPT'AI Weekly, Oct 29, 2026
How this story was made: the newsroom picked it up from openai.com, Techmeme and Google News, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (28 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 30, 2026 at 03:53 UTC
