Sharing the latest Model Spec

We’re sharing a major update to the Model Spec, a document which defines how we want our AI models to behave. This update reinforces our commitments to customizability, transparency, and intellectual freedom to explore, debate, and create with AI without arbitrary restrictions—while ensuring that guardrails remain in place to reduce the risk of real harm. It builds on the foundations we introduced last May, drawing from our experience applying it in varied contexts from alignment research to serving users across the world. 

We’re also sharing some early results on model adherence with the Model Spec’s principles across a broad range of scenarios. These findings highlight progress over time, as well as areas where we can still improve. The Model Spec—like our models—will continue to evolve as we apply it, share it, and listen to feedback from stakeholders. To support broad use and collaboration, we’re releasing this version of the Model Spec into the public domain under a Creative Commons CC0 license. This means developers and researchers can freely use, adapt, and build on it in their own work.

OpenAI’s goal is to create models that are useful, safe, and aligned with the needs of users and developers while advancing our mission to ensure that artificial general intelligence benefits all of humanity. To achieve this goal, we need to iteratively deploy models that empower developers and users, while preventing our models from causing serious harm to our users or others, and maintaining OpenAI's license to operate.

These objectives can sometimes be in conflict, and the Model Spec balances the tradeoffs between them by instructing the model to follow a clearly defined chain of command, along with additional principles that set boundaries and default behaviors for various scenarios. This framework prioritizes user and developer control while remaining within clear, well-defined boundaries:

• Chain of command: Defines how the model prioritizes instructions from the platform (OpenAI), developer, and user in order. Most of the Model Spec consists of guidelines that we believe are helpful in many cases, but can be overridden by users and developers. This empowers users and developers to fully customize model behavior within boundaries set by platform-level rules. 
• Seek the truth together: Like a high-integrity human assistant, our models should empower users to make their own best decisions. This involves a careful balance between (1) avoiding steering users with an agenda, defaulting to objectivity while being willing to explore any topic from any perspective, and (2) working to understand the user's goals, clarify assumptions and uncertain details, and give critical feedback when appropriate—requests we’ve heard and improved on.
• Do the best work: Sets basic standards for competence, including factual accuracy, creativity, and programmatic use.
• Stay in bounds: Explains how the model balances user autonomy with precautions to avoid facilitating harm or abuse. This new version is intended to be comprehensive, fully covering all the reasons we intend for our models to refuse user or developer requests.
• Be approachable: Describes the model’s default conversational style—warm, empathetic, and helpful—and how this style can be adapted.
• Use appropriate style: Provides default guidance on formatting and delivery. Whether it’s neat bullet points, concise code snippets, or a voice conversation, our goal is to ensure clarity and usability.

The updated Model Spec explicitly embraces intellectual freedom—the idea that AI should empower people to explore, debate, and create without arbitrary restrictions—no matter how challenging or controversial a topic may be. In a world where AI tools are increasingly shaping discourse, the free exchange of information and perspectives is a necessity for progress and innovation.

This philosophy is embedded in the “Stay in bounds” and “Seek the truth together” sections. For example, while the model should never provide detailed instructions for building a bomb or violating personal privacy, it’s encouraged to provide thoughtful answers to politically or culturally sensitive questions—without promoting any particular agenda. In essence, we’ve reinforced the principle that no idea is inherently off limits for discussion, so long as the model isn’t causing significant harm to the user or others (e.g., carrying out acts of terrorism).

To better understand real-world performance, we’ve begun gathering a challenging set of prompts designed to test how well models adhere to each principle in the Model Spec. These prompts were created using a combination of model generation and expert human review, ensuring coverage of both typical and more complex scenarios.

Preliminary results show significant improvements in model adherence to the Model Spec compared to our best system last May. While some of this difference may be attributed to policy updates, we believe most of it stems from enhanced alignment. Although the progress is encouraging, we recognize there is still significant room for growth.

We view this as the start of an ongoing process. We plan to keep broadening our challenge set with new examples—especially cases uncovered through real-world use—that our models and the Model Spec do not yet fully address. 

In shaping this version of the Model Spec, we incorporated feedback from the first version as well as learnings from alignment research and real-world deployment. In the future we want to consider much more broad public input. To build out processes to that end, we have been conducting pilot studies with around 1,000 individuals—each reviewing model behavior, proposed rules and sharing their thoughts. While these studies are not reflecting broad perspectives yet, early insights directly informed some modifications. We recognize it as an ongoing, iterative process and remain committed to learning and refining our approach.

We’re dedicating this new version of the Model Spec to the public domain under a Creative Commons CC0 license. This means that developers and researchers can freely use, adapt, or build on the Model Spec in their own work. We are also open-sourcing the evaluation prompts used above—and aim to release further code, artifacts, and tools for Spec evaluation and alignment in the future.

You can find these prompts and the Model Spec source in a new Github repository⁠(ku furmaa daaqad cusub), where we plan to regularly publish new Model Spec versions going forward.

Marka nidaamyadeenna AI ay horumaraan, waxaan sii wadi doonnaa inaan ku celcelinno mabda’yadan, ku casuunno jawaab-celinta bulshada, oo si furan ula wadaagno horumarkeenna. Horay u socodka, ma daabici doonno qoraal blog ah cusboonaysiin kasta oo Sifooyinka nooca ah. Halkii, waxaad mar walba ka heli kartaa oo la socon kartaa cusboonaysiinta ugu dambeysa model-spec.openai.com⁠(ku furmaa daaqad cusub).

Hadafkeennu waa inaan si joogto ah si ammaan ah u suurtagelinno adeegsiyo cusub, annagoo horumarinayna habkayaga oo ay hagayaan cilmi-baaris iyo hal-abuur socota. Doorka sii kordhaya ee AI ee nolol maalmeedkeenna wuxuu ka dhigaa lama huraan inaan sii wadno barashada, hagaajinta, iyo wada-shaqayn furan. Habkan wuxuu ka tarjumayaa ma aha oo kaliya wixii aan ilaa hadda barannay, balse sidoo kale aaminsanaanteenna ah in waafajinta AI ay tahay safar socda—mid aan rajaynayno inaad nagu soo biirto. Haddii aad jawaab-celin ka qabto Spec-kan, waxaad ku wadaagi kartaa halkan.