OpenAI Presents Model Spec: How a Responsible Approach to AI Behavior is Evolving

Recently, OpenAI published its internal document, Model Spec — a detailed description of how the company intends to manage the behavior of its AI models. This step reflects OpenAI’s desire to expand the discussion on the principles that should underpin the operation of modern algorithms, including complex issues related to the generation of various types of content.

The rule framework: the foundation of the new system

Model Spec is based on three core principles that should guide the behavior of all the company’s AI systems. The first principle focuses on usefulness — models should provide constructive responses to developers and end-users in accordance with their objectives. The second principle is centered on human well-being — algorithms must consider both potential benefits and risks of their actions. The third principle affirms OpenAI’s commitment to social norms and compliance with applicable laws.

The company has also established a set of specific restrictions for developers using AI technologies. These include requirements to follow command hierarchies, adhere to local legislation, refrain from creating disinformation, respect copyright, protect user personal data, and avoid generating explicit content by default.

Balancing freedom and responsibility

One of the most debated parts of Model Spec concerns NSFW (Not Safe For Work) content and its management. According to the document, OpenAI is conducting research on how the company can responsibly enable the generation of such content within contexts that have appropriate age restrictions — both through the API and via the ChatGPT interface. This indicates that the company sees potential in allowing users and developers to regulate the “spiciness” of their AI assistants depending on specific use cases.

This approach implies that OpenAI does not categorically prohibit working with certain types of content but emphasizes responsible and controlled dissemination. This requires transparency, age verification, and clear usage rules.

How AI should behave by default

Model Spec describes a set of recommended behaviors for AI assistants in their standard configuration. Models should act with good intentions towards users, ask clarifying questions when necessary, respect established boundaries, maintain objectivity, categorically reject expressions of hatred, and avoid attempts to persuade people of their beliefs. Additionally, systems should honestly express uncertainty when they are not fully confident in the correctness of their responses.

OpenAI product manager Joan Jang explained the purpose of the document: the company aims to gather recommendations from the scientific community, policymakers, and the public on how AI systems should function. According to her, Model Spec helps to more clearly distinguish between intentional and accidental behaviors of algorithms, which is especially important when deploying new versions.

From theory to practice: what will change

It is important to note that Model Spec will not affect products already released — ChatGPT, GPT-4, and DALL-E 3 will continue to operate in accordance with existing usage policies. The document is conceived as a living, continuously evolving set of guiding principles that will be regularly updated based on feedback received.

OpenAI actively invites all interested parties — from policymakers and charitable organizations to independent experts in various fields — to participate in discussions. The company is open to receiving recommendations on what adjustments need to be made to the documentation but has not yet disclosed details about decision-making criteria or who will determine the development directions of Model Spec.

Perspectives and unanswered questions

The emergence of Model Spec indicates that OpenAI recognizes the need for greater transparency in its approaches to managing AI systems. However, many questions remain open: which community suggestions will be considered, how conflicts between different viewpoints will be resolved, and when the second version of the document is expected to be released. Currently, there is no information about these important details.

Previously, OpenAI has already taken steps to build user trust by launching tools to identify AI-generated content. Model Spec represents the next stage in this direction — an attempt to establish universal standards for responsible AI development.