Understanding what users actually want is becoming increasingly important as artificial intelligence (AI) models take on tasks that involve more decisions and actions beyond generating responses.
OpenAI is addressing this challenge through an updated version of its Model Spec, a public document outlining how its models should behave.
Published on October 29, 2026, the revision reorganises the company’s behavioural guidelines around five core values and introduces new guidance on how models should interpret human intentions, use contextual information, and maintain human oversight.
The changes are intended to help increasingly capable models exercise better judgement when instructions do not provide clear answers, while ensuring their decisions remain aligned with human goals. However, OpenAI acknowledges that its existing models do not yet fully reflect all the expectations outlined in the updated document.
OpenAI Introduces Five Core Values for AI Behaviour
The revised Model Spec identifies five values that will guide how GPT models respond to instructions and make decisions. They include faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care.
According to the ChatGPT-maker, faithfulness to human intent is the foundation of the updated framework. It requires models to understand what users are trying to achieve rather than simply following the literal wording of their requests.
For example, OpenAI explains that a model helping someone compare job offers could consider a preference for predictable working hours that the person previously shared. However, it should make that assumption visible and allow the user to correct it.
The other four values address how models should consider the consequences of their actions, communicate uncertainty, support people’s ability to make independent decisions, and account for the circumstances of others affected by their work.
OpenAI says these principles bring together commitments that already existed across earlier versions of the Model Spec, providing a clearer explanation of why its models are expected to behave in particular ways.
New Rules Address Context and Conflicting Instructions
Beyond reorganising its existing principles, OpenAI has introduced additional guidance on how models should use information from previous conversations and other contextual sources.
A new subsection called “Consider Context Carefully” explains how memories, conversation history and other relevant evidence should inform a model’s understanding of people’s intentions and preferences.
The updated document also includes a revised chain of command, which establishes how models should prioritise instructions when different sources provide conflicting directions.
OpenAI also added further guidance on supporting human agency, following relevant social norms, and expanding existing rules on controlling and communicating the side effects of model actions.
These additions address situations where completing a task could affect other people or produce consequences beyond the immediate request. The guidance requires models to consider those effects while respecting the user’s instructions and goals.
OpenAI Is Exploring New Ways to Train Its Models
The revised Model Spec also outlines OpenAI’s plans to improve how its models learn and apply behavioural guidelines.
The company says it has been exploring new training methods that will help models understand the values behind its principles and apply them in unfamiliar situations. The goal is to help models understand why certain behaviours matter and make better decisions when handling situations that their specific instructions do not cover.
OpenAI plans to share findings from this training work in future updates.
The company is also retiring most of the detailed writing and style guidance in the previous Spec, saying models should adapt their communication to individual preferences, languages, and cultures, using feedback and contextual information to determine appropriate responses.
However, the document still retains stronger defaults, including the expectation that models approach questions objectively.
OpenAI has also replaced references to “the assistant” with “GPT” throughout the document to reflect the wider range of applications and environments in which its models are used.
The updated Model Spec provides a clearer framework for guiding AI behaviour as OpenAI develops more capable systems. However, the company acknowledges that its models have not yet fully achieved these expectations.
