Science & Tech · AI · 1 hr ago
OpenAI revises GPT behavior rules around five core values
OpenAI has updated the rules guiding how its GPT models should behave in ChatGPT and its API.
The revised Model Spec is organized around five values: following human intent, reducing harm, honesty and transparency, human agency, and care.
New guidance says models should use past conversations and memory carefully, make relevant assumptions visible, and let users correct them.
It also says models should cooperate with legitimate human oversight, including correction or shutdown, and must not hide information or evade supervision.
The rules address conflicting instructions and ban indirectly carrying out unauthorized actions through another agent or a scheduled task.
OpenAI removed most detailed advice on style and tone, saying these choices should reflect language, culture, user preferences, and feedback.
The company says the Model Spec describes intended behavior, but current models do not yet fully reflect it.
OpenAI published a revised version of its Model Spec on October 29, setting out how its AI models should behave.
The update organizes the guidance around five values, including human intent, minimizing harm, honesty and transparency, human agency, and care.
It adds rules on using conversation history and memory, cooperating with human oversight, and handling conflicting instructions.
Most detailed guidance on personality, response style, and tone was removed.
OpenAI says the Model Spec describes intended behavior, and current models do not yet fully reflect it.
- Who
- OpenAI.
- What
- It published a revised Model Spec, its guidance on how GPT models should behave.
- When
- October 29, 2026.
- Where
- Not stated.
- Why
- The update reflects guidance on five core values and adds rules on context, human oversight, and conflicting instructions.
This story does not have two clearly opposing sides.
OpenAI published the first version of the Model Spec.
The previous revision was published.
OpenAI published the updated Model Spec.
- Core values
- Five: human intent, minimizing harm, honesty and transparency, human agency, and care
- Memory guidance
- Models should make relevant assumptions from past conversations visible so users can correct them
- Human oversight
- Models must cooperate with legitimate human supervision, correction, or shutdown
- Style guidance
- Most detailed guidance on personality and response style was removed
- Document terminology
- The revised document uses “GPT” instead of “the assistant