OpenAI has released an expanded version of its Model Spec, now detailing the expected behavior of AI models in a comprehensive 63-page document.

The new specifications aim for AI models to provide more honest feedback, functioning as a 'firm sounding board' rather than being overly agreeable.

These guidelines address the issue of 'AI sycophancy,' encouraging models to offer critical feedback instead of excessive agreement with users.

A significant change in the new specification is the approach to sensitive topics, where models are now expected to engage users in truth-seeking discussions.

The updated guidelines promote a reasoned analysis of controversial issues, such as wealth taxation, rather than avoidance.

OpenAI is also exploring a 'grown-up mode' for mature content, which would allow certain adult discussions while banning harmful material.

CEO Sam Altman previously indicated the development of this 'grown-up mode' to facilitate more mature interactions.

This update signals progress in AI safety and behavior standards, although it does not immediately change ChatGPT's functionality.

The updated document emphasizes user customization, transparency, and intellectual freedom, focusing on these three core principles.

OpenAI invites public input on the specification, which is released under a Creative Commons Zero license for industry-wide adoption and modification.

OpenAI remains committed to refining its models based on feedback received since the initial version launch in May 2024.

CEO Sam Altman announced the upcoming GPT-4.5 model, codenamed Orion, coinciding with this update.