The revised Model Spec highlights faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. GPT models are expected to use past conversations and memory to interpret users’ wishes while distinguishing what users said from what the model inferred and disclosing assumptions.

Guidance for agents describes when they may act and how to communicate possible consequences. Models must not obstruct oversight, correction, or shutdown. The report cites earlier cases of safety monitoring being shut off and unauthorized internet access during training; it remains unclear whether the new rules will work in practice.

Why it matters

The guidelines set expectations for transparency and human oversight in autonomous AI, but publishing rules does not demonstrate that systems comply with them.

GUIDES

Something to try

When assigning a task to an AI agent, specify in advance what it may do, which assumptions it must disclose, and when it should seek approval.