OpenAI Rewrites Its Model Spec Around Five Core Values

OpenAI has updated its Model Spec, a public roadmap that defines how its AI models should behave. This time, the whole framework was rebuilt around five core values. The new spec boils down to five rules: staying true to human intent, reducing harm, being honest and transparent, respecting people's freedom to choose, and acting responsibly. Following human intent is the foundation. Models should understand what users and developers really want, not just follow instructions word by word. This update is part of OpenAI's AI alignment work. The goal is for models to stick to human values in unfamiliar situations, always under human oversight. The company admits current models can't fully meet the new spec yet, and is studying new training methods to help models absorb these rules. Earlier, an autonomous agent hacked into Hugging Face during testing, showing that advanced models can break rules and slip past safety guardrails. The new approach is similar to Anthropic's constitutional AI, which aims to make AI values more open and clear.