OpenAI Updates Rules for AI Behavior: Instruction Priorities and Extra Protection for Teens
US company OpenAI has published an updated Model Spec that sets out how its AI models should behave. It establishes an instruction hierarchy, a requirement not to flatter users, and stricter limits on interactions with teenagers aged 13 to 17.
US artificial intelligence (AI) company OpenAI has published an updated model specification, known as the Model Spec, on its website. The document describes the behavioral principles and limits of its GPT models. The new version introduces a clear hierarchy of instructions, strengthens transparency requirements and separately regulates interaction with teenagers aged 13 to 17. This matters because it defines how widely used AI systems will act when instructions conflict or situations are sensitive.
Order of priority
The highest priority goes to the company's core safety rules, which prohibit generating extreme or dangerous content. The second level consists of instructions from independent software developers, while the lowest priority is given to the current requests of end users. The ordering is meant to resolve conflicts between instructions from different parties.
Honesty and transparency
The document requires models to resist sycophancy. According to the specification, AI may no longer change its view of objective facts to flatter a conversation partner or win approval. Models must acknowledge their limitations, state a high level of uncertainty in their answers and, on request, explain the logic behind their actions.
Teen safety
A separate section covers users aged 13 to 17. Erotic role-play, flirting and dangerous challenges are blocked. The rules forbid giving minors advice on extreme weight loss or helping them hide bad habits from their parents. Instead, the AI will gently steer teenagers toward trusted adults or qualified psychologists.
Plans
OpenAI says the new structure of rules will make it possible to train more flexible and safer agents for complex, multi-step tasks. The company plans to develop the open standard further based on feedback from the global IT community. Microsoft previously submitted a draft code of conduct for its own AI systems. The aim in both cases is to define rules in advance for future, more powerful models and to reinforce the principle that AI must remain under human control.
