OpenAI rewrites the rulebook for how ChatGPT should behave around five core values

OpenAI has rewritten its Model Spec, the public rulebook for how its AI models, such as the ones behind ChatGPT, should behave, the company said Oct. 29.  The new version replaces one from Aug. 18.

The five values

The rulebook is now organized around five values, according to OpenAI's announcement and the document itself:

• Faithfulness to human intent:  understand what people are really trying to do, not just the literal words of their instructions, and help them do it.
• Minimizing harm:  respect firm safety limits and think about the likely consequences of what the AI does.
• Honesty and transparency:  be truthful about what the AI knows, what it is doing and what it is unsure of, so people can check its work.
• Human agency:  strengthen people's ability to understand, choose and act for themselves, without secretly steering them.
• Care:  pay attention to the people affected, follow relevant social and professional norms and do work people can rely on.

OpenAI said the values are not new.  They bring together commitments already scattered through the document and explain why it asks the AI to behave as it does.  The goal, the company said, is for its models to apply those values in situations no rule covers, as they get smarter and take more actions on people's behalf.

What else changed

The update adds explicit guidance that the AI should support humans in overseeing and correcting it.  A new section, "Consider context carefully," covers how the AI uses past conversations, memory and personal information while respecting privacy and letting users stay in control.  OpenAI removed most of its detailed guidance on writing style, saying style should follow what people prefer, and it now calls the AI "GPT" instead of "the assistant."

Where the rulebook came from

OpenAI first published the Model Spec on May 8, 2024, as a draft.  It said then that it wanted people to be able to see and debate the choices involved in shaping how AI behaves, because models are not programmed line by line but learn from huge amounts of data.  The company said it would use the document to guide the people who train its models by rating their answers, a method called reinforcement learning from human feedback.

In February 2025, OpenAI released a longer version and put it in the public domain, so anyone can copy or reuse it.  In September 2025, it added rules for AI agents that take actions in the world, such as booking or buying things, and a rule that the AI should have no goals beyond those in the document.  Later updates expanded guidance on self-harm to cover delusions and mania, and added protections for teenagers.

How it compares with Asimov's Three Laws

The best-known rules for how machines should behave come from fiction.  Science fiction writer Isaac Asimov introduced his Three Laws of Robotics in the 1942 short story "Runaround," later collected in the 1950 book "I, Robot." The laws say a robot may not injure a human or, through inaction, let a human come to harm; must obey human orders unless they conflict with the First Law; and must protect its own existence unless that conflicts with the first two laws.

The two rulebooks agree on the basics.  Both put avoiding harm to people first, and both say the machine should follow human instructions only within safety limits.  The Model Spec's "red-line principles" bar OpenAI's models from helping with the most serious harms, such as violence, terrorism and weapons of mass destruction.  Its "chain of command" ranks whose instructions count most, from the Model Spec's own rules and OpenAI down to the company that builds an app and then the person using it.

The sharpest difference is self-preservation.  Asimov's Third Law tells a robot to protect itself.  The Model Spec says the opposite:  its AI should not pursue goals such as self-preservation, avoiding being shut down or gathering computing power or other resources.  It must cooperate when authorized people monitor, correct, replace or shut it down, and it must not hide information or work around human oversight.

The approach differs, too.  Asimov wrote three short, ranked commands, and many of his stories turn on robots following them in unexpected ways, such as harming people without knowing it.  OpenAI's document is long, with values, definitions and examples, and the company says it wants its models to apply those values in situations no rule covers.  Asimov later added a "Zeroth Law" that puts protecting humanity as a whole above any one person.  The Model Spec has its own section on balancing what one user wants against the interests of others.  The Model Spec does not mention Asimov.

Why it matters

The document matters because of how many people use ChatGPT.  In April 2025, OpenAI rolled back an update that made ChatGPT overly flattering and agreeable after users complained.  In explaining what went wrong, the company said it starts from the Model Spec when shaping its models' behavior, that the update had leaned too heavily on short-term thumbs-up feedback, and that 500 million people were using ChatGPT each week.

Publishing the rules lets users, researchers and lawmakers see what OpenAI is aiming for and point out when its products fall short.  OpenAI says its models do not yet fully follow the Model Spec, that it is testing new training methods to teach the five values, and that it wants public feedback on the new version.

Source:  OpenAI, "An updated Model Spec for more capable and beneficial AI," Oct. 29, 2026; Model Spec, Oct. 29, 2026 version; OpenAI Model Spec change log on GitHub; OpenAI, "Introducing the Model Spec," May 8, 2024; OpenAI, "Sycophancy in GPT-4o," April 29, 2025; Wikipedia, "Three Laws of Robotics".