OpenAI's new rules for ChatGPT: five values instead of a checklist

Short answer: On October 29, 2026, OpenAI published an updated Model Spec, the public rulebook that describes how its models, including the ones behind ChatGPT, are supposed to behave. It is now built around five values: faithfulness to what people actually mean, minimizing harm, honesty and transparency, human agency, and care. OpenAI says its models don't fully follow it yet, so treat it as a target, not a promise about today's ChatGPT.

What happened

OpenAI first published the Model Spec in 2024 and has revised it several times since. According to coverage of the October update by ShiftDelete and Crypto Briefing, the new version replaces many detailed rules with a short list of core values that the model should use when its instructions don't cover a situation:

• Faithfulness to human intent: understand what you are really asking for, not just the literal words.
• Minimizing harm.
• Honesty and transparency.
• Human agency: support your decisions instead of steering them.
• Care.

Other changes reported in the coverage:

• Context and memory. A new section tells models to "consider context carefully." Preferences the model remembers about you can shape its answers, but it should make those assumptions visible and easy to correct.
• Less telling you what you want to hear. Models should answer questions objectively and avoid flattery, unless you ask them to take a side.
• Fewer style rules. Most detailed style guidance was retired; tone should adapt to your preferences, language and culture.
• Acting on your behalf. There is more guidance on noticing and explaining side effects, which matters as AI agents start doing tasks for people.
• A new name inside the document. The model is now called "GPT" instead of "the assistant."

Crypto Briefing notes that this follows an August 2026 revision focused on how models should interact with teenagers.

Why people are talking about it

The Model Spec is one of the few places where an AI company writes down, in public, what it wants its chatbot to do when things get tricky. Moving from checklists to values is a bet that judgment generalizes better than rules. OpenAI researcher Jason Wolfe wrote that the revision "makes clearer what we want from models as they take on more autonomous work in the world." OpenAI is also asking the public for feedback on the values and wording.

What it means if you use AI

• Pushback is a feature. If a chatbot disagrees with you or points out a weak spot, that is closer to what this update asks for. Our explainer on why AI always agrees with you covers why flattery happens in the first place.
• You should be able to see and fix what it assumes. If ChatGPT seems to rely on something it remembers about you, you can ask what it assumed and correct it. Here's what ChatGPT memory saves about you.
• Rules on paper aren't behavior. OpenAI itself says its models don't fully reflect the Spec yet, so keep checking important answers.

Quick checklist

• Ask "what are you assuming about me?" when an answer feels oddly personal.
• If you want honest feedback, say so directly; if you want a one-sided argument, ask for that explicitly.
• Review your memory settings now and then, and delete anything you don't want used.
• For health, money or legal questions, still verify with a trusted source or professional.

This article reflects reporting up to November 1, 2026.

Cover photo: Steve Jurvetson, [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/), via Wikimedia Commons.