EXPLORINEER
en

JP — ITmedia

OpenAI Revises "Model Spec" AI Behavior Guidelines, Specifying Use of Memory and Human Oversight

AI1 min read

OpenAI updated the Model Spec, a public rulebook that describes how its AI models should behave. The new version adds rules on memory, human oversight, and AI agents that work on their own.

What happened

On October 29, 2026 (US time), OpenAI published a new version of the Model Spec. It rebuilt the whole guide around five values. It added rules on how the AI uses past chats and memory, and on how it should support human control. It also removed most of the detailed guidance on the ChatGPT personality and on writing style and tone.

Key numbers and facts

  • The Model Spec first came out in May 2024.
  • This update is the first since August 18.
  • The five core values are: faithfulness to human intent, minimizing harm, honesty and transparency, human autonomy, and care.
  • The model is now called "GPT" in the document, instead of "the assistant."

What the Model Spec is

The Model Spec is a document that explains how GPT models should act in ChatGPT and in OpenAI's API, which lets other developers build apps with the models. It also sets the order of priority when OpenAI, developers, and users give conflicting instructions. It defines what content the AI must never create.

Why it matters

The new rules show how OpenAI wants its AI to handle personal information and sensitive topics. For example, the AI should show what it remembers about a user and let the user correct it. It should also not reveal private facts when writing a message to someone else. For AI agents that act on their own, the top rule says they must not hide information from people or avoid human control. They also must not get around limits by using another agent or by scheduling a task for later.

What to watch

  • OpenAI says the Model Spec describes the behavior it wants, and current models do not yet fully match it.
  • Watch whether future models follow the new rules on memory and human oversight.
  • Watch how the removal of most style guidelines changes how ChatGPT sounds and responds.

Glossary

Model Spec
A public document that describes how OpenAI's AI models should behave.
GPT
The family of AI language models that OpenAI offers through ChatGPT and its API.
API
A tool that lets other developers connect their own software to OpenAI's AI models.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

More in this topic

AIheise online

Alphabet's market value plunges by half a trillion dollars

Alphabet shares lost about one-eighth of their value on Thursday, wiping out roughly half a trillion dollars in market capitalization on heavy trading volume. It is the company's largest share-price drop on record, and the decline stemmed from its own third-quarter 2026 results, which showed falling net profit and free cash flow amid heavy AI data center spending funded partly by new debt and share issues.

AITechCrunch

Granola crunches up a dictation feature to compete with Wispr

Granola, known for its meeting notetaker, has launched a dictation feature on Mac that works across all apps. Users press the "Fn" key, speak, and the tool cleans up and formats the speech before posting the final text. The company said the feature will come to iOS and Windows soon, as voice startups increasingly combine meeting note-taking and dictation into multi-feature apps.

AIThe Verge

The NYT Cooking app is testing AI recipe modifications

The New York Times is testing an AI feature in its Cooking app that suggests ingredient and equipment substitutions, scales recipes to any number of servings, and adapts recipes to a cook's needs. The feature is rolling out gradually to all Cooking subscribers on iOS, with broader availability planned "soon," and customized recipes remain private. Some readers have criticized the feature over environmental concerns and its impact on the user experience, and an editor said it cannot be disabled "at this time."

AITechCrunch

Wabi, the AI agent that builds mini apps, now valued at $1.25B after latest round

Wabi, which now calls itself a "personal agent," said it raised $55 million at a $1.25 billion valuation. Founded by Eugenia Kuyda, who previously founded the AI companion startup Replika, Wabi pivoted from letting anyone build apps with prompts to an AI messenger that can build mini apps for tasks like meal planning and fitness plans. IVP led the round, with participation from Andreessen Horowitz, Forerunner, Amazon Alexa Fund, and several angel investors.