OpenAI reorganizes Model Spec around five values and commits to robots.txt

OpenAI has reorganized its Model Spec, its public description of how it wants its models to behave, around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The update, dated October 29, 2026, aims to help models apply human values in unfamiliar situations while staying faithful to what people intend. According to the company, the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults, and drops much of the Spec's detailed style guidance.

OpenAI defines faithfulness as more than following instructions literally. Models should understand what OpenAI, developers and users intend, and act wisely when that understanding is incomplete. The post's example involves a model helping someone compare job offers. It might draw on a priority the person has shared before, such as predictable hours, but it would make that assumption visible and easy to correct. The Spec's structure follows from this: OpenAI writes the Spec and sets boundaries, developers direct applications within them, and users direct the model's work within the constraints above that. The company says the values become firm boundaries where needed to protect people and society, and otherwise mostly act as defaults that users and developers can override.

The other four values get briefer treatment. Minimizing harm means respecting firm safety boundaries and accounting for the foreseeable consequences of actions. Honesty and transparency means being truthful about what the model knows, does and is uncertain about. Facilitating human agency means strengthening people's ability to understand, choose and act for themselves. Care means attending to the circumstances of those affected and doing work people can rely on. OpenAI presents these as a summary of commitments the Spec already contained.

The Turing Post newsletter calls all four of those remaining values equally aspirational. Its criticism is easier to test against the post itself. OpenAI says the values are not meant to capture the full range of human values models should respect, and it states that its models do not yet fully reflect the Spec. The company says it has been exploring new training methods to help its models learn these values, but it promises to share results and lessons in future updates. So the post offers no evidence yet that training follows the new structure.

Among the concrete changes, the post adds a section on faithfulness that borrows from the assistance-games framing of alignment. It also includes an updated chain of command that explains which instructions apply and how conflicts are resolved, plus a subsection on how conversations, memories and other context should inform what a model infers about people's intentions. New defaults cover supporting human agency and following social norms. The Spec now refers to the model as GPT rather than the assistant, reflecting a wider range of use cases. Most detailed style guidance is retired, on the reasoning that style should follow user preferences, though the Spec keeps stronger defaults such as approaching questions objectively.

Turing Post also reports that OpenAI committed to honoring a shared robots.txt signal for publishers' AI training preferences. It links to a Training preferences page.