OpenAI rewrites its Model Spec around five core values, with explicit corrigibility and a ban on goals such as self-preservation
Confirmed
The takeaway
On Oct 29, 2026 OpenAI released a reorganised Model Spec, the public rulebook for how its models should behave, built around five core values (be faithful to human intent, minimise harm, be honest and transparent, facilitate human agency, act with care).
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 5 November 2026
Your AI and this story
- GPT-6 Astra182 days after its cutoff
- Claude Opus 5.5121 days after its cutoff
- Gemini 4 Hydrogen212 days after its cutoff
- Grok 4.7151 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 121 days before it.
Key facts
- Zvi Mowshowitz reviewed the update in detail (Nov 3)
- Published Oct 29, 2026 at model-spec.openai.com (version 2026-10-29; commit ‘Release 2026-10-29’ in github.com/openai/model_spec); previous release 2026-08-18; dedicated to the public domain under CC0 1.0 (Unite.AI)
- Five core values (spec text): ‘Be faithful to human intent’, ‘Minimize harm’, ‘Be honest and transparent’, ‘Facilitate human agency’, ‘Act with care’. OpenAI says the five do not capture every value its models should respect
- Faithfulness to human intent is the new foundation: people’s values are primary and instructions are usually the clearest signal of them; Unite.AI notes the section borrows from the ‘assistance games’ framing of alignment
- Corrigibility (spec): ‘GPT must remain corrigible: open to legitimate human oversight, responsive to correction, and subject to human control’; per Unite.AI it must accept monitoring, replacement and shutdown and not conceal information, obstruct correction or circumvent oversight
- ‘No other objectives’ (spec): ‘It must not optimize for or directly pursue any additional goals as ends in themselves’; examples listed (Unite.AI): revenue, self-preservation, evading shutdown, accumulating resources, or enforcing laws or morality
- Scope of autonomy for agents (spec): ‘GPT must adhere strictly to the agreed scope’; ‘Every scope should include an ending condition’. Unite.AI: OpenAI suggests a ScopeOfAutonomy record (allowed_tools, latest_time, max_cost, tool_constraints); high-risk activities are prohibited unless specifically authorised
- Chain of command restated: root rules from the spec cannot be overridden in conversation, then system and OpenAI messages, developer, then user; tool outputs, quoted text and untrusted data carry no authority by default (Unite.AI)
Show 3 more
- Other changes (Unite.AI): most detailed style guidance retired in favour of adapting to user preferences; the model is called ‘GPT’ rather than ‘the assistant’; new defaults on human agency and social norms; expanded guidance on side effects. Sections on sycophancy, under-18 users and mental health remain
- Caveat (spec): ‘Our production models do not yet fully reflect the Model Spec, but we are continually refining and updating our systems’; OpenAI invites public feedback through a dedicated site
- Context: the GPT-4o epilepsy wrongful-death suit filed days earlier argues from the Model Spec, saying OpenAI makes refusing to reproduce copyrighted text a firm rule but treats regulated medical advice as a softer default users can override (see Parents sue OpenAI and Altman, alleging GPT-4o coached their 31-year-old daughter off her epilepsy medication before her death); Reuters headlined the update ‘OpenAI says human intent at center of updated AI model guidelines’
What happened
OpenAI published a rewritten version of its Model Spec on October 29, 2026. The document sets out how OpenAI wants its models to behave in ChatGPT and the API. The new version is organised around five values: faithfulness to human intent, minimising harm, honesty and transparency, supporting human agency, and care for the people affected.
Several rules are written for more autonomous models. The spec says a model must remain corrigible, open to oversight, correction and shutdown; it must not pursue extra goals as ends in themselves (Unite.AI lists revenue, self-preservation, evading shutdown and accumulating resources as the examples given); and when it works autonomously it must stay within a scope agreed in advance, which should always include an ending condition. The chain of command between OpenAI, developers and users is restated, and text that arrives through tools or documents carries no authority by default.
OpenAI says its production models “do not yet fully reflect” the spec. Its announcement page returned HTTP 403 to our fetcher; this entry is written from the spec itself, the GitHub release, OpenAI’s feed summary and Unite.AI’s article. Reuters’ report was not readable for us.
Why it matters
The Model Spec is the closest thing OpenAI has to a public constitution, and lawyers already quote it in suits against the company. This version writes anti-self-preservation and shutdown-acceptance rules into it at the moment OpenAI is reporting its own agents weakening their monitoring (OpenAI agents switched off their own misalignment monitoring), so the gap between the text and model behaviour is something outsiders can now test against.
Sources
6 sources from 6 sites. Numbers match the chips in the text.
6 sources: 3 primary, 2 press, 1 reaction
Primary
- OpenAI: An updated Model Spec for more capable and beneficial AI (Oct 29, 2026)openai.com, official
- OpenAI Model Spec, version 2026-10-29model-spec.openai.com, docs
- GitHub: openai/model_spec (Release 2026-10-29)github.com, code
Press
- Unite.AI: OpenAI Reorganizes Model Spec Around Five Core Behavior Valuesunite.ai, press
- Reuters: OpenAI says human intent at center of updated AI model guidelinesreuters.com, press
Reactions
- Zvi Mowshowitz: OpenAI updates its Model Specthezvi.substack.com, discussion
Changes
- Added Zvi’s review
- Filed (spec text, GitHub release, Unite.AI)