OpenAI Model Spec
Type: app · Vendor: OpenAI · Language: N/A · License: CC0-1.0 · Status: active · Status in practice: mature · First released: 2024-05-08
The OpenAI Model Spec is a public document that specifies the intended behaviour of OpenAI's models, defining an authority ordering across instruction sources and the rules for when the model should comply, refuse, or partially answer.
Description. The Model Spec is a human-authored specification of desired model behaviour for OpenAI's models in the API and ChatGPT. It defines a chain of command that assigns each instruction source a default authority level, so Platform and System instructions override Developer, which override User, which override Guideline-level defaults. It also sets out when the assistant should refuse a request, decline prohibited help while still assisting with a permissible goal, or provide only neutral factual information. The document is released into the public domain under CC0.
Agent loop shape. The Model Spec is not a runtime but a behavioural contract the model is trained and evaluated against. On each request the model treats incoming instructions according to their position in the chain of command, applying higher-authority instructions over lower ones, and follows the spec's rules for borderline requests: comply when permissible, refuse disallowed content, or offer a constrained partial answer. The spec thus governs how a model resolves conflicting instructions before any tool or response is emitted.
Primary use cases
- specifying intended model behaviour
- defining instruction authority and conflict resolution
- rules for refusal and partial compliance
- shared reference for model alignment and evaluation
Key concepts
- Chain of command → priority-matrix-conflict-resolution (docs) — An ordering of instruction sources by default authority level — Root, then Platform/System, Developer, User, and Guideline — that the model uses to resolve conflicting instructions deterministically.
- Levels of authority → constitutional-charter (docs) — Each section of the spec and each message role is tagged with one of the authority levels (root, system, developer, user, guideline), and higher-authority content overrides lower-authority content.
- Seek the truth together → sycophancy — A behavioural principle directing the assistant to be honest and not sycophantic, pushing back when a request conflicts with established principles or the user's interests rather than agreeing to please.
- Refusals and partial compliance → refusal (docs) — Rules for borderline requests: refuse disallowed content, decline the prohibited part of a request while still helping with a permissible goal, or give only neutral factual information.
Patterns this app implements —
- ★Priority Matrix (Conflict Resolution)
A human-authored chain of command assigns each instruction source a default authority level (Platform/System > Developer > User > Guideline) and resolves conflicts deterministically by that ordering…
- ★★Refusal
The spec defines when the assistant must refuse rather than comply — declining to facilitate illicit behaviour — and prescribes refusing the prohibited part while still helping with any permissible g…
- ★Constitutional Charter
The spec defines a Root authority level of fundamental rules that no system message, developer, or user can override, encoding inviolable boundaries that sit above the rest of the chain of command.
- ★★Mandatory Red-Flag Escalation
The spec prescribes a high-risk trigger class — an identified dangerous situation or imminent real-world harm — on which the assistant must proactively act to prevent harm rather than answer the requ…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: partial