AI Research Atlas

Model Spec (draft)

OpenAI · 8 May 2024

OpenAI publishes a draft Model Spec, which sets out written objectives, rules and default behaviors for its models, with developer instructions outranking user ones.

An early public behavior specification from a frontier lab (Anthropic's published constitution is a related precedent), used to guide RLHF and floated as possible direct training input. Builds on the April 2024 instruction-hierarchy paper.

Date
Wednesday, 8 May 2024
Lab
OpenAI
Kind
paper
Access
research preview

Source is an expert summary of OpenAI's post; the Model Spec has been revised several times since.

Sources

  1. simonwillison.net/2024/May/8/model-spec/
  2. simonwillison.net/2024/Apr/23/the-instruction-hierarchy/

This record was checked against its sources on 6 October 2026. How we check