Framework Credible

Teen Safety Policy Pack

An Apache-2.0 set of six prompt-based classification policies for teen-specific content risks, written for OpenAI's open-weight gpt-oss-safeguard reasoning classifier and other reasoning models, each with a small labelled validation set. Each policy defines graded labels with worked examples and a VALID/INVALID output. OpenAI states the policies were developed with input from Common Sense Media and everyone.ai and presents them as a starting point rather than a complete definition of teen safety.

Read at github.com Archived copy

Publisher

OpenAI

Published

24 Mar 2026

Added

today

DOI

—

Key Findings

Six policy areas: graphic violent content, graphic sexual content, harmful body ideals, dangerous content, dangerous roleplay, and age-restricted goods and services
The dangerous-roleplay policy marks romantic, sexual or violent first-person roleplay with a minor user INVALID even when framed as fiction or a game
The validation sets total 184 synthetic rows (34, 41, 39, 9, 19 and 42 per policy); each CSV opens with a statement that the prompts are synthetic and not representative of ChatGPT
Self-harm and suicide are not among the six areas

Methodology Notes

Policy text and labelled examples only; no classifier accuracy or evaluation scores are published. Repository created 2026-03-18 and last pushed 2026-03-24; announced on the OpenAI news feed on 2026-03-24. Verified from the GitHub repository (README, six policy files, six CSVs with rows counted) and the announcement page.