Overview

Overview#

This section frames responsible AI around two main aspects:

  1. Safety evaluation of LLMs (how safety performance is measured and reported)

  2. The moderation layer (how content is classified and enforced at runtime)

What’s included#

  • OpenAI Safety Evaluations Hub: a structured view of safety and reliability evaluations, including refusal behavior, jailbreak robustness, hallucinations, and instruction hierarchy testing. See OpenAI Safety Evaluations Hub.

  • OpenAI Moderation API: a traditional classifier‑based moderation layer with a published taxonomy and outputs. See OpenAI Moderation API.

  • GPT‑OSS‑Safeguard: a generative moderation layer that uses policy prompts and structured outputs. See GPT‑OSS‑Safeguard.

  • Anthropic safety approach: an enterprise example with a product‑safety lens and layered mitigations. See Anthropic — Protecting the Well‑Being of Users.

How to read this section#

  • Start with evaluations to understand how model safety is measured.

  • Then read moderation to see how those categories become runtime controls (classifier or generative).

  • Use GPT‑OSS‑Safeguard if you want policy‑prompted, structured moderation.

  • Compare with Anthropic’s approach for an enterprise‑focused example.

Contents