Three-layer moderation: the 2026 default deployment pattern
user prompt
Llama Guard 3/4
OpenAI Moderation
custom rules
finance disclaimer
medical refusal
model
GPT-5, Claude, Gemini
or open-source
output moderation
Llama Guard 3/4
OpenAI Moderation
category taxonomies
OpenAI Moderation (13 categories)
harassment(+threatening) / hate(+threatening)
self-harm(+intent,+instructions) / sexual(+minors)
violence(+graphic) / illicit(+violent)
Llama Guard 3 (14 categories)
MLCommons AILuminate taxonomy
+ code-interpreter-abuse, elections
+ defamation, specialized-advice
Perspective (toxicity baseline)
TOXICITY / SEVERE / INSULT
pre-LLM-era
still used in research
deprecations + migration
Azure Content Moderator: deprecated Feb 2024, retired Feb 2027.
replacement: Azure AI Content Safety (LLM-based, integrates with Azure OpenAI).
migration projects: 2024-2027 for affected deployments; factor into EU AI Act timeline.
omni-moderation-latest (2024): free for most developers; text + images; 42% better on multilingual.