Glossary term
Glossary term
Safety and Alignment
Any software or process that prevents harm to humans or systems. Harm can take many forms, including preventing data leaks or unauthorized access, or ensuring that an LLM's responses don't contain offensive material.
H
Created for this library
A SaaS team adds guardrails around its assistant so the model refuses to discuss accounts it does not have permission to access.
A bank deploys guardrails on its customer chatbot so it never quotes specific rates without retrieval from the approved source.
An LLM platform team builds reusable guardrails for personally identifiable information so individual product teams do not each implement their own.
Definition source: Google for Developers Machine Learning Glossary | Creative Commons Attribution 4.0 License