Jargon:Guardrails
1 min read
Say this at the party
“Guardrails are the rules and filters wrapped around an AI to stop it saying or doing things it shouldn't, like a crash barrier on a road.”
In plain English
Some guardrails are trained into the AI itself, so it learns to refuse harmful requests.
Others sit around it like a bouncer, checking what goes in and what comes out: blocking rude language, hiding personal details or keeping a customer service bot on topic.
They make AI safer and more predictable, but a determined person can sometimes find a way round.
Where it came from
A borrowed everyday word (barriers along a road). In AI it became common around 2023, helped by tools like NVIDIA's open-source NeMo Guardrails, and is now a standard feature name in products such as Amazon Bedrock Guardrails.
See it in action
A bank's chatbot is asked for a joke about a celebrity and politely steers back to banking. Later, someone types their full card number and the bot blanks it out before saving the chat. Both are guardrails at work.
What people get wrong
- Guardrails aren't a guarantee. Jailbreaks and prompt injection exist precisely because they can be got round.
- More guardrails isn't always better. Too many and the AI refuses perfectly reasonable requests.
Go deeper
- PaperRebedea et al., NVIDIA, 2023NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable RailsExplains the two kinds of guardrails: trained in, and added around the model.
- GuideAmazon Web ServicesDetect and filter harmful content by using Amazon Bedrock GuardrailsWhat guardrails look like in a real product: filters for topics, words and personal data.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.