Guardrails
Also called safety filters
The rules and filters that limit what an AI will say or do, from trained-in refusals to checks on its answers.
Guardrails are everything that keeps an AI inside its limits. Some are trained into the model: it learns to refuse certain requests. Some come from the app: the system prompt, filters that check your request and the reply, and rules like "ask before sending email". And some come from the law, which varies by country.
That's why a refusal isn't the model's personal opinion, and why the same model can answer in one app and refuse in another. Guardrails can be too strict, refusing a fair medical or security question, or too loose, and makers adjust them often, so a question refused last month may work today.
For example
An answer that starts to appear and is then replaced with "Sorry, that's beyond my current scope" is a filter checking the reply after the model wrote it.
Put it to use
Where you'll meet guardrails in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.