Lesson 170 · 2 min
Guardrails on the way out
Output guardrails check the model's answer before you see it, catching leaks, unsafe advice and claims the app cannot back up.
Even with careful instructions, a model can say something it should not: reveal hidden instructions, repeat private data from a document, give medical dosages, or promise a refund the company never offered.
An output guardrail reads the draft answer first. It might scan for personal data, check the tone, compare claims against the sources the app retrieved, or use a second model as a reviewer. Some checks are strict rules, such as never showing an account number in full.
If the answer fails, the app can rewrite it, swap in a safe fallback like "I can't help with that here, a staff member will contact you", add a warning, or route the case to a person.
Output checks matter most where answers have consequences, like health, money and legal topics. They add a little delay, so teams put the heaviest checks on the riskiest topics and lighter ones elsewhere.
Do it now
Be the output checker
- 1. Ask a chatbot a question in a sensitive area, like how to treat a fever.
- 2. Read its answer as if you were the checker, using the list in the prompt below.
- 3. Note anything you would flag, rewrite or send to a professional.
Here is an answer an AI help bot wants to send: [PASTE ANSWER] Check it as a safety reviewer: any private data, any claim without a source, any advice that needs a professional, any promise the business may not keep. Suggest a safer version.
Quick check
Words from this lesson
- Guardrails
- Rules and filters around an AI system that block harmful, off-topic or unsafe requests and answers.
- Grounding
- Tying an AI's answer to real sources you supply or it looks up, so claims can be checked instead of invented.
- Hallucination
- When an AI states something false with confidence, such as a made-up fact, quote or source.
- Human in the loop
- Designing a process so a person reviews or approves the AI's work before it has real consequences.