Guardrails
Checks around the AI
Priya puts checks around the Garden Guide before it reaches more players. A guardrail is a check that runs before or after the AI. An input check looks at what the player typed. An output check looks at the answer before the player sees it.
A free safety check
OpenAI's moderation models "detect harmful content in text and images", and "The moderation endpoint is free to use". Its categories include "harassment", "self-harm" and "sexual/minors". Any studio can run it on what goes in and what comes out.
Quick check
A player types something hurtful to the Garden Guide. Where can Priya's guardrail catch it?
- An input check, before the AI answers
- Only after the AI answers
- Nowhere: the AI decides
Show the answer
An input check, before the AI answers
An input check stops it before the AI reads it at all.
Shard Garden's guardrails
- TopicGame questions only. Others get "I can only help with Shard Garden."
- PrivacyAnswers never contain names, addresses or contact details.
- CareA message about self-harm shows help from real people, not an AI answer.
Pick the guardrail
A player asks the Garden Guide to do their maths homework. What happens?
- A friendly note: it only helps with Shard Garden
- It does the homework
- The player is banned
Show the answer
A friendly note: it only helps with Shard Garden
The topic guardrail keeps the guide on the game, kindly.
Fail safely
If a guardrail or the AI service is down, the guide says "The Garden Guide is resting. Try the help page." It never skips a check to answer faster.
Quick check
If the safety check is slow, it is fine to skip it and show the answer.
- True
- False
Show the answer
False
A slow answer is better than an unsafe one. When checks fail, the guide rests.
Count what is blocked
A dashboard counts blocked questions and answers each day. A sudden jump means a new trick, a bug or a change in the model, and the team looks the same day.
Quick check
On Priya's dashboard, blocked answers jump from 20 a day to 2,000 after a model update. What comes first?
- Roll back the update and read the blocked answers
- Turn off the guardrail
- Nothing: the guardrail did its job
Show the answer
Roll back the update and read the blocked answers
The guardrail caught them this time. Something changed, and the safe move is to undo it first.
Answers that feel quick
Players notice waiting. A long answer can start appearing at once.
Bolt says: Next: response time and streaming.
Fact-checked 12 October 2026
- OpenAI API docs: Moderation developers.openai.com