AI-103 Practice Question 43: Content Filter Severity Threshold

Severity thresholds are configurable per category (low, medium, high), and lowering one blocks more content in that category; the default filters at medium, so moving to low also catches low-severity hate and violence content. Critically, "content filters can be configured at the resource level. Once a new configuration is created, it can be associated with one or more deployments." Creating a stricter configuration and attaching it only to the chatbot's deployment tightens that deployment's filtering while the research assistant keeps its existing configuration untouched.

- A is wrong because a blocklist screens for literal terms and phrases you supply; it doesn't change how aggressively the hate and violence classifiers themselves fire, so it wouldn't tighten filtering on content that doesn't happen to contain the blocked strings. - B is wrong because Prompt Shields targets adversarial instructions hidden in user input or documents (jailbreaks and indirect attacks), not the hate and violence harm categories. - C is wrong because a configuration change at the resource level doesn't apply itself to every deployment automatically; a configuration must still be explicitly associated with a deployment, and doing that to every deployment would also tighten the research assistant's filtering, which the team was told not to touch.

Reference: Configure content filters (classic)