The correct option is Safety settings.

Safety settings directly control content moderation for Vertex AI generative models. They let you set strict thresholds for harm categories such as toxicity, hate speech, harassment, sexual content, and violence. For a helper aimed at children, you would configure stringent block thresholds so the model filters or blocks any responses that could be unsafe, even if that leads to simpler or less creative output. This is the only parameter family designed specifically to enforce safety policies.

Top k tunes how many of the highest probability tokens the model considers when sampling and it affects diversity rather than safety. Lowering or raising it does not reliably prevent harmful or inappropriate content.

Output length controls the maximum number of tokens returned and it can truncate responses but it does not moderate content. Shorter replies can still contain unsafe material.

Temperature adjusts randomness and creativity and a lower value can make output more deterministic, yet it does not enforce content filters. Harmful content can still appear without dedicated safety controls.

Full AWS Practitioner Certification Question