Agent Search: Implementing Responsible AI and SafeSearch for Content Filtering

The inherent risks associated with Agent Search include the potential generation of unexpected outputs. These outputs might manifest as text that is offensive, insensitive to others' feelings, or factually inaccurate. Recognizing these complexities and potential pitfalls, Agent Search has been meticulously designed with Google's overarching AI Principles at its core. These principles guide the development and deployment of AI technologies to ensure they are beneficial, fair, and safe. Nevertheless, it is paramount for developers to thoroughly understand, rigorously test, and responsibly deploy their models to mitigate these risks effectively.

Agent Search provides a dedicated SafeSearch filtering capability. This feature is specifically engineered to filter out search results that contain explicit content, particularly when performing website searches. The explicit content that SafeSearch aims to block includes, but is not limited to, adult material, violent imagery, and gory depictions. It is important to note that the SafeSearch filtering functionality is not applied to private content. The ongoing evaluation and continuous updating of these filtering mechanisms are crucial for maintaining their effectiveness.

In certain scenarios and use cases, customers may require access to a broader spectrum of information, even if it includes potentially explicit content, to gain comprehensive data insights and results from Agent Search. To accommodate such needs, Agent Search offers the flexibility to enable or disable these safety filters programmatically. This is achieved by setting the `safe_search` parameter to either 'true' (to enable filtering) or 'false' (to disable filtering) through API calls. For general search, the API call involves specifying the `safe_search` parameter in the request body.

It is crucial to understand that the `safe_search` setting primarily governs explicit content filtering for website searches. However, Agent Search's commitment to Responsible AI extends beyond this. Agent Search outputs, such as generated summaries, are systematically evaluated against a comprehensive set of safety attributes. This evaluation occurs irrespective of whether the `safe_search` setting is enabled or disabled. The system is designed to proactively filter out potentially harmful content, including defamatory, sexually explicit, malicious, or violent information, ensuring a baseline level of safety across all generated responses.

Integrating Agent Search into unique application scenarios and contexts requires careful consideration of additional Responsible AI factors and model limitations. Developers are strongly advised to adopt best practices related to fairness, interpretability, privacy, and security. These considerations are not merely optional but are fundamental to the ethical and effective deployment of generative AI technologies. By proactively addressing these aspects, developers can ensure that their Agent Search applications are not only powerful but also trustworthy and beneficial to users.