Google has announced a significant upgrade to its Gemini application programming interface (API) and AI Studio aimed at improving the accuracy of AI-generated responses. This new feature, called Grounding with Google Search, allows developers to validate AI outputs against relevant information available on the Internet, thereby fine-tuning their applications for more reliable and current data.
The announcement, made on Thursday, highlights the need for grounding methods, particularly for AI prompts that require real-time information. The Google AI for Developers support page confirmed that this feature will be integrated into both the Gemini API and Google AI Studio, platforms widely utilized by developers creating mobile and desktop AI applications.
AI models often generate outputs that may not be entirely accurate, a phenomenon known as “hallucination,” which can diminish the credibility of applications—especially when addressing topics of current events where up-to-date information is crucial. While developers can manually adjust AI models, errors can still occur without proper guiding datasets.
To address this issue, Google is introducing a method that connects AI models to verifiable information sources. This grounding process enhances the quality of information by incorporating high-quality context from documents, images, local databases, and the web. Specifically, the Grounding with Google Search feature leverages top search results to cross-verify the information provided by Gemini AI models. Google asserts that this practice will significantly boost the “accuracy, reliability, and usefulness of AI outputs.”
Additionally, this method enables AI models to extend their knowledge beyond their typical cut-off date by sourcing the latest data directly from grounding references. As a result, Gemini models can now retrieve real-time information utilizing Google’s search algorithm.
An illustrative example provided by Google demonstrates the effectiveness of this new feature. In an ungrounded response to the query, “Who won the Super Bowl this year?” the AI model inaccurately stated, “The Kansas City Chiefs won Super Bowl LVII this year (2023).” However, after applying the Grounding with Google Search feature, the response was corrected to, “The Kansas City Chiefs won Super Bowl LVIII this year, defeating the San Francisco 49ers in overtime with a score of 25 to 22.”
It’s important to note that the Grounding with Google Search feature currently supports only text-based outputs and does not accommodate multimodal responses. This development marks a significant step forward for developers looking to enhance the functionality and trustworthiness of their AI applications.
