Google is adding a new feature to the Gemini application programming interface (API) and AI Studio to help developers ground the responses generated by artificial intelligence. Announced on Thursday, the feature called Grounding with Google Search will allow developers to check AI-generated responses against similar information available on the Internet. This way developers will be able to further improve their AI apps and provide more accurate and up-to-date information to their users. Google highlighted that such grounding methods are important for signals that generate real-time information from the web.
Google releases ‘Grounding with Google Search’ feature
The Google AI for Developers support page detailed the new feature that will be available on both the Gemini API and Google AI Studio. Both of these tools are extensively used by developers who are building mobile and desktop apps with AI capabilities.
However, generating responses from AI models can often lead to hallucinations, which can negatively impact the reliability of apps. The problem can become even more significant when the app discusses current affairs topics in depth, where the latest information from the web is required. While developers can fine-tune AI models manually without a guiding dataset, errors may still exist.
To solve this, Google is introducing a new way to verify the output generated by AI. This process, known as grounding, connects AI models to verifiable sources of information. Such sources contain higher quality information and add more context to the information. Some examples of these sources include documents, images, local databases, and the Internet.
Grounding with Google Search uses the last source to find verifiable information. Developers can now use the top results from Google search to compare information provided by Gemini AI models. The Mountain View-based tech giant claims the exercise will improve the “accuracy, reliability and usefulness of AI outputs.”
This method helps AI models overcome the cut-off date of their knowledge by receiving information directly from a grounding source. So, in this case, Gemini models can get the latest information by using the output of the search algorithm.
Google also shared an example of the difference between a grounded output versus an ungrounded output. “Who won the Super Bowl this year?” A baseless answer to the question. Was “The Kansas City Chiefs win Super Bowl LVII this year (2023).”
However, after using grounding with the Google search feature, the refined response was, “The Kansas City Chiefs won Super Bowl LVIII this year, defeating the San Francisco 49ers in overtime with a score of 25 to 22.” Notably, this feature only supports text-based output and cannot process multimodal responses.
