Google is expanding an effort to allow its cloud customers to ground their enterprise AI chatbots in real-world facts, including a new partnership with Moody’s to use its financial data. Why it matters: Generative AI will make up information, but grounding these systems in known factual data can significantly lessen hallucinations. In April, Google announced an effort to ground Vertex AI results in web search as well as a plan to allow companies to ground AI systems in their own internal data. Now, Google is offering an additional option: using third-party data to help ground AI results. Initial partners in this effort include Moody’s, Thomson Reuters and ZoomInfo. “You can actually trust the model to do a task on your behalf because you have a basis for trusting it,” Google Cloud CEO Thomas Kurian told Axios. The moves come as major genAI providers seek to prove that their systems are safe and reliable enough for business use. oom in: Google is also offering more ways to make its systems more reliable and predictable. One is a new “confidence score,” in which the AI model offers a numeric indicator of how sure it is of the answer. Another new feature allows customers to tell the model to focus on finding answers in the documents or other information included in a prompt, rather than from its broad training data. “We’ve taught the model how to guarantee that when it responds, it takes what’s in the input prompt as the primary information it needs to pay attention to,” Kurian said. “It avoids being distracted by all other training data.”
About OODA Analyst
OODA is comprised of a unique team of international experts capable of providing advanced intelligence and analysis, strategy and planning support, risk and threat management, training, decision support, crisis response, and security services to global corporations and governments.