Microsoft is developing a way to eliminate hallucinations, or false responses, in artificial intelligence (AI) models. It filed U.S. Patent Application No. 18/140,658, entitled “Interacting with a Language Model using External Knowledge and Feedback,” in April 2023. The application published on October 31, 2024, and became available for public inspection.
The patent specification describes a method for improving the quality of responses from an AI language model. Specifically, the model generates a response to a user query and scores the response using information obtained from outside knowledge banks. A poorly scored response indicates that the generated response does not match the information contained in the knowledge bank and should be revised. The system then revises and rescores the response before presenting it to the user.
The application, which is currently waiting examination at the United States Patent and Trademark Office, includes three independent claims, which generally relate to the steps for generating the usefulness of a response.
*This post was authored by Daniel Lass, law clerk at Robinson+Cole. Daniel is not admitted to practice law.