Microsoft patents AI filter that blocks medical chatbots from breaking rules

1. A method comprising: monitoring, by a model safeguarding processor, an interactive session between a user and a generative Artificial Intelligence (AI) system to ensure that content exchanged between the user and the generative AI system complies with a Clinical Practice Guideline (CPG) based (CPG-based) protocol, the model safeguarding processor being independent from the generative AI system, wherein monitoring the interactive session includes: receiving, by the model safeguarding processor, model-generated content output by the generative AI system based on information provided by the user, wherein the model-generated content includes an initial inquiry for the user; processing, by the model safeguarding processor without presenting the initial inquiry to the user, the model-generated content based on the CPG-based protocol to determine that the initial inquiry is not relevant to, or otherwise lacks compliance with, the CPG-based protocol; sending, by the model safeguarding processor, a request to the generative AI system upon determining that the initial inquiry is not relevant to, or otherwise lacks compliance with, the CPG-based protocol, the request prompting the generative AI system to provide a replacement inquiry for the user; and presenting, by the model safeguarding processor, the replacement inquiry to the user in lieu of the initial inquiry responsive to confirming that the replacement inquiry is relevant to, or otherwise complies with, the CPG-based protocol, wherein the replacement inquiry prompts the user to provide user-generated content in response to the replacement inquiry.