Risk and safety evaluations for protected material (text) are now available in public preview, accessible through Microsoft Foundry UI and SDK experiences. Because foundation models are typically trained using a massive corpus of data, users are understandably concerned that models may output responses containing protected material, putting end users at risk of unintended infringement. With the Azure AI evaluation SDK, users can now simulate conversations with their generative AI model or application to try and illicit responses containing protected text (e.g. song lyrics, articles, recipes, select web content) and measure how often their AI outputs protected text in response (the defect rate). To do so, the evaluation checks the outputs against an index of third-party text content maintained on GitHub. Users can drill into evaluation details to better-understand how their application typically responds to these user prompts and the associated risks. With this information, users may decide to activate protected material detection in Azure AI Content Safety, adjust their system message, or apply other mitigations before rerunning the evaluation and deploying to production.
Learn more.