Configurable HAP/PII Explainability & Exclusion Controls for OpenScale

The Watson OpenScale HAP and PII Metric Monitor currently lacks transparency and configurability, making it difficult for customers to effectively govern and fine tune content risk in real world use cases. When the HAP and PII metric triggers, there is no visibility into which specific word or phrase caused the violation in a large field of text, leaving governance and operations teams unable to troubleshoot false positives efficiently.

Additionally, OpenScale does not provide a mechanism to exclude or allow specific customer defined terms from HAP and PII evaluation. This limitation is particularly problematic in regulated and domain specific industries such as healthcare, emergency response, and legal services, where industry standard language can mistakenly trigger HAP and PII alerts. For example, a medical chatbot recommending that a patient call 911 immediately can incorrectly trigger a HAP and PII alert due to the string 9 11 (9/11). We have seen the same effect with specific medical terms as well.

This lack of explainability and configurability results in noisy alerts, broken thresholds, and automated notifications that obscure true positives and slow response times. At a minimum, OpenScale should support the ability to view the specific terms or phrases that triggered HAP and PII violations, provide customer configurable exclusion controls for HAP and PII keywords, and offer more granular tuning to support industry specific jargon and workflows. Until these gaps are addressed, the HAP and PII monitor remains of limited value for certain customer use cases, which is disappointing for organizations that have made significant investments in the OpenScale platform.

These enhancements would enable governance teams to accurately assess risk, reduce false positives, and maintain trust in OpenScale automated monitoring and alerting capabilities, especially when reviewing long and complex outputs such as medical documentation where multiple terms could plausibly trigger a violation.