cela edit

Author: PatrickFarley

articles/ai-services/openai/concepts/abuse-monitoring.md

@@ -22,7 +22,7 @@ There are several components to abuse monitoring:
 - **Content Classification**: Classifier models detect harmful text and/or images in user prompts (inputs) and completions (outputs). The system looks for categories of harms as defined in the [Content Requirements](/legal/cognitive-services/openai/code-of-conduct?context=/azure/ai-services/openai/context/context), and assigns severity levels as described in more detail on the [Content Filtering](/azure/ai-services/openai/concepts/content-filter) page. The content classification signals contribute to pattern detection as described below.  
 - **Abuse Pattern Capture**: Azure OpenAI Service’s abuse monitoring system looks at customer usage patterns and employs algorithms and heuristics to detect and score indicators of potential abuse. Detected patterns consider, for example, the frequency and severity at which harmful content is detected (as indicated in content classifier signals) in a customer’s prompts and completions, as well as the intentionality of the behavior. The trends and urgency of the detected pattern will also affect scoring of potential abuse severity.
     For example, a higher volume of harmful content classified as higher severity, or recurring conduct indicating intentionality (such as recurring jailbreak attempts) are both more likely to receive a high score indicating potential abuse. 
-- **Review and Decision**: Prompts and completions that are flagged through content classification and/or identified as part of a potentially abusive pattern of use are subjected to another review process to help confirm the system’s analysis and inform actioning decisions. Such review is conducted through two methods: automated review and human review.
+- **Review and Decision**: Prompts and completions that are flagged through content classification and/or identified as part of a potentially abusive pattern of use are subjected to another review process to help confirm the system’s analysis and inform actioning decisions for abuse monitoring. Such review is conducted through two methods: automated review and human review.
     - By default, if prompts and completions are flagged through content classification as harmful and/or identified to be part of a potentially abusive pattern of use, they may be sampled for review by using automated means including AI models such as LLMs instead of a human reviewer. The model used for this purpose processes prompts and completions only to confirm the system’s analysis and inform actioning decisions; prompts and completions that undergo such review are not stored by the abuse monitoring system or used to train the AI model or other systems.
     - In some cases, when automated review does not meet applicable confidence thresholds in complex contexts or if automated review systems are not available, human eyes-on review may be introduced to make an extra judgment. Authorized Microsoft employees may assess content flagged through content classification and/or identified as part of a potentially abusive pattern of use, and either confirm or correct the classification or determination based on predefined guidelines and policies. Such prompts and completions can be accessed for human review only by authorized Microsoft employees via Secure Access Workstations (SAWs) with Just-In-Time (JIT) request approval granted by team managers. For Azure OpenAI Service resources deployed in the European Economic Area, the authorized Microsoft employees are located in the European Economic Area. This human review abuse monitoring process will not take place if the customer has been approved for modified abuse monitoring. 
 - **Notification and Action**: When a threshold of abusive behavior has been confirmed based on the preceding steps, the customer is informed of the determination by email. Except in cases of severe or recurring abuse, customers typically are given an opportunity to explain or remediate—and implement mechanisms to prevent recurrence of—the abusive behavior. Failure to address the behavior—or recurring or severe abuse—may result in suspension or termination of the customer’s access to Azure OpenAI resources and/or capabilities.