web
You’re offline. This is a read only version of the page.
close
Skip to main content
Community site session details

Community site session details

Session Id :
Dynamics 365 Community / Blogs / Community Help & Announcements / Spam detection agent

Spam detection agent

Views (1)
Bret Fraser Profile Picture Bret Fraser Community Manager
This FAQ provides answers to frequently asked questions about the AI technology used in Spam Detection. It includes key considerations and details about how the AI is used, how it was tested and evaluated, and any specific limitations. 
 
What is Spam Detection in the Community? 
Spam Detection is an AI-powered feature that automatically evaluates user-generated posts in real time to determine whether the content qualifies as spam. If a post is classified as spam, it can be withheld, flagged for moderation, or deprioritized before being shown to the community. This feature is designed to reduce low-quality, promotional, or malicious content that negatively impacts the community experience. The spam detection functionality is triggered at the time of post submission and is currently available for English-language content only. 
 
What are the capabilities of Spam Detection in the Community? 
The Spam Detection feature offers the following capabilities: 
  • Real-time spam classification: Automatically evaluates new posts at the time of submission and identifies whether they are spam.
  • Pattern and intent recognition: Detects a wide variety of spam tactics, including promotional language, link spam, low-effort posts, and scam attempts. 
  • Context-aware filtering: Uses AI language understanding to go beyond simple keyword-based rules and analyze post structure and tone. 
  •  
How is the Spam Detection system evaluated? What metrics are used to measure performance? 
The Spam Detection system underwent testing and review before release, including: 
  • Internal benchmarking using a labeled set of real-world examples to validate accuracy and coverage. 
  • Scenario-based test cases to assess performance on common spam patterns such as shortened links, clickbait, and irrelevant promotions. 
  • Shadow testing on live posts prior to launch to compare AI judgments against human moderator decisions. 
  • Manual audit feedback loops, where moderators can report false positives and missed spam for ongoing improvement. 
Performance is measured based on: 
  • Precision and recall on internal test cases 
  • User and moderator feedback on classification accuracy 
  •  
What are the limitations of Spam Detection in the Community? How can users and moderators minimize the impact of these limitations? 
While the system improves spam filtering at scale, it has the following known limitations: 
  • Language support: The current model is optimized for English. Posts in other languages may not be accurately classified. 
  • No contextual memory: The system evaluates each post in isolation and does not consider user history or thread context. 
  • No support for media or attachments: Spam embedded in images or non-text content cannot be detected at this time. 
  • Possible false positives or misses: Some cleverly disguised spam may not be flagged, and some borderline content may be incorrectly flagged as spam. 
  • AI model variability: As with all generative AI systems, results may vary slightly, especially for edge cases. 
To minimize the impact of these limitations: 
  • Moderators can review flagged posts via the audit log. 
  • Feedback can be used to refine prompts or examples. 
  • Users can report incorrectly flagged posts for follow-up. 
  •  
What data is collected or used by the Spam Detection Agent? 
The Spam Detection Agent is designed with privacy in mind. The system does not collect, send, or store any personally identifiable information (PII) as part of its operation. When a user submits a post, only the text content of the post is evaluated by the AI model for spam detection purposes. 
  • No usernames, email addresses, IP addresses, or other personal metadata are sent to the model. 
  • The AI model processes post content in real time and does not retain inputs. 
  • The system is configured to strip or exclude any PII before sending data to the hosted large language model (LLM). 
  • Logs used for auditing or evaluation are handled according to internal security and compliance standards and are redacted of PII. 
This approach ensures that the spam detection system upholds strong privacy and responsible AI practices while improving the quality and safety of community content.

 

Comments

*This post is locked for comments