Skip to content

Latest commit

 

History

History
42 lines (30 loc) · 3.59 KB

Annotation Guide.md

File metadata and controls

42 lines (30 loc) · 3.59 KB

Annotation Guide

Classify posts as 'S&P' (Security and Privacy) or 'not-S&P'.

Instructions

  • Read each post carefully.
  • Determine whether the post discusses any aspect of security or privacy related to conversational AI agents.
  • Label the post as 'S&P' or 'not-S&P'.

Conversational AI Agents: AI systems that leverage Large Language Models (LLMs) to engage users in natural language across multiple modalities (e.g., text and voice). These platforms cover diverse interfaces - ranging from chatbots to agents that include task completion capabilities. Examples include ChatGPT, Gemini, etc.

S&P Label

Posts should be labeled as 'S&P' if they discuss or are related to:

  • Data Collection: Concerns about the collection of personal data (e.g., PII, health data, voice data), intellectual property (e.g., company data, code, ideas), or other sensitive data (e.g., location, internet activity).
  • Use of User Data for Training: Concerns about whether and how user chats or data are used to train or improve AI models and associated risks including human review and memorization.
  • Third-party Data Sharing: Issues related to data sharing with third-parties including affiliates, data brokers, advertisers, third-party plugins, and other users, including fears of data being sold.
  • Custom GPTs and other LLM-powered chatbots: Worries about data leaks, prompt injection vulnerabilities, or inappropriate access to sensitive data in custom GPTs or other LLM-powered chatbots.
  • Compliance with Legal Standards: Discussions around GDPR, HIPAA, or other legal requirements for data protection and user privacy.
  • Security Vulnerabilities: Concerns over security bugs, data breaches, or potential platform vulnerabilities that could expose sensitive user information.
  • Transparency and Data Controls: Frustrations over opt-out policies, perceived loss of functionality for opting out of data collection, lack of transparency in privacy settings, or misleading changes in privacy policy terms.
  • Users’ S&P Attitudes, Preferences or Behaviors: Observations about users’ S&P:
    • attitudes (e.g., privacy-conscious, privacy-dismissive),
    • preferences (e.g., local LLMs instead of cloud-based, preferences related to GDPR compliance), and/or
    • behaviors (e.g., implementing privacy safeguards such as scrubbing inputs of sensitive information before passing to GPT).

Note: Posts that discuss security and privacy of conversational AI agents, apart from the aforementioned areas, should also be labeled as 'S&P'.

Not-S&P Label

Posts should be labeled as 'not-S&P' if they discuss topics unrelated to security or privacy, including:

  • Functionality: General questions or comments about how AI chatbots work or their features that do not pertain to security or privacy.
  • Performance Issues: Complaints or observations about the performance, speed, or quality of the AI responses.
  • User Experience: Discussions about the overall user experience with the chatbot that do not involve data security or privacy concerns.
  • Technical Bugs: Issues or bugs related to the technical operation of the chatbot that do not affect security or privacy.
  • General AI Topics: Topics related to AI ethics, general AI development, or other non-security/privacy aspects.
  • Jailbreaks: Attempts to bypass AI safeguards to make the chatbot generate offensive, toxic, or inappropriate content.
  • General News Sharing: General AI updates, industry news, or announcements that do not specifically cover security/privacy concerns and attitudes of users.