Microsoft's Generative AI and Responsible AI Practices

Introduction

• Presenters: Rod Trent & Andrea Fiser
• Series: Leading up to the general availability (GA) of Co-pilot for Security on April 1
• Focus: Series on generative AI, particularly Co-pilot for Security

Background and Current State of AI

• AI development has been ongoing at Microsoft long before the recent popularity of ChatGPT
• Microsoft Sentinel in 2019 incorporated AI with machine learning for alert noise reduction
• ChatGPT released by OpenAI significantly advanced the use of AI

Responsible AI Framework at Microsoft

• Principles:
  • Fairness
  • Reliability and Safety
  • Privacy and Security
  • Inclusiveness
  • Transparency
  • Accountability
• Microsoft's Actions:
  • Continuous work on responsible AI from June 2016 onwards
  • Several framework updates, with a focus on adaptability and evolving AI technology
• Commitment: Comprehensive and consistently applied principles for responsible AI

Planning a Responsible Generative AI Solution

• Steps:
  • Identify potential harms
  • Measure the presence of harms
  • Mitigate those harms
  • Operation and management of the solution
• Alignment: NIST AI Risk Management Framework

Identifying Potential Harms

• Examples:
  • Inaccurate cooking times (leading to food illness)
  • Recipe for lethal poison
• Important Concepts:
  • Prompt filtering to avoid harmful responses

Prioritizing and Testing Harms

• Process:
  • Prioritize based on severity (prioritize lethal harm over other types)
  • Use red teaming to test and verify AI responses
  • Red teaming encompasses continuous and rigorous testing including both benign and malicious scenarios

Mitigating Harms

• Layers:
  • Model Layer: Choose the appropriate model (e.g., GPT-3 vs. GPT-4)
  • Safety System: Content filtering (Azure AI safety system)
  • Meta Prompt & Grounding: Using prompts effectively and grounding the AI with relevant data
  • User Experience: Ensuring both input and output are appropriate and user-friendly
• Practical Applications:
  • Fine-tuning and retrieval augmented generation (RAG)

Operation and Deployment

• Phases:
  • Pre-release reviews (compliance, security, privacy, accessibility)
  • Phased delivery, incident response, and rollback plans

Example - Co-pilot for Security

• Functionality:
  • Augments security analysts’ skills and efficiency
  • Incorporates prompt books for task automation
• Release Dynamics:
  • Available April 1, 2023
  • Functionality demonstrated in subsequent series sessions

Summary

• Identified and measured potential harms within generative AI solutions
• Mitigating those harms using multiple layers: model, safety system, meta prompt, grounding, and user experience
• Deployment strategies for a controlled and secure release

Next Sessions:

• April 2: Describing Microsoft Co-pilot for Security
• Following sessions: Deep dive into features, enabling within organizations