Unveiling Microsoft Copilot in Bing: A Deep Dive into Responsible AI Practices

Microsoft Copilot in Bing represents a significant leap in AI-enhanced web search, offering users summarized search results, a conversational chat experience, and the ability to generate creative content. This article provides an in-depth look at how Microsoft approaches responsible AI in Copilot, covering key aspects like privacy, safety, transparency, and control.

The Foundation: Microsoft's Commitment to Responsible AI

Microsoft is deeply committed to responsible AI. The development of Copilot in Bing is guided by Microsoft's AI Principles and Responsible AI Standard. This commitment involves collaboration with responsible AI experts within Microsoft, including the Responsible AI Office, engineering teams, Microsoft Research, and Aether.

Key areas of focus include:

• Fairness: Ensuring AI systems don't discriminate against individuals or groups.
• Reliability & Safety: Guaranteeing AI systems perform consistently and safely under various conditions.
• Privacy & Security: Protecting user data and preventing unauthorized access.
• Inclusiveness: Designing AI systems that are accessible and beneficial to everyone.
• Transparency: Providing clear explanations of how AI systems work and make decisions.
• Accountability: Establishing mechanisms for addressing issues and ensuring responsible use.

Understanding the Core Components

To understand how Microsoft ensures responsible AI in Copilot in Bing, it’s essential to understand the following key terms:

• Large Language Model (LLM): An AI model trained on massive text datasets to predict sequences of words. LLMs enable tasks like text generation, summarization, and translation.
• Classification: As an important measure to ensure safety, machine learning models are used to sort data into information categories, to help detect or flag user submissions with potential harmful content.
• Grounding: Copilot bases its responses on web content pulled from Bing search results that are weighted for relevance, quality, credibility, and freshness. It provides website links for verification.
• Metaprompt: A program that guides the behavior and ensures it aligns with Microsoft AI Principles and user expectations.
• Mitigation: Methods to lower potential risks associated with using AI-driven features.
• Prompt: User input (text, voice, image or more) used to communicate information that would create a AI powered generation experience.
• Red Teaming: Testing used by experts to evaluate systems and test planned mitigation.

How Copilot in Bing Works

Copilot in Bing leverages advanced technologies to deliver an innovative search experience. Here's a breakdown of the process:

1. Prompt Input: Users enter a prompt into Copilot.
2. Data Processing: The system combines the prompt with recent conversation history, a metaprompt, and top-ranked search results.
3. LLM Generation: The LLM generates a response based on the combined data.
4. Response Presentation: The response is presented in various formats (traditional links, AI-generated summaries, images, etc.) with citations to source websites.

Within prompts, grounded responses are favored, where inputted information supports the LLM statement so the response remains relevant to what is drawn from the web. In constrast, a non-grounded response is when something is presented by LLM that cannot be founded in an input source.

Identifying, Measuring, and Mitigating Risks

Microsoft employs a rigorous approach to identify, measure, and mitigate potential risks associated with Copilot.

Discovery

Initial work began in the summer of 2022 with exploratory analysis of GPT-4, in partnership with Open AI for extensive red team testing. This rigorous exploration aims to evaluate the functionality of the newest technologies in development to identify potential areas of abuse, by intentionally generating harmful responses.

Measurement

The team creates templates to capture conversational structures. It reviews whether dialogue contains harmful content, by adapting and implementing guidelines. The responsible AI index measures Copilot's effectiveness.

Mitigation Strategies

• Staged Release: A phased rollout allows Microsoft to monitor user behavior, identify potential abuse, and implement necessary mitigations before a wider release.
• AI-Based Classifiers: Classifiers categorize text to flag potentially harmful content (self-harm, violence, hate speech, etc.).
• Limiting Conversation Turns: Limiting the number of turns in a chat session prevents repetitive or unhelpful responses.
• Prompt Augmentation: When user prompts are unclear, the system uses the LLM to add details that will provide more certainty that users will get the response they seek.
• Content Credentials: Content Credentials uses cryptographic methods that mark AI images created with Designer’s feature to tag all AI generated images on its origin or source of creation, by displaying the origin, time, and date.
• Terms of Use: As a use agreement, there are identified limitations, as well as consequences to violations.

Safeguarding Privacy

Protecting user privacy is a core principle in the development of Copilot in Bing. The platform is designed to collect and use personal data only when necessary and not retain it longer than required. Microsoft protects user privacy through transparency, control and minimization of user data.

Visual Search Privacy

Copilot employs facial blurring techniques when users upload images in chat prompts. This protects the subject's privacy. Also, digital representations of people in uploaded images are not stored or shared with third parties.

Protecting Children and Youth

Microsoft takes extra precautions to protect children and young people using Copilot in Bing. All Microsoft accounts for users under 13 (or as defined by local laws) are blocked from accessing the experience. SafeSearch is set to "Strict" for these users to prevent exposure to inappropriate content.

Transparency and Giving Users Control

Microsoft believes that trust in technology depends on transparent access to data usage and user control. Copilot in Bing offers insights into how and why the product works, while keeping the user in control.

• AI Disclosures: Copilot provides multiple touchpoints for AI disclosures, including notifications that users are interacting with an AI system, to avoid over reliance on AI.

Continuous Feedback, Monitoring, and Action

Microsoft continuously monitors Copilot in Bing and has processes for addressing potential abuse or violations of the Terms of Use. This includes tools for users to submit feedback and report issues to Microsoft's operations teams.

Conclusion

Microsoft's Copilot in Bing represents a significant advancement in AI-powered web search. By prioritizing responsible AI practices, Microsoft aims to deliver a powerful and beneficial tool while mitigating potential risks and ensuring a safe and trustworthy experience for all users. As the technology evolves, Microsoft remains committed to continuously improving its approach to responsible AI and adapting to new challenges.