英文_谷歌（Google）_2025负责任人工智能进展报告（英文版）_16页_5mb

Responsible AI Progress Report Summary (February 2025)

Core Content of the Report

This report outlines Google's comprehensive approach to responsible AI, emphasizing a lifecycle-based governance model that includes govern, map, measure, and manage. The initiative is guided by AI Principles, which prioritize bold innovation, responsible development, and collaborative progress. The approach is informed by industry frameworks such as the NIST AI Risk Management Framework and aligns with ISO/IEC 42001 compliance.

Main Components of the Responsible AI Approach

Govern: Full-stack AI Governance

• Google employs a full-stack governance model, covering responsible model development, deployment, and post-launch monitoring.
• AI Principles guide all AI-related decisions and are integrated into various frameworks and policies:
  • Secure AI Framework: Focuses on security and privacy.
  • Frontier Safety Framework: Proactively prepares for risks posed by future AI models.
  • Prohibited Use Policies: Govern how users interact with AI models and features.
• Pre- and post-launch processes ensure alignment with these principles, including:
  • Model and application requirements
  • Leadership reviews
  • Documentation of training techniques and model behavior

Map: Identifying and Understanding AI Risks

• A scientific approach is used to map AI risks through research and expert consultation.
• Risk Taxonomy is developed based on NIST guidelines and Google's own experiences, covering:
  • Safety, privacy, and security risks
  • Transparency and accountability risks
• AI Misuse Research is a key focus, including:
  • Analysis of potential misuse by government-backed threat actors
  • Identification of novel risks in AI capabilities
• External Domain Expertise is leveraged through:
  • Workshops and demos with global safety engineering centers and industry events
  • Trusted testing groups with secure access to models and applications

Measure: Assessing Risks and Mitigations

• Red Teaming is a central part of risk measurement, both internal and external.
• Multi-layered Red Teaming includes:
  • Security-focused red teaming to simulate real-world attacks
  • Content-focused red teaming to identify vulnerabilities
  • External partnerships (e.g., DEF CON, Escal8) and AI-assisted red teaming using AI agents
• Model and Application Evaluations are conducted to assess alignment with policies and frameworks, using:
  • Benchmarks for safety, privacy, and security
  • Automated testing tools and AI-assisted evaluations
  • Synthetic testing data to scale evaluation processes

Manage: Mitigating Risks

• Google implements content safety, security, and privacy mitigations, including:
  • Safety filters to block harmful content
  • System instructions to guide model behavior
  • Safety tuning to ensure high-quality, responsible outputs
• Security Mitigations include:
  • Use of the SAIF Risk Self Assessment framework
  • Data sanitization and access controls to prevent data poisoning and model exfiltration
  • AI agents to detect and block security threats
• Privacy Mitigations are based on a contextual integrity framework, ensuring appropriate information sharing.
• Phased Launches and Post-launch Monitoring are used to test and refine mitigations before wider deployment.
• User Feedback Mechanisms are integrated into applications to promote safety and quality reporting.

Key Outcomes and Achievements

• 300+ research papers published on responsible AI and safety topics.
• Partnered with numerous external organizations such as the Frontier Model Forum, Partnership on AI, and MLCommons to advance responsible AI practices.
• Invested $120 million globally in AI education and training.
• Achieved a "mature" rating for Google Cloud AI through NIST AI Risk Management Framework evaluation.
• Certified Gemini app, Google Cloud, and Google Workspace under ISO/IEC 42001.
• 19,000 security professionals have completed the SAIF Risk Self Assessment.

Model Cards: Promoting Transparency

• Model cards are used to document and provide transparency about AI models.
• They include:
  • Model Details (developer, date, type, training data)
  • Intended Use (primary and out-of-scope use cases)
  • Factors (demographic, technical, and environmental)
  • Metrics (performance, risk, ethical considerations)
  • Evaluation Data and Training Data details
  • Quantitative Analyses and Ethical Considerations
• The Gemma family of models serves as a case study, demonstrating the use of model cards to ensure safety and reliability.

Case Study: AlphaFold 3

• AlphaFold 3 was developed with extensive research into biosecurity risks and potential benefits.
• Collaborated with 50+ external experts in DNA synthesis, virology, and national security.
• Conducted ethics and safety assessments and external evaluations to ensure responsible deployment.
• Partnered with EMBL to provide free tutorials and promote equitable adoption.

Conclusion

Google's responsible AI strategy is a continuous, collaborative, and iterative process. It emphasizes transparency, accountability, and risk management across the AI lifecycle. Through model cards, red teaming, external partnerships, and rigorous evaluation, Google aims to ensure that AI technologies are developed and deployed responsibly, with a focus on safety, privacy, and security.