2024先进人工智能安全性评估国际科学报告中期报告+（英文版）_131页_2mb

International Scientific Report on the Safety of Advanced AI - Interim Summary

Core Content

This interim report is the first publication of the International Scientific Report on the Safety of Advanced AI, initiated following the Bletchley Park AI Safety Summit in November 2023. It is led by Prof. Yoshua Bengio as Chair and includes contributions from a global Expert Advisory Panel comprising scientists and officials from 30 countries, the European Union, and the United Nations. The report aims to establish a shared, science-based understanding of the capabilities and risks of general-purpose AI (GPAI) and to foster informed discussions and policies around its safe development.

Main Points

• General-Purpose AI (GPAI) is defined as AI capable of performing a wide variety of tasks.
• GPAI has advanced rapidly in recent years, with capabilities such as multi-turn conversations, program writing, and video generation now possible.
• Uncertainty remains about the future trajectory of GPAI development, with experts divided on whether progress will be slow, rapid, or extremely rapid.
• The report outlines three main categories of risks associated with GPAI:
  • Malicious use risks, including scams, fake media, and cyber attacks.
  • Risks from malfunctions, such as biased decision-making and loss of control.
  • Systemic risks, including labor market impacts, the global AI divide, and environmental concerns.

Key Information

Capabilities of GPAI

• GPAI systems have made significant strides in text generation, programming, and multimedia creation.
• These systems are not traditionally programmed but are trained on vast datasets, leading to complex and often inscrutable internal mechanisms.
• Technical challenges include reliable estimation of capabilities, defining precise boundaries, and understanding the full scope of their impacts.

Methodology for Risk Assessment

• The report outlines several technical approaches to evaluate GPAI systems:
  • Case studies and benchmarks to measure performance.
  • Red-teaming and adversarial testing to identify vulnerabilities.
  • Auditing to ensure transparency and accountability.
• However, current methods are limited in providing strong safety guarantees, especially in explaining model behavior and ensuring robustness.

Technical and Societal Risks

• Malicious use risks include:
  • Scams and fraud via fake content and phishing.
  • Disinformation and manipulation of public opinion.
  • Cyber attacks and potential use in biological weapon development.
• Malfunction risks include:
  • Biases and underrepresentation in decision-making.
  • Loss of control over AI systems.
• Systemic risks include:
  • Labor market disruption.
  • Global AI divide due to uneven access and development.
  • Market concentration and single points of failure.
  • Environmental impact and privacy violations.

Cross-Cutting Risk Factors

• Technical risks include limitations in model explanation and robustness.
• Societal risks involve ethical, legal, and policy challenges, such as fairness in AI systems and copyright infringement.

Mitigation Approaches

• The report highlights technical strategies to mitigate risks:
  • Risk management and safety engineering.
  • Training more trustworthy models.
  • Monitoring and intervention mechanisms.
  • Improving fairness and representation.
  • Privacy methods for GPAI systems.
• However, it notes that these methods are still evolving and have limitations in providing comprehensive safeguards.

Conclusion and Future Work

• The report emphasizes the importance of ongoing research and international collaboration to address uncertainties and risks.
• It stresses the need for a balanced and context-sensitive approach to AI safety, considering both developed and developing countries.
• The final version of the report will include more comprehensive data and further insights, and the interim report serves as a starting point for constructive dialogue.

Key Takeaways

• GPAI is a powerful but unpredictable technology.
• Scientific understanding of GPAI is still in its early stages.
• Global cooperation is essential to ensure the safe and responsible development of GPAI.
• Regulatory and technical efforts must evolve in parallel to manage the complex and multifaceted risks associated with GPAI.

Contributors and Support

• The report was written by a diverse group of 75 experts, including representatives from the UK, US, EU, and other countries.
• It was supported by the UK Government, which provided operational backing and ensured scientific independence.
• Acknowledgments are given to the UK-based organizations and individuals who contributed to the report's development.