兰德-人工智能对安全的风险与工作的未来（英文）-2017.12-23页

Summary of The Risks of Artificial Intelligence to Security and the Future of Work

Core Content

This document explores the risks and implications of artificial intelligence (AI) in two major domains: security and the future of work. It emphasizes the need for interdisciplinary policy analysis to understand how AI can reshape societal structures and introduce new vulnerabilities.

Main Points

1. Introduction and Context

• The rapid advancement of AI has led to significant changes in how society interacts with technology.
• AI is increasingly integrated into decision-making processes, from simple devices like Roomba robots to advanced systems like IBM's Watson.
• The report highlights the importance of identifying blind spots and biases in AI systems, especially in critical areas like criminal justice, cybersecurity, and surveillance.
• A structured interdisciplinary approach was used to analyze AI's potential impact across various sectors.

2. Research Methodology

• A diverse team of RAND researchers from multiple disciplines (e.g., economics, psychology, political science, engineering) was assembled.
• The team engaged in brainstorming, small group discussions, and whole-group debates to define AI and identify its application areas.
• A working definition of AI was established: "autonomous, non-biological learning system."
• The team identified key areas for future AI impact: security, employment, decision-making, and health.
• A future-casting exercise was conducted, which highlighted consistent AI applications across subgroups, including security and employment.

3. Security Risks

National Security

• AI in national security poses risks such as automated decision-making errors and cybersecurity vulnerabilities.
• AI can be used in information warfare, where adversaries may manipulate AI systems with disinformation.
• Malware can be enhanced with AI to become more strategic and harder to detect, as seen in Stuxnet.
• Data diet vulnerability refers to AI systems being limited by the quality and bias of their training data.
• AI could be exploited to create automated double agents that manipulate political discourse and public opinion.

Domestic Security

• AI is already being used in government surveillance, raising concerns about privacy erosion and legal rights.
• Predictive policing and algorithmic bias in criminal justice systems, such as the COMPAS algorithm, have been shown to produce systemic inequities.
• The use of traffic cameras and robots in law enforcement has sparked debates about due process and the presumption of innocence.
• The legal status of AI systems is an emerging issue, with scholars exploring legal personhood for artificial agents.

4. Future of Work

• The future of work is a central concern, focusing on how AI affects the supply and demand for human labor.
• AI has the potential to displace workers in routine-based tasks, leading to job polarization—a shift in job demand away from middle-skill jobs.
• Frey and Osborne (2013) estimated that 47% of U.S. jobs could be displaced by automation, but this has been challenged by the OECD (Arntz et al., 2016), which suggests only 9% of jobs are at risk for full automation.
• The report also discusses the deskilling effect, where automation reduces the need for certain skills and shifts training focus.
• Microwork has emerged as a new form of labor enabled by AI platforms, such as TaskRabbit, Uber, and Amazon Mechanical Turk.
• Concerns about income inequality and the concentration of AI benefits among a few "superstar firms" are highlighted.

5. Near- to Medium-Term Trends

• AI is already disrupting traditional work patterns, creating new opportunities and challenges.
• The deskilling effect is a real concern, but it mirrors historical patterns of technological change.
• The task-based approach to analyzing automation risks is more accurate than previous occupation-based models.
• Moravec's paradox shows that humans often misjudge the difficulty of tasks for AI, overestimating tasks requiring creativity or social intelligence and underestimating perception and routine tasks.

6. Framework for Assessing Occupational Susceptibility to Automation

• The report proposes a framework based on RAND's research on managing occupational surprise.
• Two key factors determine an occupation's susceptibility to automation:
  1. Amount of chaos in the tasks: occupations involving unpredictable or complex scenarios are less likely to be automated.
  2. Task definition and routine:
     • AI excels at well-defined, repetitive tasks.
     • Planning and chaotic environments remain challenging for AI.

Key Information

• AI in security raises concerns about disinformation, surveillance, and legal rights.
• AI in employment could lead to job displacement, deskilling, and increased inequality.
• Blind spots in AI include data bias, algorithmic fairness, and ethical implications.
• Policy challenges are growing as AI becomes more embedded in both public and private sectors.
• Interdisciplinary collaboration is essential for understanding and addressing AI's impact on society.

Conclusion

The report underscores the urgent need for policy frameworks to address the risks and opportunities AI presents. It calls for a nuanced understanding of AI's role in security and labor markets, emphasizing the importance of ethical, legal, and social considerations in AI development and deployment.