通用人工智能预测与情景分析_领域现状_方法论缺口及战略意义_75页_504kb

Artificial General Intelligence Forecasting and Scenario Analysis Summary

Core Content

This report provides an analysis of the current state of methodologies for forecasting the arrival of artificial general intelligence (AGI), highlighting methodological gaps and strategic implications. It emphasizes the need for robust forecasting infrastructure and outlines a framework for decisionmakers to navigate uncertainty in AI development.

Main Viewpoints

• AGI Timeline Shifts: Over the past five years, expert forecasts for AGI have shifted significantly from mid-century to the near term. Prediction markets and compute-centric models now estimate AGI to arrive in the 2030s, while the most recent expert survey places the median at 2047 for high-level machine intelligence (HLMI). The "full automation of labor" estimate has also moved earlier, from 2164 to 2116.

• Definition of AGI: AGI is defined as systems capable of performing most economically valuable work at or above human level across a wide range of domains. This definition focuses on task performance rather than autonomous goal-pursuit or economic transformation, though alternative definitions are acknowledged and discussed.

• Forecasting Limitations: The forecasting field is immature, lacking resolved forecasts for calibration, benchmarks resistant to saturation, real-time model capability insights, and independent validation. These limitations affect the reliability of current methods.

• Value of Forecasts: AGI forecasts offer two types of value: predictive signal for near-term action and synthesis of information for long-term preparation. They should be treated as scenario-structuring tools rather than point estimates.

Key Information

• Forecasting Approaches: The report synthesizes expert surveys, prediction markets, compute-centric models, and scenario analysis. These methods vary in their assumptions and outputs, leading to divergent forecasts.

• Sources of Disagreement: Disagreement among forecasters often stems from different definitions of AGI and varying targets (technical capability, operational deployment, societal transformation). Even with consistent definitions, substantial disagreement persists.

• Strategic Implications: Decisionmakers must prioritize flexibility, adaptiveness, and robustness in strategy development. They should prepare for a range of possible AI futures, including short-timeline trajectories, and base their actions on concrete, observable indicators.

Recommendations

1. Treat Forecasts as Scenario-Structuring Tools
   Planners should focus on plausible, consequential, and challenging scenarios rather than precise probabilities. This includes scenarios that lack adequate preparation, such as those resembling a "Taiwan contingency" or "9/11-style attack."

2. Build Adaptive Capacity with Clear Reassessment Triggers
   Decisionmakers should develop contingency plans for near-term AGI scenarios and establish explicit triggers for reassessment. These triggers should be tied to observable indicators, such as AI systems completing multiweek software engineering projects or AI research automation exceeding 50%.

3. Link Forecasts to Strategic Choices
   Forecasts should be structured around conditional questions that explore how different investment, strategic, or diplomatic scenarios affect timelines and capabilities. Integrating forecasting with scenario planning methods can help surface these contingencies.

4. Invest in Methodological Diversity and Novel Approaches
   Forecasting should incorporate a broader range of methods and disciplinary perspectives, including econometricians, cognitive scientists, and complex systems researchers. This helps surface blind spots and challenge shared assumptions.

5. Institutionalize Independent Validation and Stress-Testing
   Influential forecasting models should undergo adversarial scrutiny, similar to climate modeling or macroeconomic projections. This includes red-teaming assumptions, conducting sensitivity analyses, and systematically comparing predictions with outcomes.

6. Invest in Independent, Ongoing Capability Evaluation
   Current benchmarks degrade quickly due to contamination, saturation, and optimization pressure. Decisionmakers should support evaluation infrastructure that continuously develops new assessments and tracks real-world task performance.

7. Develop Monitoring Infrastructure for Compressed Timelines
   Short-timeline scenarios require higher-frequency, more heterogeneous monitoring approaches. This includes tracking leading indicators in real time and detecting discontinuities rather than just extrapolating trends.

8. Strengthen Internal Monitoring of AI Automation in R&D
   Frontier laboratories should develop standardized monitoring systems to track AI's role in accelerating research. This presents a collective action problem, requiring coordination through industry commitments or government-facilitated information-sharing.

9. Improve Forecasting Ecosystem Maturity
   The forecasting ecosystem is immature and lacks systematic validation. Improving it is critical for providing reliable inputs to strategic and policy decisions.

Strategic and Policy Implications

• Uncertainty in Decisionmaking: Given the compressed and uncertain timelines, strategies must be flexible and adaptive. Static investments may not be sufficient if development accelerates beyond expectations.

• Cost Asymmetry: Policies should account for the asymmetry in costs between different AI futures. Decisionmakers must weigh the risks and benefits of various scenarios when shaping policy.

• Institutional Independence: Forecasting institutions should be independent of the organizations they are forecasting about to ensure objectivity and avoid bias.

• Need for Systematic Integration: Forecasting should be systematically integrated with scenario planning to improve decision-relevant preparation and response options.

Conclusion

This report underscores the importance of understanding and improving AGI forecasting methods to support effective strategic and policy decisions. It advocates for a shift in focus from predicting exact timelines to preparing for a range of possible AI futures, emphasizing the need for robust, flexible, and adaptive strategies.

Funding and Acknowledgments

The research was independently initiated and conducted within the Center on AI, Security, and Technology at RAND. Funding came from operational income and gifts/grants from philanthropic supporters. The report acknowledges contributions from numerous experts and teams involved in its development and review.

Summary Reference

• Grace et al. (2025): "Thousands of AI Authors on the Future of AI," Journal of Artificial Intelligence Research, Vol. 84, October 2025.