Research

Foresight and Forecasting

Forecasting involves using quantitative techniques to predict future outcomes. At TFI, our efforts focus on how to best assess the likelihood of specific risks and opportunities associated with advanced AI technologies. We leverage computational models, statistical analysis, and probabilistic reasoning to build tools to determine potential futures and evaluate how various variables might influence these possible trajectories. Our forecasting tools can be utilized in a number of different ways:

• Risk assessment: by providing probabilistic estimates of specific events occurring within defined timelines
• Impact and risk evaluation: by estimating the societal, economic, and environmental impacts of technological developments and/or associated policies.
• Causal inference and complex systems analysis: by integrating complex systems modeling, our forecasting approach can provide a clear picture of interdependencies and causal relationships within systems.

Foresight is not just about predicting specific events but rather understanding a broad spectrum of possible scenarios, including those that are less likely but potentially very high impact. It is a structured and systematic way to gather evidence about future events to anticipate challenges and prepare for the unexpected. Foresight involves envisioning multiple potential futures by comprehensively evaluating identifying influential factors, trends, and early sign indicators. Our foresight efforts are geared toward recognizing emerging trends, technological innovations, and societal shifts that could influence short, medium, and long-term outcomes. It specifically includes:

• Expert elicitation: engaging with subject matter experts to gain deep insights and diverse perspectives.
• Visioning workshops: facilitating discussions that help us understand different risk perceptions, desirable futures, and relevant questions to forecast for effective decision-making.
• Monitoring developments across technology, social behaviors, economics, etc., to anticipate changes.

Scenario Discovery and Scenario Mapping

Scenario discovery is a novel process that uses simulation-generated databases to model the spectrum of plausible futures and unique combinations while refining and isolating individual scenario elements that meet user-defined constraints, such as where a policy fails to meet its objective or is below a particular threshold. The process identifies scenarios based on current knowledge and predictive insights to uncover “unknown unknowns” and critical factors and events that could influence future developments. Through the unique combination of computational modeling, consultation with experts, and data analysis, we can find ways to systematically identify and categorize key variables that could unleash different futures. Specifically, it can help:

• Identify broad patterns and trends that are likely to influence futures
• Uncover specific parameter combinations that are highly predictive of policy-relevant scenarios
• Identify areas of critical uncertainties
• Identify overlooked or hidden variables
• Develop a comprehensive understanding of future challenges and opportunities

Scenario mapping involves creating detailed outlines of each identified scenario, elaborating on how various elements could interact over time. Instead of focusing on outcomes, it provides visualization of future trajectories. Scenario mapping can help:

• Provide a visual representation of complex pathways and points of intersection – showing how choices can lead to different futures, and provide a basis for strategic conversations
• Assess the impact of potential decisions – allowing for strategic planning across a range of possible futures

Structural Risks

Structural risks are the externalities that arise from developing and deploying advanced AI technology within the broader sociotechnical system and how the technology shapes or is shaped by the broader environment. These are the complex societal-scale risks that can have unpredictable consequences across individual, systemic, and global scales.

TFI works to develop strategies to map and measure the dangers of AI structural risks with the intent of modeling key forces and indicators for predictive analysis. Our approach maps structural risks in unique ways, specifically by leveraging simulations such as agent-based modeling, by employing scenario discovery for more powerful and nuanced explorations of this space, and by eliciting feedback from external partners. It is imperative to fully understand the scope of structural risks in order to have any hope of mitigating them. For example, our work on the Reasoning Under Uncertainty Trap unpacks how an initially linear harm from an AI misuse, when considered in a broader structural context of human-human, human-system, and human-AI feedbacks, in fact can lead to rapid, non-linear, deleterious outcomes.

AI Governance and Evaluations

AI Governance refers to the systems and processes established to oversee the development, deployment and impact of AI technologies. TFI’s research in this area includes identifying the most effective policies, standards, and regulatory framework to ensure that AI operates within safe boundaries. It also involves actively engaging with various stakeholders, including policymakers and subject-matter experts to optimize outcomes. At TFI, we believe that effective governance is essential to mitigating societal-scale risks associated with AI.

Evaluations focus on systematically assessing AI systems at all stages of development and deployment to ensure their safety, reliability, and alignment with human values. This process involves pre-deployment risk-assessments, testing, and reviews to verify the safety and integrity of systems, as well as post-deployment monitoring, third party audits, and red-teaming exercises to improve the robustness and reliability of AI systems against potential threats. The evaluations process may also propose solutions to rectify identified problems or vulnerabilities and include crisis management plans. At TFI we strive to highlight the importance of this critical and ever-evolving component of risk assessment, and to contribute valuable insights for decision-makers.