 |
Center for Nonlinear Studies |
We ask experts and accurate forecasters to predict the effect of AI progress on the likelihood of catastrophic pathogen outbreaks, particularly via AI-enhanced access to bioweapons. A primary objective is to inform capability scaling policies (CSPs) for leading AI labs and policymakers, with a focus on the risk posed by AI-enabled chemical and biological weapons. Despite general concerns, few details exist on the specific capabilities that could raise biosecurity risks, the thresholds for intervention, or necessary countermeasures. To address these gaps, we design an idealized evaluation scenario comparing the risk of pathogen synthesis between two groups—one with and one without access to large language models (LLMs). We ask experts to forecast the risk of a non-natural catastrophic pathogen outbreak (causing more than 100,000 deaths or more than $1 trillion in damage) under two AI progress scenarios: (A) no major AI advancements by 2026, and (B) a 10x increase in the proportion of STEM graduates able to synthesize pathogens, driven by LLMs. Expert elicitation suggests a median increase of more than 50,000 expected deaths (or greater than $500 billion in damages) within three years in scenario B vs. scenario A. These findings highlight the need for detailed, scenario-based evaluations to inform policy design and to identify "red-line capabilities" in AI development that warrant precautionary regulation.
Bio: Ezra Karger is an economist in the microeconomics research group at the Federal Reserve Bank of Chicago and the Research Director of the Forecasting Research Institute, where he develops incentive-compatible methods for forecasting unresolvable questions, explores the limits of forecasting in low-probability domains, and conducts large-scale surveys of experts, often focused on long-run geopolitical outcomes. In his role as an economist, he also uses large datasets to construct high-frequency indices that track policy-relevant economic indicators.
Host: Sara Del Valle (A-1)