AI Threat to Strategic Stability: MAD, Loss Framing, and US Public Preferences

• In Event: Artificial Intelligence & International Security

Abstract

A number of scholars have raised concerns that artificial intelligence (AI) may threaten strategic stability in the future (e.g., Kasperse and King 2019; Horowitz 2019). These scholars proceed from the definition of stability rooted in MAD (Mutually Assured Destruction), and subsequently worry that integrating AI into NC4ISR will upset that strategic stability by either incentivizing a first strike during a deterrence crisis (Avin and Amadae 2019) or that technical errors in the AI will make accidental nuclear first use more likely (Sauer 2019). At some point in the not-to-distant future, the U.S. will face a political debate about how to weight the potential costs and benefits of integrating AI into NC4ISR (Horowitz 2016; Zhang and Defoe 2019). Yet, existing scholarship has yet to consider, let alone empirically examine, how the public processes the risks AI may pose to strategic stability.
We, therefore, examine the validity of scholars’ concern about the threat AI poses to stability by conducting a series of survey experiments that provide subjects drawn from the US public with prospective U.S.-China scenarios. Our results show, first, that when confronted with AI’s strengths and weaknesses, the public strongly opposes integrating AI into US systems that can locate Chinese nuclear weapons. Second, we find that the public becomes more willing to support AI-controlled NC4ISR when presented with either of the two potential loss frames considered in existing scholarship—i.e., China integrates AI into its NC4ISR and China pulls ahead of the US in the overall balance of power. Yet, neither frame leads to anything close to a majority of Americans to support such a move. Third, we find that MAD logic neither motivates the US public’s opposition to integrating AI nor does it affect their susceptibility to loss frame thinking. While most Americans believe MAD will enhance US national security, they do not embrace the utility of mutual vulnerability and technological restraint. Ultimately, then, our results moderate the likelihood that the US will integrate AI into NC4ISR, but not for the reasons advanced by IR scholarship to date.

Authors

• Zachary Zwald, University of Houston

• Ryan P. Kennedy, Ohio State University