(iPoster) Algorithmic Diplomacy

• In Event: Poster Session: Various Interesting Foreign Policy

Abstract

Diplomats and soldiers have already started deploying artificial intelligence to implement foreign policy. The use of such tools raises essential questions about algorithmic governance and the potential bias of AI-enabled recommendations. In this study, we evaluate six publicly available foundation models on their performance across a battery of questions teasing out potential biases across a spectrum of two foreign policy axes, which we call the Models of Foreign Policy Framework (MFPF), i.e., Liberal to Conservative and Isolationism to Interventionist. Additionally, we divide the potential responses between the Diplomacy, Information, Military, and Economic levers of national power. The paper presents two analyses: (1) a “within” DIME category analysis using a neural networked propensity scoring matching approach to estimate an MFPF score presented on a two-axis scatter plot. This "within DIME" plot will show how foundational models vary in their recommendations about using each element of national power. (2) A “between” DIME category analysis using a maximum likelihood estimator creates expected value scores across the DIME and MFPF interactions and will be presented as a set of six radar plots. These "between DIME" plots will show how the foundational models vary in their recommendations about whether the nation-state should use different levers of national power. These analyses highlight the potential underlying bias towards different policy bundles, which may inadvertently impact forging policy actors’ strategic decisions. The results discuss how bias may play a part in shaping algorithmic AI-driven foreign policy decisions, with special attention paid to the potential impact on future strategic diplomatic action recommendations.

Author

• Brandon DeBruhl, University of Southern California