Historical analogies as markers of decisions: an LLM-assisted analysis in foreign policy

Historical analogies recur in leaders’ discourse during crises as policymakers rely on familiar frameworks to navigate complex foreign-policy decisions and to create a sense of continuity amid unprecedented challenges. To date, scholars have typically treated such analogies either as cognitive “cozy” frames that simplify choice under pressure or as rhetorical devices that legitimate action. This article advances a third perspective: historical analogies as markers of emerging choices. It asks whether introducing and later repeating a specific past-to-present mapping can indicate the consolidation of a policy direction before it becomes visible in formal announcements or overt actions. The article applies an LLM-assisted workflow to texts from three presidents. An OpenAI GPT-based screening procedure is applied to 1100 official documents: 499 from U.S. President Bill Clinton, 375 from Russian President Vladimir Putin, and 226 from Chinese President Xi Jinping to identify candidate analogies. Once all outputs were manually verified and coded as cognitive, rhetorical, or signaling, the procedure yielded a small set of case-linked positive instances: six Clinton documents (1993–1995), three Putin documents (2013–2020), and five Xi documents (2015–2024). The workflow is designed to identify a limited number of highly relevant instances for close analysis instead of estimating the overall prevalence of analogical rhetoric. Across the three cases, the detected analogies align with subsequent trajectories; Clinton’s references to U.S. President Harry Truman appeared to foreshadow NATO’s enlargement, Putin’s Versailles frame aligned with his policy of reshaping Europe’s security landscape, and Xi’s “century of humiliation” has framed his Taiwan question. The findings suggest that historical analogies can provide an additional open-source indicator for experts to assess the intentions of states on the global stage. Methodologically, the article offers a replicable protocol for combining LLM-based screening with expert verification to support transparent analyses across large textual corpora.