On March 10, the Fletcher Russia and Eurasia Program hosted Adam N. Stulberg and Stephan De Spiegeleire. The two distinguished scholars presented their new project, RuBase, a collaborative initiative that uses AI methods to systematically evaluate Russian rhetoric surrounding deterrence, nuclear threats, and coercive diplomacy. With the goal of deciphering the meaning behind Putin’s statements, they also discussed the broader applicability of their methods for producing rigorous, efficient, and accurate analyses of geopolitical events.

Adam N. Stulberg is the Sam Nunn Professor and School Chair at the Sam Nunn School of International Affairs at Georgia Tech. He is an expert on Russian security and energy politics, contemporary great-power competition, and “gray zone” conflict in Eurasia. He previously served as a Political Consultant at RAND from 1987 to 1997 and as a Senior Research Associate at the Center for Nonproliferation Studies (CNS) at the Middlebury Institute of International Studies at Monterey from 1997 to 1998. He has also worked closely with former Senator Sam Nunn, drafting policy recommendations and background studies on future directions for the U.S. Cooperative Threat Reduction Program and on building regional and energy security regimes in Central Asia and the South Caucasus.

Stephan De Spiegeleire previously worked for the RAND Corporation for nearly ten years, with additional stints at Stiftung Wissenschaft und Politik and the WEU’s Institute for Security Studies. He began his career as an expert in Soviet policy before moving into work on strategic defense management, security resilience, network-centrism, capabilities-based planning, and the transformation of defense planning.

Background

The central question guiding the research was whether Russia’s frequent invocation of “red lines” represents meaningful strategic signaling, political theater, or some other form of coercive communication. By combining traditional strategic analysis with large-scale AI-driven text analysis, the project aims to move beyond the anecdotal interpretations often seen in other reporting and provide a more systematic understanding of Russian messaging.

Stulberg began with an overview of Russian “red line” messaging. Since the start of the war in Ukraine, Russian officials such as Vladimir Putin and members of the Russian elite have repeatedly warned that Western actions could cross Russian “red lines.” However, these statements often appear inconsistent or cryptic, as well as divorced from specific nuclear threats. The term itself is used frequently but often without precise definition, and the language surrounding red lines is ambiguous and formulaic while nuclear threats appear more systematic. The statements also at times appear disconnected from actual Russian military behavior or strategic logic of deterrence and brinkmanship.

Stulberg highlighted how this creates confusion over whether such statements should be interpreted as credible deterrent threats or as rhetorical signaling. In Western strategic theory, deterrence typically relies on clarity and credibility: threats must specify what actions are prohibited and what consequences will follow. Russian statements, however, often lack this clarity (as applied to threats, ines, and consequences), making them difficult for Western policymakers to interpret. This led to the central question the researchers sought to answer about “red line” rhetoric: are these statements ineffective political theater, as they are often portrayed in Western media, or are Russian policymakers operating with a different conception of coercive diplomacy in which such statements serve a strategic function?

Western Debates

The speakers then examined the current state of Western debate regarding “red lines” and nuclear-threshold signaling, which typically draws on classical deterrence theory from thinkers such as Thomas Schelling. They identified three major schools of thought: strategic bluffs, clear and substantive threats, and calculated brinkmanship.

The “strategic bluff” school argues that Russian threats are largely bluffs. According to this view, Russia uses rhetorical escalation to deter Western involvement but has little intention of following through, and the West should therefore remain firm and continue expanding military assistance to Ukraine. The second school, “clear and substantive threats,” argues that Russian statements may signal a genuine willingness to escalate, including the potential use of nuclear weapons. This perspective emphasizes caution and restraint in order to avoid triggering escalation. The third school, “calculated brinkmanship,” argues that the threats are designed to push adversaries toward compromise while leaving room for diplomatic off-ramps. This view aligns most closely with the classical Schelling tradition of “threats that leave something to chance,” though it does not necessarily prescribe a specific course of action. The problem with all three interpretations, the speakers argued, is that they assume the Russian government is operating according to Western strategic logic, which may not be the case.

Mirror Imaging Critique

The speakers also emphasized the danger of “mirror imaging,” in which Western analysts project onto adversaries the same frameworks and incentives that guide their own decision-making. Much Western scholarship assumes that Russian leaders use red line rhetoric in the same way Western policymakers would. However, drawing on their extensive background in Soviet and Russian studies, the presenters argued that Russian strategic culture may approach coercion and signaling in fundamentally different ways. Rather than issuing clear deterrent threats, Russian messaging may intentionally blur the boundaries of escalation. This ambiguity can shape the strategic environment by influencing how opponents interpret risk and uncertainty. The question of which school of thought is correct was a significant driver of Stulberg and de Spiegeleire’s decision to build an AI model to analyze Russian statements.

Russian Coercion Methods

The presenters suggested that Russian red line rhetoric should instead be understood through the concept of “reflexive control.” Developed during the Soviet era, reflexive control is a strategic approach focused on influencing an adversary’s perceptions and decision-making processes. Rather than coercing an opponent through clear threats, the goal is to manipulate the information environment so that the opponent voluntarily chooses actions that benefit the strategist. Within this framework, ambiguity and confusion are not weaknesses but strategic tools. Russian messaging may therefore be designed to introduce uncertainty into Western decision-making, shape perceptions of escalation risk, and encourage Western restraint without requiring the credible enforcement of threats. The speakers did note, however, that based on their analysis, this strategy appears ultimately self-defeating. At the same time, Putin’s explicit nuclear saber-rattling appears more consistent with “threats that leave something to chance.”

Analytical Methodology

The second portion of the presentation, led by Stephan De Spiegeleire, focused on the technical methodology used to analyze Russian rhetoric. The research team assembled an unprecedented corpus of official Russian communications using automated scraping tools, ultimately collecting approximately 250,000 documents from 36 official Russian sources, including the Kremlin website, government agencies, and official Telegram channels.

The team then pioneered the use of large language models to analyze these texts in multiple stages. First, in an initial filtering phase, the AI system scanned document segments to identify those potentially containing Real Red Line Statements (RRLs) or Nuclear Threat Statements (NTS). The flagged sections were then classified along approximately 30 different dimensions (source, target, intensity, theme etc.). Human researchers conducted rigorous quality-control checks throughout this pipeline to validate the AI classifications.

Once completed and validated, this process yielded a searchable, curated dataset of 1,924 Red Line statements and 357 nuclear threat statements that enabled systematic analysis of Russian rhetorical patterns. To deepen the analysis, the researchers integrated this with approximately 30 heterogeneous datasets documenting the war (conflict events; diplomatic signals; sanctions; military aid; territorial control; cyber operations; equipment losses; refugee flows; etc.). The combined data, comprising over 52M rows, was transformed into a temporal knowledge graph containing 1.1 million relational “triples” representing structured relationships among entities, actions, and contexts. This dataset was then analyzed through a graph neural network with 5.1 million parameters across 210 weekly snapshots to detect patterns over time. This temporal analysis allowed the researchers to determine whether Russian rhetoric was triggered by or responded to external events. They emphasized that this method made it possible to move beyond isolated statements and instead examine the broader complex dynamics of rhetorical escalation, representing a significant advancement over previous approaches in the field.

Findings

The speakers then outlined their findings and the conclusions practitioners might draw from them. One of the most striking results was that Russian red line rhetoric appears overwhelmingly reactive rather than proactive. The statistical analysis suggested, with 99.9 percent confidence, that Russian red line statements tend to occur in response to Western actions rather than preceding them. This may indicate that RRL and NTS statements do not operate according to the typical Western understanding of coercive diplomacy, but instead function as rhetorical responses to events themselves.

The presence of RRL and NTS statements was also identified as reflecting two distinct categories within Russian rhetoric. Red Line Statements (RRLs) tend to be formulaic responses to Western political actions or military aid to Ukraine, while Nuclear Threat Statements (NTS) are more strongly correlated with Ukrainian military actions, particularly attacks on Russian territory, and appear to carry greater weight.

The researchers also identified a recurring pattern they described as a “self-defeating rhetoric cycle.” In this cycle, the West takes an action, such as providing military assistance to Ukraine; Russian leaders respond with red line warnings; Western media outlets amplify these warnings; and Western governments then respond by increasing military support. Rather than deterring Western behavior, Russian rhetoric may inadvertently reinforce Western resolve, suggesting that red line statements may fail to achieve their intended coercive effects.

Policy Implications and Conclusions

Adam Stulberg and Stephan de Spiegeleire opened their discussion of the implications of these findings by emphasizing the risks of misunderstanding adversary rhetoric. Western analysts often assume that threats must be credible and clearly defined to influence behavior. However, the Russian approach may prioritize ambiguity and psychological influence over traditional deterrence logic. As a result, Western responses to Russian rhetoric may at times misinterpret the purpose of these statements, at the risk of potentially blundering into escalation that neither side is seeking to manipulate.

Several policy lessons followed. First, analysts and policymakers should avoid mirror imaging and be cautious about assuming that adversaries share the same strategic logic. Second, analytical methods should improve. Traditional approaches to intelligence analysis often rely heavily on selective interpretation of statements, whereas AI-driven analysis can help identify patterns across large datasets and reduce the risk of cherry-picking evidence, particularly when analysis is grounded in thousands of linked statements and events. Third, the researchers emphasized the value of open data sharing, arguing that accessible and consistently maintained datasets can help social scientists work more efficiently and pursue more reliable findings.

To conclude the presentation, the speakers answered questions from the audience, including whether actual military events could be linked to rhetoric through a similar model, how confident they were in their conclusions, and how tone might be analyzed across other databases.

This article is republished from Tufts Russia and Eurasia Program. Read the original article.