Framing Influence: How Russian Trolls Use Morality and Emotion to Mimic U.S. Political Discourse

Abstract

Efforts to counter foreign influence on social media increasingly assume that harmful activity can be identified and mitigated through automated analysis of language. This dissertation shows why that assumption is flawed. Analyses of Russian Internet Research Agency activity on Twitter demonstrate that the rhetorical features most strongly associated with foreign influence substantially overlap with legitimate domestic political speech, making content-based detection analytically informative but operationally unsafe for enforcement.Based on these findings, this dissertation argues that policy responses should not rely on content-only detection or automated moderation to counter foreign influence. Instead, effective responses should prioritize transparency, attribution, and structural platform interventions that reduce the reach and impact of coordinated foreign influence behavior without requiring fine-grained judgments about the permissibility of individual messages. These recommendations reflect the high false-positive risks associated with content-based enforcement under realistic prevalence conditions and the corresponding normative and legal costs of suppressing lawful political expression. Russia’s information strategy integrates cyber operations with psychological influence to shape foreign publics and political processes. Against this backdrop, this dissertation examines the rhetoric of Russian influence campaigns using content-only features and interprets the results through a comparative U.S.–EU policy lens. Study 1 analyzes engagement in Russian-attributed tweets and estimates the independent and joint effects of Moral Foundations Theory language, partisan word use, and emotions. Moral framing increases engagement beyond partisanship and emotion, with Authority and Care showing the strongest positive associations, while partisan intensity and discrete emotions such as anger and fear provide additional but smaller effects. These results establish that moral rhetoric is not just correlated with, but incrementally predictive of, higher engagement in Russian messaging. Study 2 asks what Russian messaging most closely resembles at the tweet level. Tweet-level classifiers trained on a balanced U.S. sample using Moral Foundations, emotions, and partisan intensity achieve about 51 percent accuracy and indicate that Russian tweets are most often classified as non-elite content, resembling politically engaged or random users more than politicians. A tweet-level TF–IDF model using a 10,000-term unigram vocabulary (the 10,000 most frequent words in the U.S. training corpus, each represented as a separate feature) performs substantially better on the held-out U.S. test set for the same three-class author-type task (accuracy ≈ 67 percent, macro-F1 ≈ 0.67). When this model is applied to Russian tweets, it assigns roughly 46 percent to the random-user class, 49 percent to the engaged-user class, and 5 percent to the politician class. Together, these tweet-level results indicate that Russian messages occupy a broad, non-elite band of civic discourse: they exhibit moral framing and institutional vocabulary associated with domestic political conversation, yet they do not consistently separate from highly engaged U.S. users on content alone. Study 3 aggregates the same content-only features to the account level and evaluates whether they can support prevalence-aware intervention thresholds, that is, thresholds that take into account how rare Russian accounts are in the broader population and therefore how easily detection errors can overwhelm true positives. Account-level rhetorical models using percent-present Moral Foundations Theory and emotions, alongside the mean partisan word count per tweet, achieve accuracy around 77–78 percent with macro-F1 ≈ 0.80 for U.S. accounts. An account-level TF–IDF model, a representation that converts each account’s combined tweet text into a vector of term weights based on how distinctive each word is within the overall data, performs slightly better, reaching roughly 83 percent accuracy and macro-F1 ≈ 0.87. Across both models, Russian accounts consistently appear non-elite: the rhetorical model predicts approximately 55 percent random, 43 percent engaged, and 2 percent politician, while the TF–IDF model classifies about 86 percent of Russian accounts as random and 14 percent as engaged, with almost none labeled as politician. Feature comparisons show that partisan and institutional terms concentrate most heavily in elite accounts; Russian accounts use these terms less frequently than politicians but at levels that overlap with engaged users. Moral Foundations profiles also differ modestly, with Russian accounts showing slightly lower Fairness and Loyalty and somewhat higher Sanctity- and Disgust-related activation than domestic users. In the final part of Study 3, I simplify the classifier’s account-level predictions to make them suitable for policy evaluation. The model assigns each account a probability of belonging to the politician, engaged user, or random classes. To translate these into more interpretable measures, I create two composite scores. The political-style score combines the probabilities of the politician and engaged user classes and reflects how similar an account is to typical political actors. The random-style score is the model’s predicted probability that an account resembles a random user. Using these measures, I evaluate policy risk under realistic base-rate conditions by estimating how many U.S. accounts would be mistakenly flagged if platforms attempted to detect Russian accounts using content-only signals. The results show that content-only rhetorical/lexical models (as constrained here) are insufficient for enforcement thresholds. Setting thresholds to capture 70–90 percent of Russian accounts produces extremely high false-positive rates among U.S. users. Under political-style scoring, thresholds in this range flag nearly all U.S. politicians and engaged users, along with more than 70 percent of ordinary accounts. When realistic Russian prevalences are introduced at 0.1, 0.5, or 1.0 percent, the positive predictive value remains below 1 percent, meaning that hundreds of U.S. accounts would be incorrectly flagged for every Russian account correctly identified. Random-style scoring performs slightly better for political elites but still labels most ordinary and engaged users as Russian-like and yields comparably low positive predictive value. Together, these prevalence-aware results show that even reasonably accurate content-only models cannot, on their own, produce operational thresholds that avoid sweeping up large amounts of legitimate domestic speech. The concluding comparative case study interprets these patterns under U.S. constitutional constraints and the European Union’s regulatory approach. In the EU, the Digital Services Act framework emphasizes systemic risk assessment (ongoing evaluation of platform-wide risks such as disinformation), transparency obligations (requirements that platforms disclose how algorithms work and how content is moderated), and researcher access (mandated data access for external auditors). These tools focus on platform processes rather than individual pieces of content. Feasible adaptations for the U.S. context include using rhetorical models for offline analysis and investigative triage and combining content-based features with structural or behavioral indicators, such as coordination patterns, bot-like posting rhythms, provenance metadata, or network ties, that provide more reliable signals than text alone. By contrast, direct enforcement based solely on rhetorical style presents a high risk of over-capture, meaning that legitimate domestic accounts, particularly politically active or morally expressive users, would be swept up alongside the very small number of Russian accounts. This risk is amplified by low base rates for foreign actors and by First Amendment protections for the right to receive information, which make content-based restrictions especially difficult to justify. In sum, these findings show how moral, emotional, and partisan rhetoric shape engagement in Russian influence campaigns; how Russian tweets and accounts resemble the communication patterns of non-elite domestic users; and why effective policy responses must pair rhetorical insights with additional signals and institutional safeguards to mitigate harm without suppressing lawful political expression. All analysis code for this study, including data-processing pipelines, modeling scripts, and figure generation, is available in a GitHub repository located at https://github.com/mfeehan/PhD-Code.