The manually operated troll factories in Russia represent a bygone chapter in Russian propaganda. Yevgeny Prigozhin’s “Internet Research Agency” in St. Petersburg produced and operated fake social media profiles. Finnish investigative journalist Jessikka Aro managed to interview employees of the troll factory in an undercover operation and gained insight into the mechanism of this informational sweat shop. Shortly after the Russian invasion, Russia launched a qualitatively new disinformation campaign. Deceptively similar “lookalikes” were created based on the source code of websites such as FAZ, Spiegel, Welt, Süddeutsche Zeitung, and others. The address of the fake site was only slightly altered through so-called “typosquatting,” e.g., faz.ltd instead of faz.de. The fake web address was circulated and advertised by bots on social media. The fake content was created by AI. Although the language is correct, the software makes embarrassing mistakes that can be traced back to machine translation routines from Russian. For example, the fake articles repeatedly referred to “Bundeskanzlerin” Scholz because Russian does not distinguish between the male and female forms of “chancellor” and the AI's learning process was still stuck in the Merkel era. The German foreign minister was often referred to as “Berbock” – a retranscription of the Russian spelling of Baerbock. Today, we can observe Russian efforts to corrupt the large language models themselves. These language models are used to train artificial intelligence. This has fatal consequences: If the fundamental knowledge for the explanation of the world is already stored in the Kremlin version, there is no longer any need to give individual reports a pro-Russian spin. Instead, the task is completed in the very first step: anyone who now launches an AI-supported search query on the internet will receive a politically sterling answer. Finally the paper will address the question of the political nature of the Russian regime. Is it a “spin dictatorship” as Sergei Guriev and Daniel Treisman argue? Or is it rather a Neo-Soviet model that relies on the “volonté générale” that has to be manufactured by AI technologies and sophisticated reception modes?

The rapid evolution of Large Language Models (LLMs) has a profound impact on the processing of information in the Russian authoritarian context. Pro-Russian narratives and Kremlin propaganda find their way into the outputs of LLMs and AI-based search engines. LLMs are shaped in several stages. They use specific data gathering methods, go through pre-training, supervised finetuning and finally apply retrieval mechanisms to produce the final answer that users eventually receive. These pathways can be manipulated and influenced at every stage. The processing of information in LLMs and AI-based search engines represents an opportunity for Russian political engineers who aim to shape the public sphere according to their ideological preferences. The complete training process of a LLM is often opaque due to their proprietary nature. However, there exist open source models where not only the model weights but the complete training process has been described. We take the Olmo model developed by The Allen Institute for Artificial Intelligence and its training process to exemplify the entryways of LLM manipulation. The scales are staggering: For example the pre-training data of Olmo starts with more than 9 billion obtained documents that are then processed and filtered. The Institute’s transparency also allows to showcase a realistic estimation of the costs involved in the training process which has implications for where political engineers are most likely to start their manipulation first. Moreover, in the age of AI it is the technological that interacts with AI that becomes more and more essential. While the LLM’s training data influences the output, retrieval augmented generation (RAG) systems gain an outsized influence on what the LLM is actually used for. The paper reconstructs the conceptual framework of how these pathways can be manipulated following a utilitarian perspective where the number of readers or users represents the defining metric. Counterintuitively, the findings suggest that it is not the LLM that plays the most important role in the influencing of the information space, but the retrieval and crawler mechanisms.

Little is known about the national specifics of LLMs. In order to test the political spins that result from AI queries, this paper compares three Western LLMs with the Russian LLM models Alice AI by Yandex and Gigachat by Sberbank, and the Chinese model Quwen developed by Alibaba. The Western models were chosen based on the following factors: Mistral is the only preeminent European model, Claude by Anthropic and GPT by OpenAI are currently the leading models for coding. Quwen is one of the most recently developed Chinese LLM, while Gigachat and Alice are the only Russian models with comparable parameter size, the “settings that control and optimize a LLM output and behaviour (Belicic & Stryker, 2025).” Moreover, we will elaborate on the results we found when comparing LLMs from China, Europe and the US. We would expect that words, such as “king” and “queen” would appear closer on a visualised map than non-related terms. Hence, we will test differences between buzzwords from Kremlin narratives, such as Europe and Russophobia, Ukraine and Fascism, and NATO and Imperialism. Finally, the findings will be visualized in three dimensions: First, the vector distance in the models based on their origin will be shown. Second, the difference in the answers will be put on display depending on the languages in which the questions were asked. Third, the different outputs for Germany, the UK, and Russia will be highlighted by using a VPN client. Belcic, I., & Stryker, C. (2025, August 6). What are LLM parameters? IBM Think. https://www.ibm.com/think/topics/llm-parameters Verma, S., Bhatnagar, P., & Sharan, A. (2024). Text representation: A journey from the traditional vector space model to LLM. Special Proceedings of the 26th Annual Conference (pp. 127–141). ISBN 978-81-950383-5-0. ____________________________________________________________ Perplexity was used to create the citation style. Grammarly was used to edit the text.

This article examines how Russia articulates its identity, role, and international status vis-à-vis China in the Russian Far East (RFE) following the launch of the Special Military Operation (SVO) in February 2022. Treating the RFE as a critical regional arena where domestic and foreign policy intersect, the study analyzes how Russian political elites discursively manage cooperation with China while asserting sovereignty, regional authority, and great-power recognition. The analysis is grounded in constructivist International Relations theory, which conceptualizes identity, role, and status as co-constitutive and performatively enacted through discourse. Empirically, the article employs a computational–constructivist framework that integrates large language models (LLMs), retrieval-augmented generation (RAG), and density-based semantic clustering to analyze a corpus of 499 official statements, speeches, and social media posts produced by senior Russian political actors between 2022 and 2025. Rather than treating computational tools as theory-neutral, the study explicitly aligns them with constructivist epistemology, using them to operationalize interpretive concepts at scale while preserving contextual grounding and theoretical coherence. The findings reveal a highly consistent elite discourse in which Russia presents itself simultaneously as a cooperative and reliable partner to China, a sovereign and independent regional power, and a stabilizing actor in Northeast Asia. Cooperation in infrastructure, energy, and regional development is foregrounded, yet framed as a strategy for managing asymmetry rather than accepting subordination. Russia’s status claims oscillate between assertions of equality and implicit recognition of China’s growing economic and demographic weight, producing a relationship best characterized as managed interdependence. Historical memory and civilizational narratives are repeatedly mobilized to legitimize cooperation and mitigate status anxiety, stabilizing an otherwise indeterminate hierarchy. The article contributes to constructivist scholarship by demonstrating how identity, role, and status can be empirically operationalized through computational discourse analysis without sacrificing interpretive depth. Methodologically, it advances debates on text-as-data in International Relations by showing how generative AI can enhance, rather than replace, theoretically grounded qualitative analysis. Substantively, it offers new insights into how great powers reproduce global identities through localized regional discourse under conditions of geopolitical constraint.