Förderung für Forschung mit Azure-Services

Organisational unit: Faculty of Computer Science, Research Group Security and Privacy

Abstract:

We will investigate how optimization techniques used for deploying AI models on edge devices affect their security against various types of attacks. Many real-world systems apply quantization and related methods to reduce model size and accelerate inference, but the security implications of these transformations are still not well understood. In this project, we will systematically compare full-precision models with quantized variants to measure how optimization changes both predictive performance and robustness against prompt injections and other types of attacks.

Our evaluations will cover different model classes, including large language models (LLMs) and vision-language models (VLMs), and will include extensive attack simulations under realistic threat assumptions. We will study trade-offs between accuracy, computational efficiency, and resilience against attacks, and we will analyze whether certain quantization settings or model families show consistent security patterns. The results will provide practical guidance for deploying smaller and more energy-efficient AI systems while maintaining reliable security properties in real-world environments.

Organisational unit: Department of Government

Abstract: 

This project examines the spatial distribution and strategic use of election posters during the 2025 German federal election, focusing on the district of Friedrichshain-Kreuzberg in Berlin. While previous research has analyzed poster design and messaging, the geographic placement of campaign posters has received limited attention due to the challenges of collecting and processing spatially detailed data. To address this gap, the project builds on a manually collected dataset of more than 10,000 georeferenced photographs of election posters. 

As an initial step, a random subset of 500 photographs has been manually double-coded, providing validated annotations on poster presence, recurring poster motifs, and substantive content categories. This manually coded subset serves as both a validation benchmark and a training resource for the next phase of the project, which focuses on scaling the analysis to the full dataset using automated methods. Specifically, the automated pipeline proceeds in three steps. First, election posters are detected and extracted from raw photographs using pre-trained computer vision models. Second, the extracted poster images are assigned to recurring poster motifs, reflecting repeated campaign designs, using image-based similarity measures and classification models. Third, the substantive content of each poster motif is classified, including party affiliation, issue focus, and degree of personalization. This final classification step combines classical machine-learning approaches with AI-supported methods, including image-based analysis (e.g., face detection to identify candidate-centered posters), text extraction via OCR, and natural language processing enhanced by LLM-based coding to interpret and classify textual elements on the posters. 

To implement this hybrid pipeline, the project employs a combination of offline local data collection and cloud-based computation in Microsoft Azure. Pre-trained computer vision models, classical machine-learning classifiers, and AI models (including LLMs) are applied for poster and face detection, motif recognition, and content classification. The largescale processing of high-resolution images and multimodal model application exceeds available local computing resources and cannot be efficiently supported by existing institutional IT services, making Azure’s scalable AI and compute services essential. The project’s innovative contribution lies in transitioning from manual coding to a largely automated, reproducible, and scalable workflow for analyzing offline campaign strategies. The project team is currently preparing a follow-up research proposal that aims to extend this approach to substantially larger image datasets in future election campaigns.

Organisational unit: Faculty of Business, Economics and Statistics, Department of Business Decisions and Analytics

Abstract:

Computing delivery routes, scheduling production lines, and allocating humanitarian supplies are all challenging combinatorial optimization problems that organizations must routinely solve. Traditionally, experts design algorithms through a time consuming process of trial and error. Recently, our research group developed methods that automatically generate optimization approaches, called heuristics, using large language models (LLMs) given a problem specification and data. To date, this work has focused on routing problems, such as package delivery. This project extends our technique to broader classes of optimization problems. It will employ iterative heuristic discovery powered by extensive LLM queries, exploring different models and search algorithms to produce high quality heuristics that match or surpass state of the art human designed methods. Our LLM based discovery process democratizes access to advanced optimization, enabling organizations to tackle complex problems without specialized consultants. Moreover, the heuristics produced by our approach deliver higher quality solutions in less time than competing human crafted techniques.

Organisational unit: Faculty of Business, Economics and Statistics

Abstract: 

How political elites and news media frame intergroup relations — as cooperative or competitive — has far-reaching consequences for polarization, public attitudes, and policy preferences. While surveys capture how individuals perceive intergroup dynamics, no systematic, high-frequency measure exists for the narratives they are exposed to. We launch a research agenda to construct the first cross-national index of how intergroup relations are framed in public discourse.

Using large language models, we develop a classification framework that captures both the nature and intensity of intergroup framing in political and media texts across countries, policy domains, and media formats. The resulting index will enable researchers to study how the framing of intergroup relations in public discourse covaries with — and potentially shapes — political outcomes, policy shifts, and economic shocks.

Organisational unit: Faculty of Psychology, Department of Occupational, Economic, and Social Psychology

Abstract:

Background: Health information is often communicated under severe time pressure in busy clinical settings and in language that exceeds many patients’ health literacy levels. As a result, patients frequently leave medical encounters uncertain about critical decisions affecting their health — including decisions about vaccination. Generative AI-chatbots powered by Large Language Models (LLMs) offer a promising way to extend access to clear, tailored, and on-demand medical information beyond medical settings. However, despite their rapid adoption, there is little evidence that currently available chatbots can support informed decisionmaking while reliably adhering to established health-communication guidelines.

Research set-up: By embedding AI-assisted communication within a controlled decision environment, this project moves beyond speculative claims and provides rigorous causal evidence on whether—and under what conditions—generative AI can meaningfully improve vaccination decision-making. To this end, we employ an established vaccination decision-making paradigm, in which participants face a realistic decision scenario involving a fictitious infectious disease for which a vaccine is available, requiring them to weigh disease risk against potential vaccine side effects. Participants’ choices carry tangible consequences within the experiment, affecting their accumulated “fitness points,” which are converted into monetary payoffs and serve as a proxy for health outcomes. To inform their decisions, participants are given access to a chatbot constrained to a carefully curated, scenario-specific information base. Specifically, using Azure Open AI services combined with Azure Container Apps, a GPT-4o model will be trained on epidemiological information regarding the disease, the vaccine, and the broader decision context.

Research program: In Study 1, we evaluate whether a generative AI-chatbot prompted to use simple, plain language can improve users’ understanding and decision-making in a vaccination context. Study 2 moves beyond language alone by testing how both user framing and chatbot behavior shape outcomes: participants are either instructed to interact with the chatbot as if it were a human health professional or not, and the chatbot is either instructed to follow established medical communication guidelines or not. Across both studies, we assess outcomes central to effective health communication, including trust in the chatbot, knowledge about the disease and vaccination, decisional conflict, and perceived usability. Crucially, we also directly evaluate the chatbot’s adherence to simple language and medical communication guidelines, respectively. By systematically isolating these design factors, this project generates much-needed empirical evidence on how generative AI can—and cannot—be configured to support vaccination decision-making.

Organisational unit: Department of Government

Abstract:

This project examines Austria’s unusually large and hard to oversee body of constitutional legislation, exploring its causes and consequences. By integrating comprehensive data from the Austrian Legal Information System, parliamentary records, and Constitutional Court materials, the project aims to explain how political actors choose between ordinary and constitutional legislation and how these choices shape long-term policy constraints. While related questions have often been addressed descriptively, largely by scholars of law, this project advances the field by applying quantitative research designs and by analyzing Austria’s constitutional corpus from a political science perspective. In earlier project phases, we assembled several datasets covering all federal laws, the universe of constitutional provisions, and all cases of constitutional review in Austria since 1945. Because these data originate from a variety of sources and span multiple decades, much of the material is unstructured and lacks consistent metadata. Recently, we successfully applied a Large Language Model (LLM) to classify all constitutional-rank articles into policy areas, enabling us to track changes in constitutional regulatory density in different policy areas over time. Currently, we are starting to analyze all legislative initiatives in Austria since 1945 – nearly 20,000 initiatives available as machine-readable files but, for most of the period, without meaningful metadata. The objective is to analyze the factors that determine whether an initiative is adopted with constitutional rank, enacted as ordinary legislation, or fails in the legislative process. Additional funding for Microsoft Azure, which we would use on OpenAI in Azure, would enable us to use it to categorize these initiatives into different policy areas and enhance them with additional variables necessary for our analyses.

Organisational unit: Faculty of Informatics, Research Group Security and Privacy

Abstract: 

Website phishing remains one of the most widespread and damaging forms of online fraud, enabling attackers to steal user credentials, financial data, and other sensitive information by imitating legitimate online services. As attackers increasingly adopt realistic visual designs (e.g., logos, layouts, branding elements) alongside adaptive content and dynamic behaviors, traditional rule- or URL-based filters are becoming less effective. Conventional machine-learning approaches for phishing detection rely on extensive feature engineering and require frequent retraining.

This project aims to assess the potential of generative AI, particularly multimodal models, to reason jointly over textual, structural, and visual inputs to detect malicious websites. Such models reduce the need for handcrafted features and offer more adaptive and generalizable detection mechanisms. We will systematically benchmark a diverse set of open-source (e.g., Llama, DeepSeek, Qwen, Gemma) and proprietary (e.g., GPT, Grok) (vision) language models (LLMs/VLMs) for phishing website detection and classification. The evaluation will cover multiple modalities such as HTML content, URLs, and website screenshots, and employ publicly available datasets comprising tens of thousands of samples.

While locally deployed (smaller) vision models provide advantages in control, adaptability, and cost predictability, large cloud-scale models typically offer superior performance and scalability. To quantify these trade offs, we will conduct extensive experiments under several configurations:

1. Mixture-of-experts setups, combining complementary LLM/VLM models,
2. Hybrid integration with classical ML approaches, using OCR and image-based
   feature extraction,
3. Combining LLMs/VLMs with classical ML, where LLM outputs serve as features for
   downstream classifiers, and
4. Knowledge-augmented reasoning, incorporating external sources such as RAG and
   GraphRAG.

Selected models will be fine-tuned on phishing-specific data, with all experiments incorporating trace logging and transparent documentation to ensure reproducibility, interpretability, and fair benchmarking. The project will systematically assess performance, cost, and scalability trade-offs, laying the foundation for advanced AI-assisted cybersecurity research. All results, including detailed setup instructions, will be published open access.

Organisational unit: Faculty of Physics, Research Group Nanomagnetism and Magnonics

Abstract: 

Spin-wave dispersion analysis is the fundamental first step in characterizing magnetic eigenmodes in a given material and geometry. It covers a broad range of research fields, including quantum magnonics and microwave spintronics. Importantly, dispersion calculations also form the basis for developing complex magnonic data-processing units, including inverse-designed devices, an approach increasingly relevant for AI inspired and neuromorphic computing [1]. Despite its importance, such simulations typically require specialized expertise and locally installed software, limiting accessibility and reproducibility. 

This project aims to develop and operate an open-access, cloud-based spin-wave dispersion calculator that allows users to configure, run, and analyze simulations through a web interface. The software will be actively used by two research units at the University of Vienna: the Nanomagnetism and Magnonics group (Univ.-Prof. Dr. Andrii Chumak) and the Physics of Functional Materials group (Univ.-Prof. Dr. Dieter Suess). It will be also openly accessible to external collaborators and researchers worldwide, in line with modern principles of open academic knowledge exchange. The platform is implemented using Microsoft Azure services. An Azure Static Web App provides the frontend for defining material parameters, geometries, and excitation conditions [2, 3]. Simulations are executed in an Azure Container App using the open-source TetraX micromagnetic solver [4] to compute spin-wave dispersion relations [5], while simulation data and logs are stored via Azure File Share to ensure reproducibility. 

During the funding period, the platform will be extended beyond basic dispersion calculations toward more complex simulations of magnetization dynamics and collective excitation phenomena. These developments will require increased computational resources and will utilize Azure Virtual Machines for scalable CPU and GPU computing. In parallel, container orchestration will be migrated to Azure Kubernetes Service (AKS) to improve resource management and scalability for multiple concurrent users. By combining cloud-based infrastructure with a well-established opensource simulation engine, MaDiVie will provide a reliable, extensible, and reproducible research tool for the international magnonics community.
 

References

[1] Chumak, A. V., et al., Advances in magnetics roadmap on spinwave computing, IEEE Transactions on Magnetics 58, 1–72 (2022). 

[2] Frontend website: https://www.madivie.at  

[3] Frontend GitHub repository: https://github.com/GIGAluckman/WebDispersionCalculator 

[4] Körber, L., et al., Finite-element dynamic-matrix approach for spin-wave dispersions in magnonic waveguides with arbitrary cross section, AIP Advances 11, 095006 (2021). 

[5] Backend GitHub repository: https://github.com/GIGAluckman/WebDispersionCalculator-backend 

Organisational unit: Vienna University Library and Archive Services, Library - Research and Publication Services

Abstract:

For several years, a digital library and a digital bibliography on literature in translation have been established at the Vienna University Library (Department of Repository Management PHAIDRA Services) under the name DLBT. The Phaidra-add-on DLBT uses the YARM package as technical infrastructure and Phaidra as a long-term repository.

The DLBT not only collects data on translated literature, but also texts (and other assets) on their reception. Currently, over 65,000 translations, adaptations and reception documents are listed in the DLBT and more than 24,000 digital copies are available for research.

As part of the project ‘iConTxt - AI for the DLBT’ (7 European partners, funded by the Taalunie), led by Herbert Van Uffelen, the possibilities and limits of using artificial intelligence for the DLBT are being explored and a new YARM package (iConTxt) is being developed for the DLBT. The new package uses AI to support users in entering data and to improve the quality and accessibility of the information available in the DLBT. During the quality improvement process, the metadata are checked and ‘improved OCRs’ of the scanned texts are created.

iConTxt creates further English translations and English summaries for every reception text in the DLBT and creates relationships to the translations listed in the DLBT. In combination with the results of specific web queries, the DLBT-specific ‘knowledge pool’ generated by iConTxt will be used for Retrieval Augmented Generation of new information on authors, translators, publishers and translated titles mentioned in the DLBT.

Organisational unit: Faculty of Business, Economics and Statistics - Vienna Center for Experimental Economics

Abstract: This project aims to develop and validate a robust framework for stress testing chatbots powered by large language models (LLMs) designed to serve as instruction bots in experimental economics. These bots are intended to provide participants with accurate and unbiased instructions, ensuring experimental integrity and minimizing noise in collected data. While general-purpose tools exist for stress testing commercial chatbots, the unique demands of experimental economics require the development of a specialized stress testing protocol.

The framework involves the creation of two LLM-based agents: the Stress Tester Bot (STB) and the Evaluator Bot (EB). STB acts as a synthetic subject, attempting to “break” the Instruction Bot (IB) by eliciting biased, false, or irrelevant responses. EB evaluates these interactions, identifying problematic outputs from IB to ensure its reliability and robustness.

Two potential approaches for developing STB and EB are considered. The first approach leverages multi-agent pipeline structures using larger, off-the-shelf LLMs (e.g., 4o or o1), prioritizing accessibility, flexibility, and scalability. This method eliminates the need for fine-tuning and data collection, making it an appealing and computationally efficient option for the broader social science community. However, whether this approach can achieve desirable performance for stress testing without the need for fine-tuning and data collection remains an open empirical question.

The second approach involves fine-tuning smaller LLMs (e.g., 4o-mini) on data collected from incentivized experiments where subjects are instructed to “break” the IB. While this approach is likely to yield more reliable and tailored outcomes, it requires extensive data collection, computational resources, and fine-tuning for each specific experiment, potentially limiting its scalability and accessibility, particularly for resource-constrained researchers.

The project will initially pursue the multi-agent pipeline approach and only consider fine-tuning if the first approach fails to meet reliability standards. Funding is sought to access computational resources, via Azure infrastructure and API credits, to support development and testing. This research addresses a critical gap in the field, providing a scalable and objective standard for stress testing LLM-based tools in experimental economics, ultimately enhancing the reliability of experimental outcomes and promoting the broader adoption of instructional chatbots in social science research.

Organisational units:

• Vienna University Library and Archive Services – Department for Bibliometrics and Publication Strategies,
• Vienna University Library and Archive Services – Library-Communications,
• Zentraler Informatikdienst – IT-Support for Research

Abstract:

Both projects delve into the diverse potential of AI to analyze and reshape scientific abstracts, creating innovative services for the University of Vienna.

The project “SciTextMatch” identifies cross-university collaboration potential via two complementary pipelines: a document-based pipeline and a candidate-based pipeline.

The document-based pipeline compares publication titles and abstracts from selected principal-investigator (PI) candidates across universities. Similarities computed with TF-IDF (a statistical method) and SPECTER2 (an embedding model) produce a ranked list of PI–PI pairs, which is validated with OpenAI’s GPT-4.1 and mapped to the themes of upcoming Horizon calls.

In parallel, the candidate-based pipeline uses OpenAI’s o3 deep-research model to conduct a Bing-grounded analysis of the full (non-sensitive) PI lists from participating institutions, following a predefined analysis protocol to identify PI–PI candidate pairs.

Finally, outputs from both pipelines are reconciled: matches confirmed by both are accepted automatically, while discrepancies are flagged for human review.

The podcast project “Talking Abstracts: Echoes of Knowledge” distills complex research findings into engaging content, making it accessible to various audiences. Internally, it helps in the efficient identification of relevant content for Communications, thereby enhancing the university’s outreach efforts. It promotes interdisciplinary networking and supports the rectorate in strategic planning and decision-making.

This initiative is driven by a cross-departmental team: Janos Bekesi (IT Support), Martin Gasteiner (UB Communications), and Christian Gumpenberger, Lothar Hölbling, Ines Konnerth, and Lilian Nowak (Team Bibliometrics and Publication Strategies).

Organisational unit: Centre for Teacher Education – Department for Teacher Education

Abstract:

The Computational Empowerment Lab is part of the Digitalisation in Education unit and is located at the Centre for Teacher Education at the University of Vienna. The Lab sees itself as an intellectual and practical space for the realisation of ideas, research and projects to promote computational empowerment. This is done using a wide range of technologies. In the first stage of the project, a self-trained chatbot (self-trained text-based AI) will be developed and implemented, which is specially designed for questions about CE-Lab equipment and CE-Lab questions from students. This chatbot is intended to provide students with fast and reliable help by responding to frequently asked questions and problems in the CE-Lab environment. By integrating this chatbot into the CE-Lab, the existing support offer for the use of the various technologies is to be expanded and direct and efficient support for students is to be ensured, resulting in an improved learning experience and a reduction in the workload of teaching staff. In the second stage of the project, which is considered a “nice to have”, didactic recommendations around the CE-Lab will be developed on the basis of (own) scientific and didactic literature, specialised articles and brochures (e.g. the Computational Empowerment in Practice brochure). These recommendations are intended to help students gain a deeper understanding of the use and handling of lab technologies. The reason for the realisation of this project lies in the existing bias, lack of transparency and lack of comprehensibility of existing models. This project aims to create a pilot for a specialised educational LLM (Large Language Model) that addresses these challenges and provides transparent and comprehensible support in the education sector.

Organisational unit: Faculty of Informatics – Research Group Security and Privacy

Abstract:

The detection of grievances in social media texts is critical for understanding their role in shaping public discourse and their potential to escalate into collective unrest or harmful actions. Grievances are psychological responses to perceived injustices and have been shown to be precursors to radicalization or mobilization for contentious actions (Scrivens, 2022). This project develops a hybrid AI framework that combines Microsoft Azure’s Cognitive Services, including Text Analytics and the Azure OpenAI Service, with on-premises infrastructure to detect grievances in high-risk social media texts. This approach ensures scalability, flexibility, and compliance with data protection regulations. By integrating Azure Machine Learning for model training and deployment, and utilizing Azure’s Explainable AI tools, the project prioritizes interpretability and transparency in detecting grievance-related patterns. The outcomes will provide actionable insights into grievance propagation and its impact on digital discourse, contributing to computational social science and the ethical design of AI systems. This framework will also serve as a replicable model for combining cloud-based AI and local infrastructure in high-stakes text analysis applications, providing valuable insights for policymakers, platform moderators, and researchers in managing online discourse.

Organisational unit: Faculty of Psychology – Department of Cognition, Emotion, and Methods in Psychology

Abstract:

Large-scale surveys are crucial for understanding the cultural dynamics of populations over time. However, applying these tools to study long-term historical trends remains a recent development. Historical psychology has emerged as a field that uses past fiction as “cognitive fossils” to uncover insights into societal values, emotions, and moral frameworks of bygone eras [1].

Traditional computational methods, such as bag-of-words and topic modeling, have revealed intriguing trends, such as rising sentiments towards cooperation preceding the French Revolution and English Civil War, followed by a decline [2]. However, these methods fail to capture the contextual and nuanced use of language in historical texts [3]. Advances in large language models (LLMs) now offer tools to overcome these limitations. While foundational LLMs excel in modern language analysis, they are not optimized for historical texts, where linguistic conventions differ. Recent theoretical work proposes that these challenges can be addressed by developing Historical Large Language Models (HLLMs) [4].

This project aims to develop and compare two approaches to analyzing historical texts:

1. Fine-Tuning LLMs with Historical Corpora: This involves building HLLMs tailored to English, French, and German fiction from the 16th to 19th centuries [4]. It involves modifying the model’s parameters using techniques like LoRA (Low-Rank Adaptation) and PEFT (Parameter-Efficient Fine-Tuning) to adapt the foundation model efficiently to task-specific needs. Using Azure’s infrastructure, we will fine-tune the open-source Llama 3.3-70B on a pre-processed corpus of ~15 million tokens. The fine-tuning process will leverage NVIDIA A100 GPU (e.g., NC24ads A100 v4). After fine-tuning, the HLLM will classify texts across psychological dimensions, replicating Martins & Baumard (2020) [2] - which used bags of words.
2. Optimizing Prompt Engineering: This involves leveraging GPT-4 - the gold standard tool for text annotation in psychological sciences [5] - with refined instructions to analyze historical corpora without retraining [6], such as requesting that the annotation consider the year the text was produced. Prompt optimization will utilize Azure OpenAI Services to refine instructions for the LLMs iteratively. The context window constrains this approach and relies on precise and iterative prompt refinement.

This project will deliver a paper that describes both a methodological advancement and a replication of Martins & Baumard (2020). It will comprehensively evaluate fine-tuning versus prompt engineering across several dimensions: a) Comparison Against Student Classifications: The outputs of both approaches will be validated against human-annotated classifications of 3000 sentences. Metrics such as precision, recall, and F1 scores will quantify alignment with human evaluations, providing a precise measure of accuracy. b) Comparison Against Bag-of-Words Benchmark: The results from both approaches will be compared with the findings from the Martins & Baumard (2020) [2] study on democratic sentiments, which utilized traditional Bag-of-Words methods. c) Comparison of Cost and Time Efficiency

References 

[1]. Baumard, N., Safra, L., Martins, M. & Chevallier, C. Cognitive fossils: Using cultural artifacts to reconstruct psychological changes throughout history. Trends in Cognitive Sciences (2024).

[2]. Martins, M. D. & Baumard, N. The rise of prosociality in fiction preceded democratic revolutions in Early Modern Europe. Proc Natl Acad Sci USA 117, 28684 (2020).

[3]. Martins, M. D. & Baumard, N. How to Develop Reliable Instruments to Measure the Cultural Evolution of Preferences and Feelings in History? Frontiers in Psychology 13, (2022).

[4]. Varnum, M. E. W., Baumard, N., Atari, M. & Gray, K. Large Language Models based on historical text could offer informative tools for behavioral science. Proceedings of the National Academy of Sciences 121, e2407639121 (2024).

[5]. Rathje, Steve, Dan-Mircea Mirea, Ilia Sucholutsky, Raja Marjieh, Claire E. Robertson, and Jay J. Van Bavel. "GPT is an effective tool for multilingual psychological text analysis." Proceedings of the National Academy of Sciences 121, no. 34 (2024): e2308950121.

[6]. Dubourg, E., Thouzeau, V. & Baumard, N. A step-by-step method for cultural annotation by LLMs. Front. Artif. Intell. 7, (2024).

Organisational unit: Faculty of Business, Economics and Statistics – Department of Accounting, Innovation and Strategy

Abstract:

We intend to develop a prototype of a teaching assistant chatbot using Microsoft Azure infrastructure. The proposed project aims to integrate cutting-edge AI technologies into the teaching process, to enhance student engagement and learning outcomes. By leveraging Azure OpenAI Service, Azure Bot Service, and Azure App Service, the chatbot will be designed to fulfill the core functions of a teaching assistant, providing timely responses to student queries, offering tailored academic guidance, and supporting administrative tasks. It will be trained on teaching materials available to students and provide library access via ResearchGate and JSTOR. The primary goal for 2025 is to develop and evaluate a functional prototype that reliably performs the intended tasks. This initial phase will involve designing, deploying, and refining the chatbot using Microsoft Azure’s robust suite of AI and cloud services. Building upon the insights gained during the prototyping stage, the project’s subsequent phase will focus on investigating the educational impacts of deploying such a chatbot in a real-world classroom environment. By the end of 2025, I aim to initiate a systematic study to observe student interactions with the chatbot and evaluate the resulting learning outcomes. Using carefully crafted experimental variations, the study will identify causal effects of chatbot use on student performance and engagement. This research phase, planned for 2026, will provide empirical evidence regarding the efficacy of AI-powered teaching assistants in higher education.

Organisational unit: Faculty of Historical and Cultural Studies – Department of European Ethnology

Abstract:

The Regulatory Language Analyzer Pilot Project presents a focused investigation into how different jurisdictions conceptualize AI governance through their regulatory frameworks, concentrating specifically on comparative analysis between EU and US approaches. It is intended as a preliminary methodological exploration for an ERC Consolidator Grant proposal to be submitted by the PI in January 2025. This 7-month study leverages Azure's ML infrastructure to develop and validate a specialized language model for analyzing regulatory AI policy documents, serving as a proof-of-concept for larger-scale cross-cultural policy analysis.

The project employs a streamlined fine-tuning process using Azure OpenAI GPT-3.5 as the base model, focusing on English-language regulatory documents from the EU AI Act and US federal and state-level AI frameworks. Through a single-stage fine-tuning approach, the model will be optimized to identify and analyze key regulatory patterns, linguistic features, and policy concepts specific to AI governance. This targeted approach allows for rapid development and validation of the core analytical capabilities while maintaining sufficient depth for meaningful comparative analysis.

The research methodology emphasizes efficiency and foundational insights, with data collection and preprocessing focused on creating a high-quality parallel corpus of EU and US regulatory documents. The model's performance will be evaluated through essential metrics including regulatory framework classification accuracy, policy intention recognition, and basic cross-jurisdictional concept alignment. Technical implementation uses Azure ML compute with NC6s_v3 infrastructure, providing the necessary computational power for model development and optimization within the condensed timeframe.

This pilot study will deliver key insights into how two major regulatory approaches to AI governance differ in their linguistic and conceptual frameworks. The analysis will focus on identifying significant patterns in how these jurisdictions express core concepts such as accountability, transparency, and risk management in AI systems. The findings will provide an empirical foundation for understanding how different legal traditions approach AI governance while establishing a methodological framework that can be expanded to include additional jurisdictions and languages in future research.

The project's deliverables will include a validated model for regulatory language analysis, documentation of key linguistic and conceptual patterns identified in EU and US frameworks, and recommendations for scaling the analysis to additional jurisdictions. This pilot serves as a crucial first step in developing more comprehensive tools for understanding how cultural and linguistic differences shape AI governance approaches, while providing immediate practical insights for policymakers working on AI regulation in these key jurisdictions.

The condensed scope enables rapid development and validation of the core analytical framework while establishing a solid foundation for future expansion to additional jurisdictions and more nuanced cross-cultural analysis. This approach balances the need for meaningful results within a limited timeframe with the potential for broader application in future research phases.