We send regular updates about events at I2SC through our mailing list.
The Interdisciplinary Institute for Societal Computing offers a regular Lecture Series to bring together researchers of different academic fields to analyze and discuss the broad topic of society and technology. The Lecture Series is designed as a laboratory of interdisciplinary research to encourage cooperation and new research approaches. The series will feature a mix of speakers from Computer Science, Social Science, and Digital Humanities.
November 6, 2026
Leah von der Heyde (GESIS - Leibniz Institute for the Social Sciences)
The How and Why (Not)s of Silicon Samples for Political Polling
November 20, 2026
Zoë Greene (University of Strathclyde)
Who should I vote for? Why citizens use AI chatbots to learn about elections and how popular models respond
November 27, 2026
Pola Lehmann (WZB - Berlin Social Science Center)
The Best of Both Worlds - Improving Human Coding with Automatic Classification Models
January 22, 2027
Vera Schmitt (Johannes Gutenberg-University Mainz)
Beyond Transparency: Explainability, Alignment, and the Illusion of Control in High-Risk Information Processing
January 29, 2027
Fabienne Lind (University of Salzburg)
Negotiating Societal Challenges in Digital Publics: AI-Supported Approaches for Social Science Research
The Lecture Series is in building E1 7, Room 3.23, on the campus of Saarland University from 12:15-13:45.
If you want to meet one of our speakers on the day of the event, please contact us: hello[@]i2sc.net
If you are interested in our events and want to stay up to date, please subscribe to our mailing list here.
For Guest lectures not from the lecture series, please check our Guest Lectures page.
November 6, 2026
The How and Why (Not)s of Silicon Samples for Political Polling
Over the past years, parallel developments have been shaping, or shaking, public opinion research: political crises demand rapid insight into fast-shifting attitudes, while recruiting participants for high-quality surveys is becoming slower and costlier. At the same time, the proliferation of generative AI has opened up new possibilities for social simulations. As a result, so-called “silicon samples” – data not collected from humans, but generated by AI based on existing human data, have become a research focus of computational social science. In this talk, I will provide an overview of the potentials and pitfalls of this approach. After a short introduction of the logic behind silicon samples and their relevance for political polling applications, I will draw on a systematic review as well as own applications in election prediction to showcase key challenges in modelling human attitudes and behaviors with AI, concerning both cross-sectional and longitudinal aspects. Presenting the Total Simulated Survey Error (TS2E) Framework as well as ongoing work, I will explain why errors might occur and how we can evaluate the quality of silicon public opinion data. I will close with the many aspects that remain to be explored when it comes to the design, evaluation, and ethics of silicon samples, and where computer science and political science can work together to answer them.
November 20, 2026
Who should I vote for? Why citizens use AI chatbots to learn about elections and how popular models respond
Free and fair elections depend on citizens having access to accurate political information. Technological innovations, however, can shape the availability and accuracy of the information citizens rely on. AI chatbots are quickly becoming a popular source of political information in the US and across Europe, yet citizens may not understand where chatbots draw their information from, how accurate they are, or how their choice of chatbot shapes their information diet. The ParliView project investigates both the supply and demand sides of this problem, with significant implications for electoral integrity. In a set of recent Irish byelections and more recently, the lead up to the November 2026 US election, we evaluated whether popular chatbots are up to the task. Using a systematic coding framework we assess their source attribution, accuracy, and bias. We pair these results with bespoke surveys in multiple European countries to understand how and which citizens use AI chatbots to learn about politics. Together, this research provides a robust and holistic analysis of the nature of AI-mediated political information, as well as empirical evidence on citizens’ political use of AI in a real election context. In doing so, we advance our understanding of AI’s role in political (mis)information as well as its implications for electoral events.
November 27, 2026
The Best of Both Worlds - Improving Human Coding with Automatic Classification Models
The continuous refinement of automated classification models offers many possibilities to content analysis in political science. But an important question remains, particularly in the context of long-standing human coding projects: Can and should computers truly substitute human coders, and how comparable are the results? Thus, instead of aiming for substitution this talk proposes combining automated and human coding to leverage the strengths of both approaches while circumventing their weaknesses. Two methods of human-machine cooperation are examined: divided coding and assisted coding. Both approaches combine human coders with a shared XML-RoBERTa model. On the example of the long-established Manifesto Project the talk explores to what extent such approaches can secure overtime comparability. The approaches are evaluated in terms of time savings, ease of coding and reliability. The results are encouraging, indicating that such cooperation can lead to significant time savings without substantially compromising reliability.
Photo by David Ausserhofer
January 22, 2027
Beyond Transparency: Explainability, Alignment, and the Illusion of Control in High-Risk Information Processing
Large language models (LLMs) are increasingly deployed as intelligent decision support systems for information processing tasks such as truthfulness and credibility assessment of information for different downstream tasks. Especially for high-risk use cases compliance with the EU AI Act is relevant, because errors, biases, or manipulation can affect public discourse and democratic processes. This talk asks whether transparency delivers the control it promises, or whether it creates an illusion of it. I argue that explanations can give a false sense of understanding and oversight, and that transparency requirements fall short unless they are tied to measurable improvements in human decision quality. I support this with findings from our research group on human-AI collaboration. First, LLMs are not neutral judges: biases and framing effects, including political leaning, systematically shape how they assess the truthfulness of information. Second, the explanations these systems give for their judgments do not necessarily reflect how the judgments were made, and their quality can degrade when models are compressed for efficient deployment. Third, explanations may increase people's trust in a system without improving the decisions they make with it. I then turn to an open question that needs input from outside computer science. In ongoing work with partners including DFKI and Aleph Alpha, we aim to align language models with a European value system. But who defines such a system, and how would we know that alignment has actually been achieved? I discuss how interpretability methods can help verify alignment rather than merely assume it, and where the limits of such verification lie. The talk closes with implications for researchers, regulators, and practitioners, and invites discussion on how social science and humanities perspectives can help define what "meaningful human oversight" should mean.
January 29, 2027
Negotiating Societal Challenges in Digital Publics: AI-Supported Approaches for Social Science Research
Digital platforms have become central arenas of public life, where societal challenges (from climate change and migration to inequality and public health) are debated and shaped. To examine discourses, dynamics, and impacts across multilingual, multi-context digital publics, social scientists increasingly rely on AI-driven computational methods. This talk introduces methodological innovations in digital methods, focusing on computational multilingual text analysis and the problem of measurement validity. It presents a validation framework that accounts for context-specific biases in automated measurement and demonstrate how it improves the robustness and comparability of findings across languages and platforms. In the second part, it showcases empirical studies that deploy the some of the tools to analyze public discourse around environmental and inequality-related policies, highlighting what AI can, and cannot, capture about actors, topics, and diffusion dynamics. The talk close with implications for research design, and the responsible use of AI in social science research.