COMPUTATIONAL LINGUISTICS
Max van Duijn, Leiden University ↗

Why (Some) Large Language Models Are Such Good Mindreaders

Well, are they? If by “mindreading” we mean performance on standardised Theory of Mind tests, then the answer is “yes” for several state-of-the-art LLMs. If a more encompassing definition of mindreading is adopted, involving also pragmatic skills, empathy, recursive and metarepresentational capacities, and more, and if a broader suite of LLMs is considered, then the answer should be more nuanced. In this talk I will discuss the empirical evidence for different aspects of mindreading in LLMs as presented in various key studies, including some from my own SIM lab. I will connect these to findings from the cognitive-developmental and evolutionary sciences, to deepen our theoretical understanding of social cognition in both humans and machines, as a basis for complementary interactions between them in the future.

THEATER STUDIES
Ulf Otto, LMU Munich ↗

Machine Phenomenology: Theatre Photography, Computer Vision, and the Epistemic Potentials of Statistical Models

Events lie at the heart of theatre. Media are therefore of interest to research not only as sources, but also as mediators. Especially in the mass media, theatre in its breadth has been recorded since the nineteenth century. Their digital availability makes it possible to investigate theatre cultures on a new scale —analogous to distant reading — that is, statistically rather than through exemplary case studies. What was shown, and how was it seen? Photographic reporting in particular, which became established after 1945, contains a contemporary history of theatre as an image history. Yet what appears in images cannot be counted in the way words can — at least not without a gaze that already separates and names what can be found in the image. Machines must learn, through the abstraction of large quantities of data, to “see something into” images. The machine’s gaze is therefore opaque (black-boxed) and biased — like the human gaze as well. Just as performance analysis always implicates one’s own perception and its social positioning, the machine-based preprocessing of theatre photographs also becomes a kind of Rorschach test for AI. Building on a research project on the visual history of theatre photography, the talk addresses the epistemic consequences of computational methods and machine learning, and argues for comprehensive data literacy in the humanities.

MEDIA STUDIES
Minha Lee, TU/e Eindhoven ↗

Mind over matter? Conversational futures with speculative minds 

We denote other beings to have minds of their own when we perceive them to have cognitive and affective capacities, to different degrees. A chatbot can be perceived to have some level of cognition, but not emotions, for instance. But, this bias can be overturned through how the agent and the environment it operates in are designed, which can be speculatively explored. The potential to shape our human perception of non-human “minds” has various ethical considerations to be discussed. The talk will be followed by a collaborative activity.

PSYCHOLOGY
Ute Schmid, University of Bamberg ↗

Requirements for Human AI Alignment in Joint Decision Making and Problem Solving

With the advance of highly performant AI systems – deep learning based classifiers and transformer based generative approaches – there is hope that human AI collaboration will support humans to master complex tasks more efficient as well as in high quality. This is especially relevant for critical tasks such as medical diagnostics or generation of program code for scientific tasks. However, a growing number of empirical studies shows that the best of humans or AI systems outperform human AI teams. Possible reasons are on the one hand over-reliance in the output of AI systems and on the other hand a mismatch between human cognitive processes and AI systems. In the talk, I will propose that we need more human aligned methods of explainable AI as well as novel methods to support human agency and oversight.  

TRANSDISCIPLINARY
Eva Wolfangel, Science Journalist  ↗

The World According to Words

If all you had were words, could you truly grasp the world behind them? Large language models try exactly that. In some situations they seem highly capable; under systematic testing, their limits become obvious. Whatever this is, it does not fully match what we usually mean by understanding. But the human case is not straightforward either. What actually counts as understanding, and how do we obtain it? Decades of research in different disciplines offer partial answers, yet many assumptions are still unsettled. Today, some findings suggest that language models merely imitate world knowledge; others hint at parallels to how humans rely on heuristics and incomplete representations. This talk explores, what LLMs reveal about the boundary between imitation and understanding, where language-only learning seems to reach its limits and which alternative approaches researchers are exploring to give machines a richer access to the world. Rather than delivering clear-cut answers (because: they are not here yet!), the talk brings together surprising results that challenge common intuitions about both machines and minds. An invitation to rethink what it means to “know” the world at all.