Elena Musi. Let’s Argue with GAI: An Investigation of LLMisinformation.
Henrike Beyer. Reasoning Two Ways: A Journey from Formal Puzzles to Natural Argumentation
Henning Wachsmuth. Toward Argumentative Large Language Models.
Bogdan Grecu. Robustness of Argument Mining Models
Nir Oren. On the Existence of an Inverse Solution for Preference-Based Reductions in Argumentation
Mark Snaith. Dialogue, Argumentation and Large Language Models: Complementary Perspectives.
Yohan Abhishek. A Unified Ontological Reasoning Layer for Heterogeneous Multi-Agents: A Systematic Review
Simon Wells. Visualising Dialogue Dynamics
Clara Seyfried. The Key to Comprehension: Exploring Categories of Argumentative Discourse Relations Across Disciplines
Ameer Saadat-Yazdi. Faithfulness vs. Coverage: What Do Summarization Metrics Actually Capture?
Keynote Details
Let’s Argue with GAI: An Investigation of LLMisinformation
This talk examines misinformation generated by Large Language Models (LLMs) from both theoretical and applied argumentation perspectives. While public and scholarly attention has largely focused on high-profile disinformation threats such as deepfakes, I argue that LLM-generated misinformation, defined as content that is misleading rather than entirely false, poses a distinct and significant risk due to three key factors.
First, the monopolisation of the information ecosystem: within a few years, LLMs have been integrated into search engines, content creation tools, educational platforms (e.g., NotebookLM), and policy-making processes (e.g., Redbox). Second, anthropomorphisation: LLMs can mimic skilled rhetoricians or politicians (Herbold et al., 2024), yet unlike human arguers --- whose reasoning is goal-directed and context-sensitive --- they operate as “stochastic parrots” (Bender, 2021), relying on probabilistic pattern matching. Third, opacity: LLMs often provide post hoc explanations that do not reflect their internal processes, making their non-deterministic outputs difficult for users to critically evaluate.
Building on the assumption that LLMs are fallacious by design (Musi et al., 2024), I argue that they currently function as unreasonable arguers across three levels: content repackaging, decision-making, and collective discussion. At the level of content repackaging, I present evidence from analyses of synthetic podcasting for science communication using NotebookLM (Musi, Yates and Harris, under review). At the decision-making level, I highlight gender bias in HR recruitment contexts (Sivakaminathan and Musi, 2025), observable both in topic selection and in the inferences underlying decisions. At the level of collective discussion, I report ongoing work comparing patterns of (dis)agreement among humans and LLMs.
Finally, I will discuss the concept of AI literacy, its current challenges (Musi, Carmi and Masotina, under review), and propose directions for designing LLMs that are argumentatively sound.
Toward Argumentative Large Language Models
Today's large language models (LLMs) are optimized toward giving helpful answers in response to prompts. In many situations, however, it may be preferable for an LLM to foster critical thinking rather than just following an instruction. While recent LLMs are said to 'reason', they barely build on established reasoning concepts known from argumentation theory. In this talk, I will give insights into recent efforts of my group in making LLMs more argumentative. Starting from basics of LLM training processes, I will present how to specialize LLMs for argumentation tasks via instruction fine-tuning as well as how to align the arguments they generate using reinforcement learning. From there, I will give an outlook on how to improve the actual reasoning capabilities of LLMs.
Dialogue, Argumentation and Large Language Models: Complementary Perspectives
Dialogue and argumentation research has long provided formal frameworks for modelling interaction, reasoning, and disagreement. Recent advances in AI, particularly Large Language Models, have created new opportunities to explore and apply these frameworks in practice. This highlights a natural point of connection: LLMs open up new ways to study, generate, and test dialogue and argumentation, while dialogue and argumentation provide tools for structuring, evaluating, and grounding LLM behaviour.
Drawing on a range of recent work, I will illustrate how these perspectives can be brought together in practice. This includes the EPSRC DiSCoAI project, which investigates how dialogue structures can be learned from real interaction; the PreFACE framework, which integrates structured knowledge and retrieval into conversational systems; and work on fallacy detection that uses LLMs to identify linguistic features for downstream classification.
Together, these strands point to a shared space in which generated language, structured dialogue, linguistic analysis, and grounded knowledge interact. The aim is not to present a unified solution, but to highlight how these perspectives can inform one another, in both research and the design of future systems.