Thierry Poibeau

Historical language models are increasingly used to infer attitudes, beliefs, or viewpoints from past societies. This paper argues that such uses rest on fragile epistemic assumptions. We show that historical language models do not simulate past minds or populations, but instead model the structure of surviving textual archives, which are shaped by systematic biases of literacy, genre, and preservation. By introducing a validity ladder that distinguishes textual, discursive, and population-level claims, we provide a framework for evaluating what kinds of historical inferences these models can legitimately support. This perspective clarifies how historical language models can contribute to research in the social sciences and humanities without encouraging overinterpretation.

Debates about artificial consciousness ask whether AI systems possess subjective experience or intentionality. This paper argues that this question, while philosophically legitimate, displaces attention from a more consequential issue: what institutional order is already being built around systems treated as quasi-agents. We identify a structural contradiction at the heart of current AI discourse --- systems are simultaneously described in the vocabulary of autonomous agency and governed as permanently controllable artifacts --- and argue that its implications for alignment, accountability, and governance deserve more direct attention than the metaphysical question alone.

With the widespread use of large language models (LLMs), many researchers have turned their attention to detecting text generated by them. However, there is no consistent or precise definition of their target, namely “LLM-generated text”. Differences in usage scenarios and the diversity of LLMs further increase the difficulty of detection. What is commonly regarded as the detecting target usually represents only a subset of the text that LLMs can potentially produce. Human edits to LLM outputs, together with the subtle influences that LLMs exert on their users, are blurring the line between LLM-generated and human-written text. Existing benchmarks and evaluation approaches do not adequately address the various conditions in real-world detector applications. Hence, the numerical results of detectors are often misunderstood, and their significance is diminishing. Therefore, detectors remain useful under specific conditions, but their results should be interpreted only as references rather than decisive indicators.

What do large language models really know—and what does it mean to live alongside them? This book offers a critical and interdisciplinary exploration of large language models (LLMs), examining how they reshape our understanding of language, cognition, and society. Drawing on philosophy of language, linguistics, cognitive science, and AI ethics, it investigates how these models generate meaning, simulate reasoning, and perform tasks that once seemed uniquely human—from translation to moral judgment and literary creation. Rather than offering a purely technical account, the book interrogates the epistemic, ethical, and political dimensions of LLMs. It explores their limitations, their embedded biases, and their role in processes of automation, misinformation, and platform enclosure. At the same time, it reflects on how LLMs prompt us to revisit fundamental questions: What is understanding? What is creativity? How do we ascribe agency or trust in a world of synthetic language? Written for scholars, students, and curious readers across the humanities, social sciences, and computer science, this is both a philosophical inquiry and a practical guide to navigating the era of generative AI. It invites readers to think critically about the promises and perils of language technologies—and about the kind of future we are shaping with them.

This paper deals with problems related to discourse analysis within the framework of corpus linguistics, through a linguistic study dealing with procedurality in discourse. The fact that the study does not concern a specific lexical item makes it difficult to collect data without any predefined idea, in other words without introducing a bias in the study. The paper proposes a method to solve these problems, involving several annotators on the same texts and merging their proposals in order to get an objective unified annotation. We show that this step is a real part of the overall linguistic analysis.