The Age of the Bookworm
Can Large Language Models Represent the World?

About the Workshop

Stochastic Parrots or World-Modellers?

What does it mean for a system trained exclusively on text to genuinely represent and understand the world?

Since the public release of ChatGPT in late 2022, large language models (LLMs) have become a fixture of academic and public life. Yet how these systems operate, and what they are fundamentally capable of, remains poorly understood by many of its users. While many are aware that such systems are trained to predict the most likely next word (or rather token) given a context, far fewer have reflected on capacities emerging from this simple training objective, and the further implications of fine-tuning via reinforcement learning through human feedback and verifiable rewards.

An interesting starting point for the debate is a thought experiment proposed by Bender and Koller (2020) involving an infinitely smart octopus intercepting underwater telegraphic communications between two islanders. Can the octopus genuinely output meaningful text about the land-world after eavesdropping on vast amounts of the islanders’ communication about what happens there, without ever engaging with that world directly? This analogy is meant to raise the following sharper question: can an LLM’s internal states and outputs be about extra-linguistic reality, independently of the meaning human interpreters project onto them?

This workshop gathers philosophers and AI researchers to address this question and other related issues.

Speakers

Participants

Programme

Preliminary Schedule

11–12 November 2026. This is a joint Geneva and Neuchâtel event, held in the Chico Mendes room at the Maison Internationale des Associations in Geneva.

Talks are scheduled for 45+15 minutes.

Practical Information

Attending the Workshop

In the Run-Up to the Workshop

Reading Group on the Philosophy of AI

The group is organised in the run-up to this workshop. The aim is to prepare together for the questions that will be debated here, by building a shared foundation of readings and discussion beforehand. Sessions will be held on Zoom, so that everyone can take part easily.

The idea is to bring together those who are interested in what these systems really do: can they represent the world, or only the regularities of our use of language? What does it mean to “understand” or to “refer” for a system trained on text alone? These questions, at the crossroads of the philosophy of mind, of language, and of AI, are at the heart of the field, and the reading group will be an occasion to work through them together on the basis of recent texts.

Syllabus

Part I — Foundations

The first two sessions will lay the groundwork needed to approach the literature with confidence.

Part II — Readings

From the third session on, the group will run as a reading group proper. At each meeting we will discuss the text of the week together. Please read the texts before each session. Where two texts are paired, read them in the order listed, since the second responds to the first.

1. 7 October

   Stochastic parrots?

   1. Bender & Koller (2020), “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data” →
   2. Søgaard (2023), “Grounding the Vector Space of an Octopus: Word Meaning from Raw Text” →

2. 14 October

   Is there a grounding problem for LLMs?

   1. Mollo & Millière (2023), “The Vector Grounding Problem” →
   2. Gubelmann (2024), “Pragmatic Norms Are All You Need – Why the Symbol Grounding Problem Does Not Apply to LLMs” →

3. 21 October

   A deep dive on the vector grounding problem

   1. Williams (2026), “Can Structural Correspondences Ground Real World Representational Content in Large Language Models?” →
   2. Du Sordet (2026), “Only Agentic LLMs Can Represent the World” →

4. 28 October

   Intentions in LLMs?

   • Williams (2026), “Intention-like Representations in Language Models?” →

5. 4 November

   Understanding in LLMs?

   • Beckmann & Queloz (2026), “Mechanistic Indicators of Understanding in Large Language Models” →

References

All texts are freely available at the links below.

• Beckmann, P., & Queloz, M. (2025). Mechanistic indicators of understanding in large language models. Preprint. PhilArchive
• Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185–5198). ACL. ACL Anthology
• Du Sordet, E. Only agentic LLMs can represent the world. PhilPapers
• Gubelmann, R. (2024). Pragmatic norms are all you need – Why the symbol grounding problem does not apply to LLMs. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 11663–11678). ACL. ACL Anthology
• Mollo, D. C., & Millière, R. The vector grounding problem. Forthcoming in Philosophy and the Mind Sciences. arXiv:2304.01481
• Søgaard, A. (2023). Grounding the vector space of an octopus: Word meaning from raw text. Minds and Machines, 33(1), 33–54. Springer
• Williams, I. Can structural correspondences ground real world representational content in large language models? Mind & Language (2026). arXiv:2506.16370
• Williams, I. (2025). Intention-like representations in language models? Preprint. PhilArchive

Organisation

Organizers & Funding