Conference Agenda

This talk raises the question of whether the Chinese Room Argument (CRA) applies to large language models (LLMs). The thought experiment at the center of the CRA is tailored to Good Old-Fashioned Artificial Intelligence (GOFAI) systems. However, natural language processing has made significant progress, especially with the emergence of LLMs in recent years. LLMs differ from GOFAI systems in their design; they operate on vectors rather than symbols and do not follow a program but instead learn to map inputs to outputs. Consequently, some have suggested that the CRA is no longer relevant in discussions surrounding artificial language understanding. Contrary to these authors, I argue that if the CRA successfully demonstrates that implementing a symbolic computation is not sufficient for language understanding, then it also shows that implementing an LLM is not sufficient for language understanding. I begin by presenting a thought experiment called “the flashcard sorter”, which resembles the Chinese Room thought experiment but is tailored to LLMs rather than GOFAI systems. I then discuss the analogy between the flashcard sorter and modern LLMs, along with four objections to my claim that the CRA applies to LLMs. I conclude by presenting a modified version of the CRA.

Large Language Models (LLMs) invite a re-evaluation of the concept of intelligence by pointing to the unclear boundary between human and machine intelligence. The representational view of intelligence posits that intelligence requires: a) the existence of semantic representations, and b) the causal role of representations in driving behavior. This paper argues that the representational view is not well-founded. First, I challenge the idea of semantic representations as a requirement for intelligence as a consequence of our non-scientific and anthropocentric conception of intelligence. On a second step, I argue that even if such representations exist within AI systems, there are not sufficient reasons to claim they causally drive intelligent behavior. Finally, I propose some observable criteria for a less suspect concept of intelligence, drawing on scientific approaches from the cognitive sciences. These criteria suggest that LLMs, despite presumably lacking semantic representations, may still exhibit some degree of intelligence worthy of recognition. I conclude that requiring mental representations, and particularly their causal role, not only unjustifiably anthropomorphizes the concept of intelligence but also imposes unnecessarily high standards for intelligence attribution, including humans themselves. Therefore, a more operational and deflationary view of intelligence is preferable for understanding both human and machine intelligence.

Despite the recent success of large language models (LLMs), philosophers and AI researchers widely reject the claim that LLMs can understand, speak, or act, arguing that it is both false and excessively anthropomorphic. My presentation aims to challenge key aspects of this skepticism. To that end, I frame questions about the ability to understand, speak, and act as subsets of a broader question, the question of rule-following. For that reason, my paper examines the extent to which LLMs can be said to follow rules. I show that the predominant attempts to deny LLMs the ability to understand, speak, and act stem from misguided conceptions about rule-following, which I refer to as “mentalism” and “neurophysiologicalism”. In opposition to these misconceptions, I develop a new non-reductionist and deflationary approach to rule-following. On this basis, I argue that today’s LLMs are in interesting, but substantive respects capable of rule-following. I also demonstrate how my deflationary account of rule-following opens up novel ways to assess the capacity of LLMs to understand, speak, and act. Lastly, I reject the charge of anthropomorphism: while my approach exhibits a philosophically unproblematic, epistemically beneficial form of anthropocentrism, it manages to mitigate anthropomorphism.

The rise of artificial intelligence (AI) has come with a new kind of representations. We argue that these ‘artificial’ representations differ from the familiar notions of scientific and mental representations and it would therefore be a mistake to refer to them as either. Both mental and scientific representations fulfill familiar functions, albeit different ones. A key purpose of mental representations is to help cognitive systems to act and survive in complex environments. Scientific representations are diverse tools that are used to describe, explain, and predict phenomena in the natural world.

In this paper, we argue that AI representations do not currently fulfill either of these functions and that it is thus not clear what is meant when researchers refer to their AI models as “representing”. We focus on three key differences to mental and scientific representations: content, the status of misrepresentations, and a use condition. Based on these insights, we present criteria that AI systems should fulfil to be rightfully ascribed representations of their own kind. We conclude that AI systems are useful despite the unclear status of their internal states. But their usefulness does not justify to interpret these states as familiar representations, in the mental or scientific sense.

This presentation links the alignment problem in artificial intelligence (AI) to the alignment process in social interaction, exploring a procedural scheme based on normative inferentialism. Alignment as a problem in the development of technologies reveals potential inconsistencies between human goals and machine operations. The current widespread use of AI models such as Large Language Models (LLMs) underlines the difficulty of pursuing universal alignments due to their data-driven nature and the variability of human values. Through comparisons with alignment in human social interaction, it will argue that alignment is a dynamic process rooted in mutual adaptation of contexts. Drawing on Robert Brandom's normative inferentialism, which emphasizes reciprocal licensing and tracking of normative statuses in the language game of giving and asking for reasons, this presentation will propose a procedural approach to the alignment problem. LLMs constrained by data and design assumptions can be aligned through reinforcement learning with human feedback and dynamic user interactions. This alignment can be performed as an ongoing negotiation that enables AI models to be socially adaptive, context-aware, and capable of meeting diverse human needs.

Debates in animal cognition research are often polarized between the romantic view that some animals have human like cognitive capacities and the killjoy view that human cognitive capacities are unique. This lack of progress towards agreement is especially surprising since debates concerning cognitive capacities like causal understanding, theory of mind, or animal belief remain largely unaffected by new empirical evidence. The goal of this talk is to a) diagnose reasons for disagreement, and b) suggest a way to make progress. Concerning a) I will discuss three potential reasons for this situation: (1) researchers subscribe to different principles of interpretations; (2) the methods to investigate animal cognition necessarily result in a problem of underdetermination, and (3), what I take to be the central problem: disagreement is largely conceptual, i.e. researchers have different views on how to best characterize cognitive capacities. On the assumption that this analysis is correct I’ll b) argue that this does not boil down to merely verbal disagreement, and that the most promising approach to solve the conceptual question is to shift our attention from the question whether or not animals have a cognitive capacity to treating them as informative concerning how to best think about these capacities.