We consider the questions of whether or not large language models (LLMs) have beliefs, and, if they do, how we might measure them. First, we consider whether or not we should expect LLMs to have something like beliefs in the first place. We consider some recent arguments aiming to show that LLMs cannot have beliefs. We show that these arguments are misguided. We provide a more productive framing of questions surrounding the status of beliefs in LLMs, and highlight the empirical (...) nature of the problem. With this lesson in hand, we evaluate two existing approaches for measuring the beliefs of LLMs, one due to Azaria and Mitchell (The internal state of an llm knows when its lying, 2023) and the other to Burns et al. (Discovering latent knowledge in language models without supervision, 2022). Moving from the armchair to the desk chair, we provide empirical results that show that these methods fail to generalize in very basic ways. We then argue that, even if LLMs have beliefs, these methods are unlikely to be successful for conceptual reasons. Thus, there is still no lie-detector for LLMs. We conclude by suggesting some concrete paths for future work. (shrink)

As we write this research paper, we notice an explosion in popularity of machine learning in numerous fields (ranging from governance, education, and management to criminal justice, fraud detection, and internet of things). In this contribution, rather than focusing on any of those fields, which have been well-reviewed already, we decided to concentrate on a series of more recent applications of deep learning models and technologies that have only recently gained significant track in the relevant literature. These applications are concerned (...) with artistic production (Sect. 2.1), the writing process (Sect. 2.2), music production (Sect. 2.3), text recognition and attribution (Sect. 2.4). After reviewing and analyzing the positive contributions as well as some of the major limitations of these technologies in each of those fields, we critically reflect (Sect. 3) on how their widespread implementation may affect humans and their creativity. In Sect. 4, we notice that deep learning models are here to stay; so, rather than embracing a negative or pessimistic stance with respect to their future applications in creative domains, we suggest a balanced approach for their assessment and beneficial usage. Finally (Sect. 5), we conclude by summarising what we have achieved and by pointing out possible future research directions. (shrink)

Word co‐occurrence patterns in language corpora contain a surprising amount of conceptual knowledge. Large language models (LLMs), trained to predict words in context, leverage these patterns to achieve impressive performance on diverse semantic tasks requiring world knowledge. An important but understudied question about LLMs’ semantic abilities is whether they acquire generalized knowledge of common events. Here, we test whether five pretrained LLMs (from 2018's BERT to 2023's MPT) assign a higher likelihood to plausible descriptions of agent−patient interactions than to minimally (...) different implausible versions of the same event. Using three curated sets of minimal sentence pairs (total n = 1215), we found that pretrained LLMs possess substantial event knowledge, outperforming other distributional language models. In particular, they almost always assign a higher likelihood to possible versus impossible events (The teacher bought the laptop vs. The laptop bought the teacher). However, LLMs show less consistent preferences for likely versus unlikely events (The nanny tutored the boy vs. The boy tutored the nanny). In follow‐up analyses, we show that (i) LLM scores are driven by both plausibility and surface‐level sentence features, (ii) LLM scores generalize well across syntactic variants (active vs. passive constructions) but less well across semantic variants (synonymous sentences), (iii) some LLM errors mirror human judgment ambiguity, and (iv) sentence plausibility serves as an organizing dimension in internal LLM representations. Overall, our results show that important aspects of event knowledge naturally emerge from distributional linguistic patterns, but also highlight a gap between representations of possible/impossible and likely/unlikely events. (shrink)

The transformers that drive chatbots and other AI systems constitute large language models (LLMs). These are currently the focus of a lively discussion in both the scientific literature and the popular media. This discussion ranges from hyperbolic claims that attribute general intelligence and sentience to LLMs, to the skeptical view that these devices are no more than “stochastic parrots”. I present an overview of some of the weak arguments that have been presented against LLMs, and I consider several of the (...) more compelling criticisms of these devices. The former significantly underestimate the capacity of transformers to achieve subtle inductive inferences required for high levels of performance on complex, cognitively significant tasks. In some instances, these arguments misconstrue the nature of deep learning. The latter criticisms identify significant limitations in the way in which transformers learn and represent patterns in data. They also point out important differences between the procedures through which deep neural networks and humans acquire knowledge of natural language. It is necessary to look carefully at both sets of arguments in order to achieve a balanced assessment of the potential and the limitations of LLMs. (shrink)

Event segmentation theory posits that people segment continuous experience into discrete events and that event boundaries occur when there are large transient increases in prediction error. Here, we set out to test this theory in the context of story listening, by using a deep learning language model (GPT‐2) to compute the predicted probability distribution of the next word, at each point in the story. For three stories, we used the probability distributions generated by GPT‐2 to compute the time series of (...) prediction error. We also asked participants to listen to these stories while marking event boundaries. We used regression models to relate the GPT‐2 measures to the human segmentation data. We found that event boundaries are associated with transient increases in Bayesian surprise but not with a simpler measure of prediction error (surprisal) that tracks, for each word in the story, how strongly that word was predicted at the previous time point. These results support the hypothesis that prediction error serves as a control mechanism governing event segmentation and point to important differences between operational definitions of prediction error. (shrink)

Many researchers and intellectuals warn about extreme risks from artificial intelligence. However, these warnings typically came without systematic arguments in support. This paper provides an argument that AI will lead to the permanent disempowerment of humanity, e.g. human extinction, by 2100. It rests on four substantive premises which it motivates and defends: first, the speed of advances in AI capability, as well as the capability level current systems have already reached, suggest that it is practically possible to build AI systems (...) capable of disempowering humanity by 2100. Second, due to incentives and coordination problems, if it is possible to build such AI, it will be built. Third, since it appears to be a hard technical problem to build AI which is aligned with the goals of its designers, and many actors might build powerful AI, misaligned powerful AI will be built. Fourth, because disempowering humanity is useful for a large range of misaligned goals, such AI will try to disempower humanity. If AI is capable of disempowering humanity and tries to disempower humanity by 2100, then humanity will be disempowered by 2100. This conclusion has immense moral and prudential significance. (shrink)

Does the capacity to think require the capacity to sense? A lively debate on this topic runs throughout the history of philosophy and now animates discussions of artificial intelligence. Many have argued that AI systems such as large language models cannot think and understand if they lack sensory grounding. I argue that thought does not require sensory grounding: there can be pure thinkers who can think without any sensory capacities. As a result, the absence of sensory grounding does not entail (...) that large language models cannot think or understand. I also consider to what extent quasi-sensory grounding can at least boost the performance of a language model. (shrink)

The present paper looks at one of the most thorough articles on the intelligence of GPT, research conducted by engineers at Microsoft. Although there is a great deal of value in their work, I will argue that, for familiar philosophical reasons, their methodology, ‘Black-box Interpretability’ is wrongheaded. But there is a better way. There is an exciting and emerging discipline of ‘Inner Interpretability’ (also sometimes called ‘White-box Interpretability’) that aims to uncover the internal activations and weights of models in order (...) to understand what they represent and the algorithms they implement. In my view, a crucial mistake in Black-box Interpretability is the failure to appreciate that how processes are carried out matters when it comes to intelligence and understanding. I can’t pretend to have a full story that provides both necessary and sufficient conditions for being intelligent, but I do think that Inner Interpretability dovetails nicely with plausible philosophical views of what intelligence requires. So the conclusion is modest, but the important point in my view is seeing how to get the research on the right track. Towards the end of the paper, I will show how some of the philosophical concepts can be used to further refine how Inner Interpretability is approached, so the paper helps draw out a profitable, future two-way exchange between Philosophers and Computer Scientists. (shrink)

Recent artificial intelligence advances, especially those of large language models (LLMs), have increasingly shown glimpses of human-like intelligence. This has led to bold claims that these systems are no longer a mere “it” but now a “who,” a kind of person deserving respect. In this paper, I argue that this view depends on a Cartesian account of personhood, on which identifying someone as a person is based on their cognitive sophistication and ability to address common-sense reasoning problems. I contrast this (...) with a different account of personhood, one where an agent is a person if they are autonomous, responsive to norms, and culpable for their actions. On this latter account, I show that LLMs are not person-like, as evidenced by their propensity for dishonesty, inconsistency, and offensiveness. Moreover, I argue current LLMs, given the way they are designed and trained, cannot be persons—either social or Cartesian. The upshot is that contemporary LLMs are not, and never will be, persons. (shrink)

Most, if not all, philosophers agree that computers cannot learn what words refers to from raw text alone. While many attacked Searle’s Chinese Room thought experiment, no one seemed to question this most basic assumption. For how can computers learn something that is not in the data? Emily Bender and Alexander Koller ( 2020 ) recently presented a related thought experiment—the so-called Octopus thought experiment, which replaces the rule-based interlocutor of Searle’s thought experiment with a neural language model. The Octopus (...) thought experiment was awarded a best paper prize and was widely debated in the AI community. Again, however, even its fiercest opponents accepted the premise that what a word refers to cannot be induced in the absence of direct supervision. I will argue that what a word refers to _is_ probably learnable from raw text alone. Here’s why: higher-order concept co-occurrence statistics are stable across languages and across modalities, because language use (universally) reflects the world we live in (which is relatively stable). Such statistics are sufficient to establish what words refer to. My conjecture is supported by a literature survey, a thought experiment, and an actual experiment. (shrink)

Landgrebe and Smith :2061–2081, 2021) present an unflattering diagnosis of recent advances in what they call language-centric artificial intelligence—perhaps more widely known as natural language processing: The models that are currently employed do not have sufficient expressivity, will not generalize, and are fundamentally unable to induce linguistic semantics, they say. The diagnosis is mainly derived from an analysis of the widely used Transformer architecture. Here I address a number of misunderstandings in their analysis, and present what I take to be (...) a more adequate analysis of the ability of Transformer models to learn natural language semantics. To avoid confusion, I distinguish between inferential and referential semantics. Landgrebe and Smith ’s analysis of the Transformer architecture’s expressivity and generalization concerns inferential semantics. This part of their diagnosis is shown to rely on misunderstandings of technical properties of Transformers. Landgrebe and Smith also claim that referential semantics is unobtainable for Transformer models. In response, I present a non-technical discussion of techniques for grounding Transformer models, giving them referential semantics, even in the absence of supervision. I also present a simple thought experiment to highlight the mechanisms that would lead to referential semantics, and discuss in what sense models that are grounded in this way, can be said to understand language. Finally, I discuss the approach Landgrebe and Smith advocate for, namely manual specification of formal grammars that associate linguistic expressions with logical form. (shrink)

Human presencing explores how people’s past encounters with others shape their present actions. In this paper, I present an alternative perspective on human embodiment in which the re-evoking of the absent can be traced to the intricate interplay of bodily dynamics. By situating the phenomenon within distributed, embodied, and dialogic approaches to language and cognition, I am overcoming the theoretical and methodological challenges involved in perceiving and acting upon what is not perceptually present. In a case study, I present strong (...) and weak dimensions of human presencing. In the former, a person uses their body in distinct ways and shapes their immediate ecology to make others present to them. In contrast, in the latter, a person’s past encounter with others powerfully shapes the projections they make onto written digital inscriptions. These findings have implications for how people act in online learning environments and how human activity shapes the machines we use every day. In this way, the paper highlights the complexity of a person as a social being and allows for different approaches to human embodiment in technology. (shrink)

Many of our spoken languages are endangered and rapidly becoming extinct. Due to this, there are attempts to preserve as many of those languages as possible. One preservation approach is combining data collection and artificial intelligence‐based language models. However, current data collection methods may only capture static data from a dynamic cognitive process. If data are not genuinely capturing the dynamic process, it raises questions about whether they capture all the essential knowledge about how a language functions. Here, we discuss (...) the implications of this issue and its importance in preserving endangered languages. (shrink)

Large language models and other generative artificial intelligence systems are achieving increasingly impressive results, though the quality of those results still seems dull and uninspired. This paper argues that this poor quality can be linked to the philosophical notion of inauthenticity as presented by Kierkegaard, Nietzsche, and Heidegger, and that this inauthenticity is fundamentally grounded in the design and structure of such systems by virtue of the way they statistically level down the materials on which they are trained. Although it (...) seems possible to create the conditions for authenticity in these systems, the resulting authenticity would be grounded in machine intelligence, not human intelligence. The argument extends the criticisms of artificial intelligence articulated by Hubert Dreyfus, updated to account for recent developments in machine learning and artificial neural networks. While more optimistic concerning the prospects for successfully creating artificial intelligence than Dreyfus had been, this paper argues that the resulting intelligence may not align well with human intelligence and may not be desirable for humans, if it is fully authentic. (shrink)

Could translation be fully automated? We must first acknowledge the complexity, ambiguity, and diversity of natural languages. These aspects of natural languages, when combined with a particular dilemma known as the computational dilemma, appear to imply that the machine translator faces certain obstacles that a human translator has already managed to overcome. At the same time, science has not yet solved the problem of how human brains process natural languages and how human beings come to acquire natural language understanding. We (...) will then distinguish between the task of translation and the responsibility of the translator. Thereafter, we will conduct a survey of the methods of machine translation (viz. RBMT, SMT, NMT, foundation models or large language models). These methods will then be critically evaluated both in general and relative to Bar-Hillel’s hypothesis about the impossibility of fully automatic, high-quality machine translation (F ahqmt ). Some concluding remarks will be made about the scope, prospects, and limits of machine translation. (shrink)

In this article, I develop a loosely Wittgensteinian conception of what it takes for a being, including an AI system, to understand language, and I suggest that current state of the art systems are closer to fulfilling these requirements than one might think. Developing and defending this claim has both empirical and conceptual aspects. The conceptual aspects concern the criteria that are reasonably applied when judging whether some being understands language; the empirical aspects concern the question whether a given being (...) fulfills these criteria. On the conceptual side, the article builds on Glock’s concept of intelligence, Taylor’s conception of intrinsic rightness as well as Wittgenstein’s rule-following considerations. On the empirical side, it is argued that current transformer-based NNLP models, such as BERT and GPT-3 come close to fulfilling these criteria. (shrink)

Block’s, 5–43 1981) anti-behaviourist attack of the Turing Test not only illustrates that the test is a non-sufficient criterion for attributing thought; I suggest that it also exemplifies the limiting case of the more general concern that a machine which has access to enormous amounts of data can pass the Turing Test by simple symbol-manipulation techniques. If the answers to a human interrogator are entailed by the machines’ data, the Turing Test offers no clear criterion to distinguish between a thinking (...) machine and a machine that merely manipulates representations of words and sentences as it is found in contemporary Natural Language Processing models. This paper argues that properties about vagueness are accessible to any human-like thinker but do not normally display themselves in ordinary language use. Therefore, a machine that merely performs simple symbol manipulation from large amounts of previously acquired data – where this body of data does not contain facts about vagueness – will not be able to report on these properties. Conversely, a machine that has the capacity to think would be able to report on these properties. I argue that we can exploit this fact to establish a sufficient criterion of thought. The criterion is a specification of some of the questions that, as I explain, should be asked by the interrogator in a Turing Test situation. (shrink)