Large Language Models

The emergence of Large Language Models (LLMs) and advancements in Artificial Intelligence (AI) offer an opportunity for computational social science research at scale. Building upon prior explorations of LLM agent design, our work introduces a simulated agent society where complex social relationships dynamically form and evolve over time. Agents are imbued with psychological drives and placed in a sandbox survival environment. We conduct an evaluation of the agent society through the lens of Thomas Hobbes's seminal Social Contract Theory (SCT). We (...) analyze whether, as the theory postulates, agents seek to escape a brutish "state of nature" by surrendering rights to an absolute sovereign in exchange for order and security. Our experiments unveil an alignment: Initially, agents engage in unrestrained conflict, mirroring Hobbes's depiction of the state of nature. However, as the simulation progresses, social contracts emerge, leading to the authorization of an absolute sovereign and the establishment of a peaceful commonwealth founded on mutual cooperation. This congruence between our LLM agent society's evolutionary trajectory and Hobbes's theoretical account indicates LLMs' capability to model intricate social dynamics and potentially replicate forces that shape human societies. By enabling such insights into group behavior and emergent societal phenomena, LLM-driven multi-agent simulations, while unable to simulate all the nuances of human behavior, may hold potential for advancing our understanding of social structures, group dynamics, and complex human systems. (shrink)

Recent debates over artificial intelligence have focused on its perceived lack of interpretability and explainability. I argue that these notions fail to capture an important aspect of what end-users—as opposed to developers—need from these models: what is needed is systematicity, in a more demanding sense than the compositionality-related sense that has dominated discussions of systematicity in the philosophy of language and cognitive science over the last thirty years. To recover this more demanding notion of systematicity, I distinguish between (i) the (...) systematicity of thinkable contents, (ii) the systematicity of thinking, and (iii) the ideal of systematic thought. I then deploy this distinction to critically evaluate Fodor’s systematicity-based argument for the language of thought hypothesis before recovering the notion of the systematicity of thought as a regulative ideal, which has historically shaped our understanding of what it means for thought to be rational, authoritative, and scientific. To assess how much systematicity we need from AI models, I then argue, we must look to the functions of systematizing thought. To this end, I identify five functions served by systematization, and show how these can be used to arrive at a dynamic understanding of the need to systematize thought that can tell us what kind of systematicity is called for and when. (shrink)

A key assumption fuelling optimism about the progress of Large Language Models (LLMs) in modelling the world is that the truth is systematic: true statements about the world form a whole that is not just consistent, in that it contains no contradictions, but cohesive, in that the truths are inferentially interlinked. This holds out the prospect that LLMs might rely on that systematicity to fill in gaps and correct inaccuracies in the training data: consistency and cohesiveness promise to facilitate progress (...) towards comprehensiveness in an LLM’s representation of the world. However, philosophers have identified reasons to doubt that the truth is systematic across all domains of thought, arguing that in normative domains, in particular, the truth is not necessarily systematic. I argue that insofar as the truth in normative domains is asystematic, this renders it correspondingly harder for LLMs to make progress, because they cannot rely on the consistency and cohesiveness of the truth to work towards comprehensiveness. And the less LLMs can rely on the systematicity of truth, the less we can rely on them to do our practical deliberation for us, as there is correspondingly more of a role for human agency in navigating asystematic normative domains. (shrink)

Sideloading is the creation of a digital model of a person during their life via iterative improvements of this model based on the person's feedback. The progress of LLMs with large prompts allows the creation of very large, book-size prompts which describe a personality. We will call mind-models created via sideloading "sideloads"; they often look like chatbots, but they are more than that as they have other output channels, like internal thought streams and descriptions of actions. -/- By arranging the (...) predictive power of the facts about a person, we can write down the most important facts first. This allows us to get a high level of fidelity in a person's model by writing down a finite number of facts and running it as an LLM prompt with instructions to create a person's chatbot. This instruction also includes an extensive general sideload prompt (we will call it a prompt-loader), which explains how the person's sideload should work, but it is universal and works as an operating system for different persons. -/- I created my personal sideload to test all this and it started working surprisingly quickly. We found that there are three main levels of facts about a person: core facts, long-term memory, and historical facts. Core facts should go to the main prompt of the LLM, long-term memory belongs to RAG, and historical facts – which a person often does not remember – should be only used to extract new relevant data. The most important thing in preparing data for sideloading is to get a full list of the core facts, and not to mix it with long-term memory and historical facts. The LLM-sideload project is inherently future-oriented. New state-of-the-art LLMs with large prompt capabilities emerge quarterly. This allows for the preparation of prompts that exceed current LLM capabilities, anticipating smooth execution in the following year. Below we present the key findings of this work in the form of an executive summary: Quality of sideloads is evaluated using three metrics: • Facts: Accuracy of the answers about the person's life • Vibe: How well the sideload captures the person's style and personality, as judged by the person themselves and their acquaintances • Brilliant insights: Ability to contribute unique, valuable ideas beyond the training data, while maintaining the author's style • Coarseness: Levels of details as percent of the total of recordable textual information from memory estimated to be 10000 pages. Similar to the jpeg level of compression. Estimated performance of Alexey Turchin's sideload: • Facts: 70% correct • Vibe: 20% accurate • Brilliant insights: near zero • Coarseness: 10% The project achieved these results in about one month of work. Vibe is harder to evaluate and regulate than facts. This approach focuses on informational personal identity, setting aside issues of observer identity and internal qualia. (shrink)

This paper seeks to answer the question: Can contemporary forms of artificial intelligence be creative? To answer this question, I consider three conditions that are commonly taken to be necessary for creativity. These are novelty, value, and agency. I argue that while contemporary AI models may have a claim to novelty and value, they cannot satisfy the kind of agency condition required for creativity. From this discussion, a new condition for creativity emerges. Creativity requires curiosity, a motivation to pursue epistemic (...) goods. I argue that contemporary AI models do not satisfy this new condition. Because they lack both agency and curiosity, it is a mistake to attribute the same sort of creativity to AI that we prize in humans. Finally, I consider the question of whether these AI models stand to make human creativity in the arts and sciences obsolete, despite not being creative themselves. I argue, optimistically, that this is unlikely. (shrink)

The problem considered in this chapter emerges from the tension we find when looking at the design and architecture of chatbots on the one hand and their conversational aptitude on the other. In the way that LLM chatbots are designed and built, we have good reason to suppose they don't possess second-order capacities such as intention, belief or knowledge. Yet theories of conversation make great use of second-order capacities of speakers and their audiences to explain how aspects of interaction succeed. (...) As we can all bear witness to now though, at the point of use chatbots appear capable of performing language tasks at a level close to that of humans. This creates a tension when we consider something like, for example, the classic Gricean theory of implicature. On a broad summary of this type of account, to utter p and implicate q requires the reflexive occurrence of an audience supposing a speaker believes that q, and the speaker believing that their audience can determine they believe it when they utter p. So taken at face value, if a chatbot doesn’t have the capacity for belief, then either in their role as speaker or audience, they would not seem capable of either generating or comprehending implicatures. Yet on the surface it does seem that chatbots are capable of dealing with (some) implicatures, and as such it raises questions about how we should then correlate this with what we think occurs in cases of implicature with chatbots. (shrink)

Human-centered explainable AI (HCXAI) advocates for the integration of social aspects into AI explanations. Central to the HCXAI discourse is the Social Transparency (ST) framework, which aims to make the socio-organizational context of AI systems accessible to their users. In this work, we suggest extending the ST framework to address the risks of social misattributions in Large Language Models (LLMs), particularly in sensitive areas like mental health. In fact LLMs, which are remarkably capable of simulating roles and personas, may lead (...) to mismatches between designers' intentions and users' perceptions of social attributes, risking to promote emotional manipulation and dangerous behaviors, cases of epistemic injustice, and unwarranted trust. To address these issues, we propose enhancing the ST framework with a fifth 'W-question' to clarify the specific social attributions assigned to LLMs by its designers and users. This addition aims to bridge the gap between LLM capabilities and user perceptions, promoting the ethically responsible development and use of LLM-based technology. (shrink)

Under what conditions would an artificially intelligent system have wellbeing? Despite its clear bearing on the ethics of human interactions with artificial systems, this question has received little direct attention. Because all major theories of wellbeing hold that an individual’s welfare level is partially determined by their mental life, we begin by considering whether artificial systems have mental states. We show that a wide range of theories of mental states, when combined with leading theories of wellbeing, predict that certain existing (...) artificial systems have wellbeing. Along the way, we argue that there are good reasons to believe that artificial systems can have wellbeing even if they are not phenomenally conscious. While we do not claim to demonstrate conclusively that AI systems have wellbeing, we argue that there is a significant probability that some AI systems have or will soon have wellbeing, and that this should lead us to reassess our relationship with the intelligent systems we create. (shrink)

In this paper, I examine why moral exemplars are useful and even necessary in moral education despite several critiques from researchers and educators. To support my point, I review recent AI research demonstrating that exemplar-based learning is superior to rule-based learning in model performance in training neural networks, such as large language models. I particularly focus on why education aiming at promoting the development of multifaceted moral functioning can be done effectively by using exemplars, which is similar to exemplar-based learning (...) in AI model training. Furthermore, I discuss the potential limitations and issues related to exemplar-applied moral education with findings from recent studies in AI research raising concerns about model biases and toxic outcomes. I attempt to propose ways to address the concerns regarding employing moral exemplars as well. As remedies, I suggest that autonomy-supporting deliberative and reflective learning processes should be utilized. Furthermore, based on the discussion, I examine how macroscopic socio-cultural aspects influence the effectiveness of exemplar-applied moral education. Suggestions for moral educators and future directions for research in moral education are briefly discussed. (shrink)

What can contemporary machine learning (ML) models do? Given the proliferation of ML models in society, answering this question matters to a variety of stakeholders, both public and private. The evaluation of models' capabilities is rapidly emerging as a key subfield of modern ML, buoyed by regulatory attention and government grants. Despite this, the notion of an ML model possessing a capability has not been interrogated: what are we saying when we say that a model is able to do something? (...) And what sorts of evidence bear upon this question? -/- In this paper, we aim to answer these questions, using the capabilities of large language models (LLMs) as a running example. Drawing on the large philosophical literature on abilities, we develop an account of ML models' capabilities which can be usefully applied to the nascent science of model evaluation. Our core proposal is a conditional analysis of model abilities (CAMA): crudely, a machine learning model has a capability to X just when it would reliably succeed at doing X if it 'tried'. The main contribution of the paper is making this proposal precise in the context of ML, resulting in an operationalisation of CAMA applicable to LLMs. We then put CAMA to work, showing that it can help make sense of various features of ML model evaluation practice, as well as suggest procedures for performing fair inter-model comparisons. (shrink)

One of the central questions in the epistemology of conversational AIs is how to classify the beliefs acquired from them. Two promising candidates are instrument-based and testimony-based beliefs. However, the category of instrument-based beliefs faces an intrinsic problem, and a challenge arises in its application. On the other hand, relying solely on the category of testimony-based beliefs does not encompass the totality of our practice of using conversational AIs. To address these limitations, I propose a novel classification of beliefs that (...) shifts the focus from the properties of the conversational AIs themselves to how recipients perceive AI output, specifically whether they take the output at face value. The proposed categories are beliefs by regarding as instruments and beliefs by regarding as testifiers. By using these complementary categories, a more comprehensive understanding of beliefs acquired from conversational AIs can be achieved while avoiding the problems associated with the traditional classification. (shrink)

Recent breakthroughs in artificial intelligence have primarily resulted from training deep neural networks (DNNs) with vast numbers of adjustable parameters on enormous datasets. Due to their complex internal structure, DNNs are frequently characterized as inscrutable ``black boxes,'' making it challenging to interpret the mechanisms underlying their impressive performance. This opacity creates difficulties for explanation, safety assurance, trustworthiness, and comparisons to human cognition, leading to divergent perspectives on these systems. This chapter examines recent developments in interpretability methods for DNNs, with a (...) focus on interventionist approaches inspired by causal explanation in philosophy of science. We argue that these methods offer a promising avenue for understanding how DNNs process information compared to merely behavioral benchmarking and correlational probing. We review key interventionist methods and illustrate their application through practical case studies. These methods allow researchers to identify and manipulate specific computational components within DNNs, providing insights into their causal structure and internal representations. We situate these approaches within the broader framework of causal abstraction, which aims to align low-level neural computations with high-level interpretable models. While acknowledging current limitations, we contend that interventionist methods offer a path towards more rigorous and theoretically grounded interpretability research, potentially informing both AI development and computational cognitive neuroscience. (shrink)

Humans have distinctive powers of reflection: no other animal seems to have anything like our capacity for self-examination. Many philosophers hold that this capacity has a uniquely important guiding role in our cognition; others, notably Hilary Kornblith, draw attention to its weaknesses. Kornblith chiefly aims to dispel the sense that there is anything ‘magical’ about second-order mental states, situating them in the same causal net as ordinary first-order mental states. But elsewhere he goes further, suggesting that there is something deeply (...) misleading about reflection’s first-person view of cognition. This chapter takes a closer look at the conditions under which reflection is misleading. Even if humans are the only reflective animals, we are no longer the only creatures that can be questioned about their claims. Chatbots can also be asked to explain themselves, and by studying the conditions under which they produce clarification as opposed to confabulation, we can develop a non-skeptical account of the epistemic value of reflection. I argue that this capacity naturally supports the distribution of complex problem solving among multiple agents. (shrink)

The purpose of this paper is to show that understanding AI hallucination requires an interdisciplinary approach that combines insights from epistemology and cognitive science to address the nature of AI-generated knowledge, with a terminological worry that concepts we often use might carry unnecessary presuppositions. Along with terminological issues, it is demonstrated that AI systems, comparable to human cognition, are susceptible to errors in judgement and reasoning, and proposes that epistemological frameworks, such as reliabilism, can be similarly applied to enhance the (...) trustworthiness of AI outputs. This exploration seeks to deepen our understanding of the possibility of AI cognition and its implications for the broader philosophical questions of knowledge and intelligence. (shrink)

Descartes famously constructed a language test to determine the existence of other minds. The test made critical observations about how humans use language that purportedly distinguishes them from animals and machines. These observations were carried into the generative (and later biolinguistic) enterprise under what Chomsky in his Cartesian Linguistics, terms the “creative aspect of language use” (CALU). CALU refers to the stimulus-free, unbounded, yet appropriate use of language—a tripartite depiction whose function in biolinguistics is to highlight a species-specific form of (...) intellectual freedom. This paper argues that CALU provides a set of facts that have significant downstream effects on explanatory theory-construction. These include the internalist orientation of linguistics, the invocation of a competence-performance distinction, and the postulation of a generative language faculty that makes possible—but does not explain—CALU. It contrasts the biolinguistic approach to CALU with the recent wave of enthusiasm for the use of Transformer-based Large Language Models (LLMs) as tools, models, or theories of human language, arguing that such uses neglect these fundamental insights to their detriment. It argues that, in the absence of replication, identification, or accounting of CALU, LLMs do not match the explanatory depth of the biolinguistic framework, thereby limiting their theoretical usefulness. (shrink)

AI-based technologies are increasingly pervasive in a number of contexts. Our affective and emotional life makes no exception. In this article, we analyze one way in which AI-based technologies can affect them. In particular, our investigation will focus on affective artificial agents, namely AI-powered software or robotic agents designed to interact with us in affectively salient ways. We build upon the existing literature on affective artifacts with the aim of providing an original analysis of affective artificial agents and their distinctive (...) features. We argue that, unlike comparatively low-tech affective artifacts, affective artificial agents display a specific form of agency, which prevents them from being perceived by their users as extensions of their selves. In addition to this, we claim that their functioning crucially depends on the simulation of human-like emotion-driven behavior and requires a distinctive form of transparency—we call it emotional transparency—that might give rise to ethical and normative tensions. (shrink)

Inchul Yum (2024) argues that the widespread adoption of language agent architectures would likely increase the risk posed by AI by simplifying the process of aligning artificial systems with human values and thereby making it easier for malicious actors to use them to cause a variety of harms. Yum takes this to be an example of a broader phenomenon: progress on the alignment problem is likely to be net safety-negative because it makes artificial systems easier for malicious actors to control. (...) I offer some reasons for skepticism about this surprising and pessimistic conclusion. (shrink)

This chapter considers the potential of artificial intelligence (AI) to exhibit creativity and imagination, in light of recent advances in generative AI and the use of deep neural networks (DNNs). Reasons for doubting that AI exhibits genuine creativity or imagination are considered, including the claim that the creativity of an algorithm lies in its developer, that generative AI merely reproduces patterns in its training data, and that AI is lacking in a necessary feature for creativity or imagination, such as consciousness, (...) agency, metacognitive awareness, or the use of mental imagery. In each case, a counterargument is provided that AI already has or soon will overcome the apparent barrier. In some instances, this is because the critical features deemed lacking in AI may not be necessary for human creativity and imagination either. In others, it is because important developmental and structural similarities between humans and generative AI have been insufficiently appreciated. (shrink)

The complexity and emergent autonomy of Generative AI systems introduce challenges in predictability and legal compliance. This paper analyses some of the legal and regulatory implications of such challenges in the European Union context, focusing on four areas: liability, privacy, intellectual property, and cybersecurity. It examines the adequacy of the existing and proposed EU legislation, including the Artificial Intelligence Act (AIA), in addressing the challenges posed by Generative AI in general and LLMs in particular. The paper identifies potential gaps and (...) shortcomings in the EU legislative framework and proposes recommendations to ensure the safe and compliant deployment of generative models. (shrink)

Except in extraordinary circumstances, patients' clinical care should reflect their preferences. Incapacitated patients cannot report their preferences. This is a problem. Extant solutions to the problem are inadequate: surrogates are unreliable, and advance directives are uncommon. In response, some authors have suggested developing algorithmic "patient preference predictors" (PPPs) to inform care for incapacitated patients. In a recent paper, Earp et al. propose a new twist on PPPs. Earp et al. suggest we personalize PPPs using modern machine learning (ML) techniques. In (...) this paper, I argue that, contrary to Earp et al.'s claim, personalized patient preference predictors are neither technically feasible nor ethically desirable. (shrink)

Many philosophers have argued that large language models (LLMs) subvert the traditional undergraduate philosophy paper. For the enthusiastic, LLMs merely subvert the traditional idea that students ought to write philosophy papers “entirely on their own.” For the more pessimistic, LLMs merely facilitate plagiarism. We believe that these controversies neglect a more basic crisis. We argue that, because one can, with minimal philosophical effort, use LLMs to produce outputs that at least “look like” good papers, many students will complete paper assignments (...) in a way that fails to develop their philosophical abilities. We argue that this problem exists even if students can produce better papers with AI and even if instructors can detect AI-generated content with decent reliability. But LLMs also create a pedagogical opportunity. We propose that instructors shift the emphasis of their assignments from philosophy papers to “LLM dialogues”: philosophical conversations between the student and an LLM. We describe our experience with using these types of assignments over the past several semesters. We argue that, far from undermining quality philosophical instruction, LLMs allow us to teach philosophy more effectively than was possible before. (shrink)

The topic of this dissertation is the nature of linguistic competence, the capacity to understand and produce sentences of natural language. I defend the empiricist account of linguistic competence embedded in the connectionist cognitive science. This strand of cognitive science has been opposed to the traditional symbolic cognitive science, coupled with transformational-generative grammar, which was committed to nativism due to the view that human cognition, including language capacity, should be construed in terms of symbolic representations and hardwired rules. Similarly, linguistic (...) competence in this framework was regarded as being innate, rule-governed, domain-specific, and fundamentally different from performance, i.e., idiosyncrasies and factors governing linguistic behavior. I analyze state-of-the-art connectionist, deep learning models of natural language processing, most notably large language models, to see what they can tell us about linguistic competence. Deep learning is a statistical technique for the classification of patterns through which artificial intelligence researchers train artificial neural networks containing multiple layers that crunch a gargantuan amount of textual and/or visual data. I argue that these models suggest that linguistic competence should be construed as stochastic, pattern-based, and stemming from domain-general mechanisms. Moreover, I distinguish syntactic from semantic competence, and I show for each the ramifications of the endorsement of a connectionist research program as opposed to the traditional symbolic cognitive science and transformational-generative grammar. I provide a unifying front, consisting of usage-based theories, a construction grammar approach, and an embodied approach to cognition to show that the more multimodal and diverse models are in terms of architectural features and training data, the stronger the case is for the connectionist linguistic competence. I also propose to discard the competence vs. performance distinction as theoretically inferior so that a novel and integrative account of linguistic competence originating in connectionism and empiricism that I propose and defend in the dissertation could be put forward in scientific and philosophical literature. (shrink)

Communicating interdisciplinary information is difficult, even when two fields are ostensibly discussing the same topic. In this work, I’ll discuss the capacity for analogical reasoning to provide a framework for developing novel judgments utilizing similarities in separate domains. I argue that analogies are best modeled after Paul Bartha’s By Parallel Reasoning, and that they can be used to create a Toulmin-style warrant that expresses a generalization. I argue that these comparisons provide insights into interdisciplinary research. In order to demonstrate this (...) concept, I will demonstrate that fruitful comparisons can be made between Buddhism and Artificial Intelligence research. (shrink)

The idea that humans might one day form persistent and dynamic relationships in professional, social, and even romantic contexts is a longstanding one. However, developments in machine learning and especially natural language processing over the last five years have led to this possibility becoming actualised at a previously unseen scale. Apps like Replika, Xiaoice, and CharacterAI boast many millions of active long-term users, and give rise to emotionally complex experiences. In this paper, I provide an overview of these developments, beginning (...) in Section 1 with historical and technical context. In Section 2, I lay out a basic theoretical framework for classifying human-AI relationships and their specific dynamics. Section 3 turns to ethical issues, with a focus on the core philosophical question of whether human-AI relationships can have similar intrinsic value to that possessed by human-human relationships. Section 4 extends to the discussion of ethical issues to the more empirical matter of harms and benefits of human-AI relationships. The paper concludes by noting potentially instructive parallels between the nascent field of ‘Social AI’ and the recent history of social media. (shrink)