# [[Generating Philosophy with AI]] >“Forty-two," said [[Deep Thought]], with infinite majesty and calm. It was a long time before anyone spoke. Out of the corner of his eye Phouchg could see the sea of tense expectant faces down in the square outside. "We're going to get lynched aren't we?" he whispered. "It was a tough assignment," said [[Deep Thought]] mildly. "Forty-two!" yelled Loonquawl. "Is that all you've got to show for seven and a half million years' work?" "I checked it very thoroughly," said [[the computer]], "and that quite definitely is the answer. I think the problem, to be quite honest with you, is that you've never actually known what [[the question]] is." The advancement of [[artificial intelligence]], particularly the development of Large [[Language Models]] (LLMs) such as GPT-4, prompts inquiry regarding their potential role in domains traditionally considered human; [[philosophical inquiry]] numbers among these domains. Demonstrating proficiency in generating fluent, coherent, and contextually relevant text, these models stimulate discussions about their capacities for understanding, reasoning, and potentially creativity. In this paper, we explore a specific aspect of this topic: the potential for LLMs to enhance human [[philosophical understanding]]. The central thesis advanced here is that LLMs might serve as effective tools to aid humans in achieving a specific type of [[philosophical understanding]] – namely, the accurate and comprehensive representation of dependence networks, following the account developed by Dellsén et al. (2024) – without needing to possess such understanding themselves. This potential arguably arises because LLMs function by extracting and recombining statistical regularities from vast datasets; many of these regularities, it seems, implicitly track conceptual or explanatory dependencies relevant to [[philosophical inquiry]]. Consequently, this analysis distinguishes between the functional utility of LLMs as aids to human cognition and the separate question of whether these systems possess understanding, beliefs, or semantic competence in the human sense. Our focus here is on the former: how might these pattern-matching systems contribute to [[the process]] of philosophical reflection and the refinement of understanding in human users? To address this, we shall first outline the target conception of understanding, justifying its use. Secondly, we will introduce LLMs and their relevant capabilities, clarifying the scope of the inquiry regarding LLM cognition. Following this, we shall explain the mechanisms by which they might facilitate the representation of dependence networks, drawing on examples from [[philosophical practice]] and considering findings from interpretability studies. Finally, we offer thoughts on the implications of using these tools in philosophy. ## 1 Understanding as [[Representing Dependence Networks]] To assess how LLMs might enhance [[philosophical understanding]], we first require a definition of this understanding. In this paper, we adopt the framework proposed by Dellsén et al. (2024). Their approach, explaining understanding in terms of representing dependence networks, resonates with the broader philosophical aim, articulated by Sellars (1962, p. 1), to grasp 'how things in the broadest possible sense of the term hang together in the broadest possible sense of the term'. Sellars conceived philosophy as pursuing a synoptic vision, integrating diverse entities, concepts, events, norms, and theoretical posits into a coherent whole by mapping their causal, logical, conceptual, and explanatory interconnections. Dellsén et al. offer a specific way to construe this ‘hanging together’ for evaluating philosophical progress, focusing on the representation of dependence relations. Their account aims not simply to analyse the ordinary meaning of 'understanding', but rather "to explicate it in a way that makes it suitable for being deployed in an account of philosophical progress" (Dellsén et al. 2024, p. 674). The central claim holds that the degree to which a subject understands a phenomenon corresponds to the accuracy and comprehensiveness of their representation of the network of dependence relations relevant to that phenomenon. These dependence relations are conceived as objective, worldly relations often underpinning explanations. As Dellsén et al. (2024, p. 675) note, drawing on Kim (1994), dependence relations constitute the ontological correlates of explanation. While causation is paradigmatic in empirical science, other candidates include constitution, grounding, mereological dependence, truthmaking, conceptual containment, and supervenience; the framework remains agnostic about the precise list, assuming understanding involves representing whichever relations obtain in the domain. Furthermore, understanding a phenomenon X requires grasping its position within a wider network; it involves representing both how X depends on other phenomena and how further phenomena depend on X (Dellsén et al. 2024, p. 675). Such representation includes facts about existing dependencies – positive dependencies – and non-dependencies – negative dependencies, e.g., that X lacks dependence on Y. Understanding, on this view, admits of degrees relative to accuracy and comprehensiveness. Accuracy concerns how correctly the representation maps the dependencies that actually exist; comprehensiveness concerns the extent to which the representation includes relevant phenomena and their dependence relations (or lack thereof). These criteria can be in tension – sometimes motivating trade-offs like idealisation or abstraction (Dellsén et al. 2024, p. 675). This account is described as "robustly factive," requiring the representation to correspond to actual dependencies, yet also "epistemically undemanding" (Dellsén et al. 2024, pp. 674, 676). Understanding X, in this specific sense, does not require having justification for, or even belief in, the propositions represented. Dellsén et al. argue this feature is appropriate: "an explication of understanding that does not imply justification is arguably better suited for spelling out a plausible understanding-based account of philosophical progress" (2024, p. 676). Their motivation appears to be accommodating early-stage theorising or preliminary hypotheses where confidence might be low, yet the represented structure could still be largely accurate. This specific notion – understanding as accurate and comprehensive representation of dependence networks, without requiring justification – guides the subsequent analysis. While richer notions of understanding exist, this particular account provides a clear target for assessing the potential contribution of information-processing tools like LLMs. Enhancing philosophical understanding, therefore, involves refining this mental representation of the dependence network. Encountering arguments or information prompts evaluation: would incorporating this input yield a more accurate or comprehensive representation? For example, exploring theories of time might refine an initial map linking temporal passage simply to change. Distinguishing subjective passage from objective ordering enhances accuracy. Incorporating dependencies linking temporal theories to metaphysics (presentism, eternalism) or physics increases comprehensiveness. Mapping negative dependencies (e.g., B-series ordering not depending on subjective experience) or specifying dependence types (conceptual, metaphysical) improves clarity. The individual adjusts their model – adding, removing, or altering represented dependencies. The resulting revised representation constitutes enhanced understanding by virtue of improved accuracy or comprehensiveness, irrespective of justification for every element. Having defined this target concept, we turn now to how LLMs might assist humans in this process. ## 2 Large Language Models Large Language Models (LLMs) are computational systems based on large artificial neural networks, primarily trained to predict the next token in a sequence given prior context. Typically utilising the Transformer architecture (Vaswani et al. 2017), they employ self-attention mechanisms to weigh the importance of different words in the context; this allows the handling of long-range dependencies relevant for coherent text. Given the vast scale of training data relative to model parameters, LLMs cannot merely memorise text. Effective prediction requires compression, forcing the model to identify and represent recurring patterns and structures efficiently. This results in internal states or latent representations capturing the statistical essence of language, including grammar, phrasing, and semantic associations. These representations are distributed across components such as embedding layers (encoding word similarities), self-attention heads (managing information flow between tokens), and feed-forward layers (processing representations). Empirical work demonstrates that specific components – for instance, attention heads – can learn functions that track syntactic dependencies (e.g., subject-verb agreement) or semantic relationships (e.g., coreference), suggesting some capacity within the architecture to represent relational structure explicitly (Clark et al. 2019). However, a distinction should be drawn between latent representations that merely correlate with dependence relations found in the data, and internal states that genuinely encode such dependencies in a robust way. Philosophical analysis often requires understanding dependence relations (like grounding or causation) that possess modal or metaphysical weight beyond mere statistical association. One might object, therefore, that the statistically derived patterns captured by LLMs might fall short of representing these deeper relations. This limitation warrants consideration and will be revisited (see §5). Contemporary LLM training often includes fine-tuning, frequently involving Reinforcement Learning from Human Feedback (RLHF) (Christiano et al. 2017; OpenAI 2022). In this process, human feedback trains a reward model predicting preferences, and the LLM is subsequently trained to maximise these scores, aligning its behaviour with qualities such as helpfulness and honesty. Butlin and Viebahn (2024) argue that this alignment process, particularly when feedback implicitly or explicitly rewards accuracy, can imbue the model's outputs with 'descriptive functions' – a functional orientation towards providing informative and factually correct responses, beyond simply generating statistically plausible sequences. This pressure towards accuracy, rather than simply politeness or coherence, strengthens the case that LLM outputs might reflect genuine dependencies. The result of pre-training (focused on prediction) and fine-tuning (focused on alignment) is that LLMs develop internal representations encoding information about how concepts related to causation, conditionality, constitution, logical entailment, grounding, etc., are interconnected in language and often reflect worldly structures. Connecting back to §1, this internal network constitutes a vast repository of candidate dependence relations. Because these representations are learned from text reflecting actual dependencies and are oriented towards informative output, LLM-generated text can serve as relevant input for a philosopher mapping dependence networks (à la Dellsén et al.). The model's output provides structured information for evaluation. Thus, the LLM functions as a tool, offering access to representations of dependencies without needing understanding itself. Acknowledging limitations such as training data biases or alignment distortions, this capacity forms the basis for exploring practical applications. ## 3 Practical Applications for Motivated Philosophical Prompters The prospect of using LLMs to enhance philosophical understanding faces immediate objections. Common experience with superficial AI-generated text, particularly in educational contexts, raises concerns about misuse hindering rather than helping genuine comprehension. Specific failure modes relevant to philosophy include cognitive outsourcing, where reliance on the LLM bypasses the user's own analytical work needed to build an internal representation of dependencies. Furthermore, users lacking sufficient background knowledge may struggle to evaluate LLM output for subtle inaccuracies, potentially incorporating flawed dependency claims – representing an evaluation impasse where those needing accurate information may be least equipped to verify it. Some pedagogical literature on AI-assisted writing explores related challenges. Acknowledging these concerns regarding misuse and interaction dynamics, this investigation focuses on a different scenario: the ‘motivated philosopher prompter’. This refers to an individual, often possessing philosophical training, who engages critically and interactively with an LLM. Their goal is not to replace their own thinking or automatically generate a final product, but rather to employ the LLM strategically as a tool – for brainstorming, exploring connections, structuring arguments, or accessing information – specifically aiming to enhance their own understanding, construed as making their internal representation of philosophical dependence relations more accurate and comprehensive (cf. §1). The following sections explore how such a user might leverage LLMs for this purpose, starting with minimal refinements (§4) and progressing towards more substantial restructuring (§5). ## 4 Minimal Enhancements: Refining Dependence Networks The previous sections outlined how LLMs might, in principle, facilitate the representation of dependence networks. This potential becomes clearer when considering ‘minimal understanding enhancements’: the incremental refinements in accuracy or scope that occur as one processes new philosophical information. While major shifts certainly occur, much cognitive work involves these smaller steps, gradually improving the fidelity and coverage of one's internal representation of dependence structures. According to the adopted framework (Dellsén et al. 2024), each instance where a represented dependence becomes more accurate (correcting a misidentification, specifying the type) or the network becomes more comprehensive (incorporating a new factor, mapping a new connection) constitutes an increase in understanding. These minimal enhancements arguably function as building blocks for more complete insight. Consider an individual reflecting on memory and past events. Their initial representation might be simple: [Past Event E] ---> [Current Memory M], where ---> signifies general dependence. This representation lacks detail. Encountering distinctions such as episodic memory (re-experiencing E, autonoetic consciousness) versus semantic memory (factual knowledge about E, noetic consciousness) (Tulving 1985) prompts refinement. The single node ‘M’ might split into ‘Episodic Memory of E’ (M_ep) and ‘Semantic Memory that E occurred’ (M_sem). Dependencies become more specific: M_ep might depend on the original first-person experience and self-awareness capacities; M_sem might depend on receiving information about E (via experience, testimony, etc.) and factual recall capacities. The revised network might include: [First-person Experience of E] ---> [M_ep] (Dependence: Grounding/Causal Basis) and [Information that E occurred] ---> [M_sem] (Dependence: Informational Basis). Negative dependencies might be added (e.g., M_ep not solely dependent on information transmission). This adjustment – replacing one dependency with a more articulated network – constitutes a minimal understanding enhancement, increasing accuracy and comprehensiveness. This process illustrates the cognitive refinement LLMs could facilitate for a motivated prompter. An LLM prompted about memory might introduce or clarify the episodic/semantic distinction. Specific prompts could yield targeted information: - "Explain the distinction between episodic and semantic memory" might elicit definitions clarifying the nodes (M_ep vs M_sem) and their properties, enhancing accuracy. - "What cognitive capacities does episodic memory depend on?" could generate information helping map the inputs/grounding conditions for M_ep, increasing comprehensiveness and specifying dependencies. - "Summarise arguments linking episodic memory to imagining the future" might yield information about the constructive episodic simulation hypothesis (Schacter & Addis 2007), helping map outputs/consequences of M_ep (its relation to prospective thought), increasing comprehensiveness. - "Are episodic and semantic memory fully independent?" could lead to discussion of interactionist views, prompting refinement of the relationship between M_ep and M_sem (adding bidirectional dependencies, qualifying distinctness), enhancing accuracy. By generating text articulating definitions, dependencies, arguments, and nuances, the LLM provides structured input relevant to refining the user's dependence network. The LLM lacks understanding and does not perform the refinement; its output serves as organised input that the user critically evaluates and integrates, facilitating their own process of achieving minimal understanding enhancements. Refining the language used to articulate ideas also sharpens the dependence network representation. If one initially represents the consciousness-brain link using ‘depends on’, asking an LLM for precise alternatives might yield ‘is constituted by’, ‘supervenes on’, ‘is realised by’, or ‘emerges from’. Each term implies a different dependence relation. Evaluating these forces clarification: choosing ‘supervenes on’ over ‘depends on’ increases the accuracy of the represented relation. This process of LLM-assisted linguistic refinement translates into conceptual clarification and a more accurate dependency map. This echoes historical instances where terminological precision reshaped understanding; Kripke's introduction of ‘rigid designation’, for example, markedly altered the perceived dependence network connecting naming, necessity, and reference. Even seemingly minor linguistic adjustments prompted by an LLM can thus function as minimal enhancements, improving representational fidelity by forcing greater precision. While distinct from the large-scale breakthroughs discussed next (see §5), these minimal enhancements build the foundation for more comprehensive understanding. ## 5 Comprehensive Breakthroughs Beyond minimal enhancements, intellectual history and individual cognition feature ‘comprehensive breakthroughs’ involving rapid, large-scale restructuring of dependence network representations; these lead to substantial shifts in perspective. Such breakthroughs occur personally (grasping a complex theory, undergoing transformative experience (Paul 2014)) and collectively (Kuhnian paradigm shifts (Kuhn 1962)). Can LLMs, processing linguistic patterns, facilitate such breakthroughs, or only minor adjustments? Certain limitations exist. Breakthroughs often involve creativity, insight, and grasping significance – capacities linked to genuine understanding, intentionality, and experience, which LLMs arguably lack. Many breakthroughs, moreover, rely on empirical grounding or embodiment, areas where LLMs are limited. However, these objections may not be decisive. If LLMs possess a descriptive function enabling accurate mapping of dependence relations (as argued in §2), this might provide relevant input for a user nearing a breakthrough. The LLM need not grasp the significance; it must merely present connections allowing the user to grasp it. It need not experience insight; it might generate the novel juxtaposition or synthesis that triggers insight in the user. The cognitive event remains human, facilitated by LLM-provided stimulus. One might object concerning creativity versus recombination; this hinges on whether the patterns underlying conceptual revolutions are discoverable. If breakthroughs involve non-random conceptual restructuring patterns, and if these patterns are representable within the high-dimensional spaces learned by LLMs from vast textual data, then LLMs might generate outputs reflecting pathways towards these restructurings (e.g., identifying analogies, revealing hidden assumptions, synthesising information novelly). The grounding/embodiment limitation remains substantial for empirical sciences or philosophies of lived experience. Yet, in abstract domains – mathematics, logic, parts of metaphysics/epistemology – where breakthroughs rely more on manipulating complex symbolic systems and abstract structures, the LLM's disembodiment is less detrimental. Its useful ability to process information, identify patterns, explore logical consequences, and structure complexity could prove beneficial. Consider Alan Guth's development of inflationary cosmology, emerging from synthesising constraints from particle physics (e.g., grand unified theories) with cosmological puzzles (e.g., horizon/flatness problems). An LLM processing vast scientific literature might potentially surface such cross-domain dependency clashes or candidate syntheses for evaluation by a theorist. Therefore, while LLMs cannot have breakthroughs, they might function as catalysts. By synthesising information, revealing latent structures, and structuring complexity, they could aid a prepared human mind in achieving large-scale network restructuring, particularly in abstract domains. This facilitation requires careful human verification; breakthroughs are epistemically productive only if the new map is more accurate, not merely novel. Iterative checking against empirical data or argumentative constraints must gate LLM suggestions before they revise understanding. A potential methodology involves ‘serious play’: systematically prompting an LLM for alternative dependence schemas for critical evaluation. The LLM's role seems less about originating insight and more about creating conditions where human insight is more likely.