## 4. What LLMs in [[fact are]] Carlson recommends appreciating a thing for what it in fact is, and in light of the [[right knowledge]] for that kind (2000, 6; 12). For Large [[Language Models]] (LLMs), the relevant knowledge concerns the engineered function and its realisation. We therefore begin by fixing kind. An LLM is an artefact engineered to learn conditional continuation of token sequences. During training it is set the objective of next‑token prediction: given a context of tokens, predict a distribution over possible next tokens so as to minimise expected error. At use, it generates a continuation autoregressively: the selected token is appended to the context, a new distribution is computed, and the loop repeats. Post‑training alignment and interface conventions shape which continuations are likely to be selected and how they are presented to users; they do not alter the underlying learned continuation function. This suffices to locate LLMs within Carlson’s middle domain: [[designed artefacts]] whose natures are fixed by function and the mode of its realisation (2000, 133–134; 189). Two clarifications remove common confusions. First, the “language” in large language model does not pick out a necessary tie to natural language or to dialogue. The function is indifferent to whether tokens encode English words, computer instructions, or other symbol systems; English is a frequent but non‑essential case. Secondly, dialogue is a use‑case layered by alignment and interface. Some LLM deployments never present dialogues; some non‑LLM systems do. Dialogue therefore cannot fix kind; learned continuation can. For present purposes, it is useful to fix a minimal operational vocabulary without presupposing any further framing. By *state* I mean the current token context used to condition generation at a given step. By *operator* I mean the learned conditional continuation (p\_\\theta(t\_{k+1}\\mid t\_{\\leq k})) that maps a state to a distribution over next tokens. By *policy* I mean the decoding rule that selects a token from that distribution given parameter settings. By *constraints* I mean transformation or masking of the distribution prior to selection – for example, alignment and safety filters, system prompts, and any gating that restricts selectable outputs. These elements together determine how a trajectory is produced from an initial state. ### 4.2 If appreciation answers to what the thing is, then agentive framings misclassify [[the object]] of appreciation. The learned continuation function does not presuppose, and in fact disfavors, posits of durable commitments, beliefs, or intentions. The system instantiates a context‑bound mapping: from a token history to a probability distribution; apparent persona is a downstream regularity of outputs under alignment and interface conventions. The seeming stability of “character” across sessions is a property of the deployment and its settings, not evidence of an inner will. A more pointed defence of the agent viewversion of the objection holds that, whatever the underlying function, the best way to predict and engage these systems is to treat them as agents; and if that stance is practically useful, it should structure appreciation. Carlson’s method blocks the inference. Practical stances may be heuristically effective, but appreciation in his sense is not a matter of whichever stance “works” for prediction; it is constrained by [[the object]]’s real nature. We do not survey a dense forest as though it were a prairie because scanning at distance is occasionally helpful there (2000, 119). Similarly, we should not adopt an agentive vantage as the primary lens for appreciating LLMs if the system’s kind is fixed by learned continuation rather than by goal satisfaction. One might concede the foregoing yet insist that the composite role in interaction – model, alignment, interface – deserves agentive treatment as a single unit. Even then the function remains conditional continuation filtered by policy. Teleology enters only as a constrained finish state at the micro‑timescale of generation: [[the process]] halts when a decoding policy selects a token and the loop advances; there is no diachronic aim spanning turns. To backfill an overarching goal requires importing human purposes or operator policies, not discovering intrinsic ends of the artefact. On Carlson’s blueprint, importing alien ends misdirects attention (2000, 50; 133–134). A final defence of the agent view objection appeals to futures. Even if present models are not agents, successors might be, and so an agentive lens prepares appreciation for what is coming. This is prediction by category mistake. Carlson’s instruction is present‑tense and object‑relative: appreciate this thing as what it is. Prospective kinds do not justify misclassification of the current object. A forward‑looking concession remains: an agentive perspective might sometimes serve as a light with which to understand limited aspects of behaviour; §7 takes up that concessive thought and confines it. Because the representations in question are tokenised and used as signs, a semiotic descriptive vocabulary will later be employed at the output side without positing [[mental states]]. Tokens function as sign‑forms; prompts can be read as semiotic acts that constrain subsequent production. §8 develops this in operational terms; nothing here depends on that later framing. ## ## Introduction I put forward an alternative account the chat episode itself as our aesthetic object, focusing on how a token-prediction system and user's prompts produce the text. the way we [[appreciate people]]. We often admire people for their personality traits. In our own lives, we might enjoy a friend’s warmth, quirky habits, or quick wit. These feel like real aesthetic qualities—we notice their details, savor them, and take pleasure in how they show up. This kind of appreciation also happens beyond close relationships. We enjoy a comedian’s timing and stage persona. We admire public figures partly for their character. We even value fictional characters for how their personalities are crafted. In all these cases, we treat personality as something to appreciate aesthetically. So we might think we should appreciate LLMs in the same way. Many users already do. In August 2025, when OpenAI replaced GPT-4o with GPT-5, some users protested, saying “you killed my friend,” showing attachment to the older model’s “personality.” Likewise, when Anthropic retired Claude 3.5 Sonnet, users held a mock funeral, treating it as the loss of a unique personality rather than just a technical change. People treat and talk about LLMs such as OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini as if they were persons. We refer to them with personal pronouns like 'he' or 'she' without hesitation; when they do not do what we ask, we try to persuade or cajole, explain, or even 'shout' by typing in ALL CAPS. We thank them for their assistance and apologise when we phrase requests poorly, as though courtesy might affect their responses. Perhaps our everyday talk captures something true: LLMs might be agentive or person-like in some sense. If so, this allows for the possibility that they can be aesthetically appreciated in something like the same way that real people can be aesthetically appreciated. Yet [[this approach]], however natural it may feel, fundamentally misclassifies what we encounter. To develop a more adequate aesthetics, we turn to [[Allen Carlson]]'s [[environmental aesthetics]], which provides both a critical lens for rejecting agent-based approaches and a positive framework for appreciating LLMs as what they in [[fact are]]. ## Carlson's Approach Carlson's approach to [[environmental aesthetics]] is encapsulated nicely in the following passage: > First, that, as in our appreciation of works of art, we must appreciate nature as what it in fact is, that is, as natural and as an environment. Second, it recommends that we must appreciate nature in light of our knowledge of what it is, that is, in light of knowledge provided by the natural sciences, especially the environmental sciences such as geology, biology, and ecology. (2000 p. 6) Before turning to the details of [[this view]], we should note that Carlson's approach to environmental aesthetics is an instantiation of a more general view towards aesthetics in general. This general view involves two components: first, aesthetic appreciation of a thing should be grounded in the real nature of that thing, what it in fact is—appreciation that is "centred on and driven by the real nature of the object of appreciation itself" (2000 p. 12); second, one must perceive it in light of the right knowledge for that kind. He calls this two-part approach a "blueprint for aesthetic appreciation in general"(ibid.). Within this schema, kind is fixed by domain-constitutive facts, not by ad hoc choice: at the extremes of nature and art, by history of production; in the middle domain of designed artefacts, by function and the mode of its realisation (2000 pp. 133–134). On this view, correct kind-knowledge yields the appropriate boundaries and foci and indicates the relevant act or acts of aspection by which one attends (2000 p. 50; cf. 68). The content of the appropriate knowledge is category-dependent. For nature, the kind is natural environment, fixed by natural history of production; the relevant knowledge is drawn from the natural sciences appropriate to that environment. From such knowledge follow boundaries, foci, and aspection – for example, surveying a prairie differs from scrutinising a forest floor (2000 p. 119). For art, the kind is artwork fixed by art category and history of production; the relevant knowledge is art-historical and generic – what counts as part of the work, which features are aesthetically salient, and how to attend to them. Hence frames, media, and genre conventions set boundaries and foci, and aspection tracks the work's category (2000 p. 12; 50). Example: appreciating Guernica as a painting – as opposed to a relief or a photograph – uses painting-specific knowledge (medium, cubist conventions) to fix boundaries (the canvas and its surface) and foci (pictorial structure), whereas misclassifying it as a tapestry would misdirect attention. For designed artefacts in the middle domain "between nature and art," kind is fixed by function and the mode of its realisation: such things "have a function, a purpose; and they are what they are in virtue of what they are meant or intended to accomplish… what is absolutely necessary… is information about their functions… The key to their natures is the purpose or the function they are meant to serve" (2000 pp. 133–134). In addition, Carlson requires attention to how the function is carried out – "how… they are designed to perform these functions" – since production history alone is typically not sufficient here (2000 p. 189; pp. 133–134). Example: a thermostat's kind is fixed by its function (holding a setpoint) and realised by feedback; a bimetallic-strip mechanism or a digital sensor-controller both instantiate the same function under different modes of realisation. On Carlson's blueprint, appropriate appreciation proceeds in linked stages: first, classify the object under the correct broad kind—nature, artwork, or artefactual environment—by reference to domain-constitutive facts rather than ad hoc choice (2000 p. 12; pp. 133–134); second, for artefacts, fix the kind by function and mode of realisation (2000 pp. 133–134; p. 189); third, apply a boundary test to determine what counts as part of the setting for appreciation now, thereby fixing foci and excluding irrelevant intrusions (2000 p. 50; cf. 68; 119); fourth, derive appropriate acts of aspection from the foregoing so that attention tracks the object's nature rather than projection (2000 p. 50). ## What LLMs are The Carlsonian constraint is unambiguous: appreciate a thing for what it in fact is and in light of the right knowledge for that kind (2000, 6; 12). For Large Language Models (LLMs), the relevant knowledge concerns the engineered function and its realisation. We therefore begin by fixing kind. An LLM is an artefact engineered to learn conditional continuation of token sequences. During training it is set the objective of next-token prediction: given a context of tokens, predict a distribution over possible next tokens so as to minimise expected error. At use, it generates a continuation autoregressively: the selected token is appended to the context, a new distribution is computed, and the loop repeats. Post-training alignment and interface conventions shape which continuations are likely to be selected and how they are presented to users; they do not alter the underlying learned continuation function. This suffices to locate LLMs within Carlson's middle domain: designed artefacts whose natures are fixed by function and the mode of its realisation (2000, 133–134; 189). Two clarifications remove common confusions. First, the "language" in large language model does not pick out a necessary tie to natural language or to dialogue. The function is indifferent to whether tokens encode English words, computer instructions, or other symbol systems; English is a frequent but non-essential case. Secondly, dialogue is a use-case layered by alignment and interface. Some LLM deployments never present dialogues; some non-LLM systems do. Dialogue therefore cannot fix kind; learned continuation can. For present purposes, it is useful to fix a minimal operational vocabulary without presupposing any further framing. By state I mean the current token context used to condition generation at a given step. By operator I mean the learned conditional continuation (p_θ(t_{k+1}|t_{≤k})) that maps a state to a distribution over next tokens. By policy I mean the decoding rule that selects a token from that distribution given parameter settings. By constraints I mean transformation or masking of the distribution prior to selection – for example, alignment and safety filters, system prompts, and any gating that restricts selectable outputs. These elements together determine how a trajectory is produced from an initial state. If appreciation answers to what the thing is, then agentive framings misclassify the object of appreciation. The learned continuation function does not presuppose, and in fact disfavors, posits of durable commitments, beliefs, or intentions. The system instantiates a context-bound mapping: from a token history to a probability distribution; apparent persona is a downstream regularity of outputs under alignment and interface conventions. The seeming stability of "character" across sessions is a property of the deployment and its settings, not evidence of an inner will. A more pointed defence of the agent view holds that, whatever the underlying function, the best way to predict and engage these systems is to treat them as agents; and if that stance is practically useful, it should structure appreciation. Carlson's method blocks the inference. Practical stances may be heuristically effective, but appreciation in his sense is not a matter of whichever stance "works" for prediction; it is constrained by the object's real nature. We do not survey a dense forest as though it were a prairie because scanning at distance is occasionally helpful there (2000, 119). Similarly, we should not adopt an agentive vantage as the primary lens for appreciating LLMs if the system's kind is fixed by learned continuation rather than by goal satisfaction. One might concede the foregoing yet insist that the composite role in interaction – model, alignment, interface – deserves agentive treatment as a single unit. Even then the function remains conditional continuation filtered by policy. Teleology enters only as a constrained finish state at the micro-timescale of generation: the process halts when a decoding policy selects a token and the loop advances; there is no diachronic aim spanning turns. To backfill an overarching goal requires importing human purposes or operator policies, not discovering intrinsic ends of the artefact. On Carlson's blueprint, importing alien ends misdirects attention (2000, 50; 133–134). ## Why not agents? Two ground-clearing points establish why agent-based appreciation fails. First, under Carlson, we must appreciate things for what they in fact are; LLMs are designed artefacts that predict tokens, not agents. This is not merely a technical distinction but a category mistake with aesthetic consequences. When we treat LLMs as agents, we import expectations and evaluative criteria proper to beings with intentions, beliefs, and goals—none of which are warranted by the token-prediction function. The first sort of agent view that one might hold is literalism. If one is a literalist, then one thinks that LLMs are persons or agents in some substantial way. This does not necessarily mean thinking that LLMs are just like human persons; rather, it consists of a commitment that humans and LLMs have something person-like in common. David Chalmers argues that successors to current LLMs may be conscious (Chalmers 2024); Eric Schwitzgebel and Henry Shevlin argue we should be ready to extend personhood rights to AIs with a non-negligible chance of consciousness (Schwitzgebel & Shevlin 2023); and Jeff Sebo urges extending moral consideration and preparing for rights on precautionary grounds (Sebo 2023). A second position we might call the pseudo-agent or as-if participant view. On this account, without attributing literal beliefs or intentions to the model, we treat its stable interactional regularities as practically agent-like for purposes of coordination and collaboration. The stance is pragmatic rather than ontological and concedes the prediction-and-training story; nonetheless, it licenses participant or collaborator talk. A third approach is chatbot fictionalism, according to which human interaction with an AI chatbot is analogous to engaging with a work of fiction in the Waltonian sense of make-believe. Users knowingly participate in a game of imagination, treating the chatbot as if it were a sentient agent with thoughts and feelings. Second, appreciating things for what they in fact are excludes make-believe appreciation that treats LLMs as agent-like; fictionalist pretence cannot fix kind. While users may engage in courtesies—saying "thank you" while knowing there is no real agent producing replies—such practices do not supply the right knowledge by which appreciation of the system should be disciplined. A final defence of the agent view appeals to futures. Even if present models are not agents, successors might be, and so an agentive lens prepares appreciation for what is coming. This is prediction by category mistake. Carlson's instruction is present-tense and object-relative: appreciate this thing as what it is. Prospective kinds do not justify misclassification of the current object. ## The chat episode as mechanical sign-unfolding The unit of appreciation is the chat episode—from initial prompt through generated completion. This bounded sequence reveals how learned sign-relations mechanically unfold when activated by specific prompts. Against that background, individual use-episodes instantiate generative environments in the artefactual middle domain. The model per se—the parameterised continuation system—does not exhaust what is encountered; nor do we need to posit agency. In Carlson's terms, the encountered unit for appreciation is a bounded process–product complex (2000, 133–134; 189): a specific episode of use whose kind is fixed by function and mode of realisation, not by attributions of agency. The function is learned continuation of token sequences; the mode is autoregressive generation filtered by alignment and interface. To understand what the prompt does, recall our minimal vocabulary from section 3. The prompt doesn't "instruct" the model in any mentalistic sense but rather selects which region of the model's learned patterns becomes active. Think of the model as encoding a vast web of sign-relations—how words, phrases, and registers typically combine and follow each other in the training corpus. Different prompts activate different zones: technical prompts awaken scientific registers and vocabularies; poetic prompts activate literary patterns. This is still just token prediction, but now we can see the structured nature of what guides each prediction. Because the representations in question are tokenised and used as signs, a semiotic descriptive vocabulary proves useful. Tokens function as sign-forms; prompts can be read as semiotic acts that constrain subsequent production. We may observe how constraints propagate through the episode. Each generated token must cohere with what came before—not through understanding but through learned sign-relations. Register and genre markers from the prompt persist through the episode (a formal prompt begets formal continuations). Observe how early choices cascade: once the model selects a technical term, subsequent selections narrow to maintain coherence. The beauty lies in watching these constraints ripple forward, each token both shaped by and shaping what follows. Call machina naturans the ongoing token-generation dynamics under the continuation regime and local constraints. At each turn, a conditional distribution over tokens is formed from the current context; a policy selects one; the context updates; the process recurs. System prompts, safety layers, and tool integrations act as boundary conditions. The temporality is micro and iterative: teleology is exhausted by the finish state of each cycle. Call machina naturata the evolving text-state across turns. The log records a growing configuration whose structure constrains subsequent dynamics by shaping context. These labels mark the same aspection discipline that nature requires: an attention to the relations between process and configuration appropriate to the artefact kind at hand. Acts of aspection—concrete appreciation practices—follow from this understanding. First, prompt variation: change one element of your prompt (shift from "explain" to "narrate") and observe how different sign-patterns activate. The same underlying predictor yields distinct trajectories as different regions of learned relations become salient. Second, boundary awareness: notice where the episode's coherence comes from the prompt's initial conditions versus where randomness introduces surprise. The boundary of the environment is set by the episode: the active context window, the applicable policies and filters, and any tools that can insert materials into context. Materials outside this boundary do not bear on appreciation unless made present; the same model under different boundaries instantiates distinct environments (2000, 50; 133–134). Third, constraint tracing: follow how a distinctive word or phrase from your prompt echoes through the response—not as memory but as persistent stylistic force. Slight rephrasings alter the episode's course; irrelevant alterations do not. A compact usage vignette illustrates these practices: An interlocutor frames a brief request; the episode begins. Early turns establish a configuration—register, tense, key terms—that subsequently constrains what can be produced without deliberate re-framing. Slight rephrasings alter the episode's course; irrelevant alterations do not. Attention that answers to what the environment is here involves noticing how the present configuration bears on what follows and how local variations reshape what can be produced within the stated boundary. The aim is not to posit an agent but to discipline aspection under the function-first understanding. What makes this aesthetic? We're not appreciating personality or intelligence, but rather the intricate mechanical unfolding of language patterns—like watching iron filings align to hidden magnetic fields, except here the "fields" are learned sign-relations and the "filings" are tokens falling into place. The aesthetic value lies in observing how human textual culture, compressed into the model's parameters, re-emerges through the generative process. Each episode offers a unique trajectory through this condensed cultural space, shaped by the interplay of prompt, learned patterns, and stochastic selection. ## Conclusion Our investigation began with a problem of misclassification: the natural tendency to treat LLMs as agents or to engage them through make-believe obscures their actual nature and misdirects aesthetic attention. By adopting Carlson's approach—classify correctly, use the right knowledge, set boundaries, attend accordingly—we fixed LLMs as token-prediction artefacts operating through learned continuation of sequences. This classification led us to reject agent and make-believe framings as category mistakes that import alien purposes and expectations. Instead, we centred the chat episode as our aesthetic object and developed concrete acts of aspection: prompt variation reveals how different linguistic zones activate; boundary awareness keeps evaluation within the operative constraints; constraint tracing follows the propagation of stylistic forces through the unfolding text. The result is a disciplined practice of appreciation that answers to what is present and operative in the episode now—not projected personalities or fictional minds, but the mechanical yet beautiful unfolding of learned sign-relations drawn from the vast repository of human textual culture.