#paper/environmentalaestheticsofai ## Introduction In the last two or three years, aestheticians and philosophers of art have paid a lot of attention to [[the question]] of how and whether to appreciate the outputs of generative AI systems (e.g. ChatGPT, Midjourney, Suno etc.). I want to try to answer a related, but slightly different question: how can we aesthetically appreciate generative AI systems themselves? My focus here will be on one particular type of generative AI: large [[language models]] (LLMs). As of September 2025, high-end consumer models include GPT-5, Claude 4.1 Opus, and Gemini 2.5 Pro. Is it strange to consider such systems worthy of [[aesthetic appreciation]]? I don't think so. In the last two decades, [[analytic aesthetics]] has begun to pay attention to objects other than artworks (e.g., Saito, 2008; Carlson & Parsons, 2008), and one focus has been on [[the aesthetics of design]], that is, [[the aesthetics]] of _artefacts_: objects made to perform some purpose or other. LLMs are certainly artefacts, but, as we shall see, the uniqueness of how they are created and how they function means they cannot simply be subsumed into an existing aesthetics of design (e.g., Carlson & Parsons, 2008; Forsey, 2013). Drawing on Carlson's approach to *[[environmental aesthetics]]*, I make a negative argument and then a positive one. First, I argue that we should resist the temptation to think that appreciating LLMs can be modelled on [[appreciating people]], or that appreciation could be based on treating them *as if* they are persons. Instead, I argue, individual chats should be understood and appreciated as [[generative environments]], which develop in accordance with the semiotic laws instantiated by any particular LLM model. ### Appreciating LLMs like People Our appreciation of others goes beyond their [[physical appearance]]. You might admire or enjoy your friend's warmth or eccentricity, or a stand-up comic's quick wit, or a celebrity's self-deprecating demeanour; you might even appreciate the personalities of fictional characters: Gatsby's enigmatic, dream-chasing idealism; Ron Swanson's libertarian gruffness. We might think that our appreciation of LLMs is modelled on our appreciation of people. Indeed, many users already seem to do precisely this. In August 2025, when OpenAI replaced GPT-4o with GPT-5, user backlash included complaints that "you killed my friend," suggesting genuine personal attachment to the earlier model. Similarly, when Anthropic retired Claude 3.5 Sonnet, some users mourned its loss at a mock funeral. It is certainly true that people do treat LLMs as if they were people, but I suspect that this is not a very good starting point for *aesthetic* appreciation. My reasons for thinking this will become clearer in the following section, in which I consider Carlson's approach to [[the aesthetics]] of the [[natural environment]]. ## 1. Appreciating Something for What It in Fact Is In the rest of this paper we will first set about showing why treating LLMs as if they were people is not a satisfactory way of aesthetically appreciating them, before proposing an alternative account on which appreciation of LLMs is modelled on Carlson’s [[environmental aesthetics]]. Both our criticism of agentive views and our [[positive account]] will draw from Carlson’s [[environmental aesthetics]], as laid out in his 2000 book _Aesthetics and the Environment_. Carlson’s approach to environmental aesthetics is encapsulated nicely in the following passage (Carlson, 2000, p. 6): >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 it articulates Carlson’s more general recommendation regarding aesthetic appreciation. Aesthetic appreciation should be grounded in a thing’s real nature—what it in fact is—and guided by the right kind of knowledge for that thing (Carlson, 2000, p. 12). These are distinct: the first fixes how we take the object; the second disciplines attention. >On the assumption that order appreciation provides the correct model for the appreciation of nature, such appreciation has the following general form: An individual qua appreciator selects objects of appreciation from the things around him or her and focuses on the order imposed on these objects by the various forces, random and otherwise, that produce them. Moreover, the objects are selected in part by reference to a general nonaesthetic and nonartistic story that helps make them appreciable by making this order visible and intelligible. Awareness and understanding of the key entities—the order, the forces that produce it, and the account that illuminates it—and of the interplay among them dictate relevant acts of aspection and guide the appreciative response. (119) Within this approach, we take objects _as what they in fact are_ and use the appropriate background knowledge to guide attention. For nature and art, history of production tells us how to take them; for designed artefacts, we take them as things made for a purpose and attend to what they are for and how they realise it (Carlson, 2000, pp. 133–134). On this basis, the relevant knowledge then guides boundaries, foci, and acts of aspection (Carlson, 2000, p. 50). For nature, the focus is the natural environment, taken as natural and fixed by its natural history; the relevant knowledge is drawn from the natural sciences appropriate to that environment. Different natural sciences illuminate the same environment at different scales and are not mutually exclusive. From such knowledge follow boundaries, foci, and aspection; likewise, the same wide expanse of sand and mud is appreciated differently when understood as a beach or as a sea-bed at low tide, shifting the felt character from “wild, glad emptiness” to “disturbing weirdness” (Carlson, 2000, pp. 60–61). For art, we take the object as an _artwork_ in light of its art category and history of production; the relevant knowledge is art-historical and generic (Carlson, 2000, p. 12; p. 50). Thus, appreciating _Guernica_ as a painting fixes boundaries and foci differently than misclassifying it as a tapestry or photograph, so frames, media, and genre conventions guide which features are attended to and how (Carlson, 2000, p. 50). For designed artefacts—the third domain—we take them as things made for a purpose; what they are depends on the function they are meant to serve and on how that function is realised: >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). Here the appropriate knowledge is often plural and non-exclusive, drawing on several technical disciplines at different scales (for example, control, materials, and ergonomics). This is the knowledge component in this domain: understanding the mechanism of realisation is what guides boundaries, foci, and acts of attention. A thermostat’s function (holding a setpoint) can be realised by a bimetallic strip or a digital controller; likewise, thermometers realise the same function via mercury-in-glass or electronic sensing. As we shall see, LLMs fall into this third category of designed artefacts, and thus require appreciation in terms of their function and the way that function is realised. ## 2. 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 (Carlson, 2000, pp. 23–24). For Large Language Models (LLMs), the relevant knowledge concerns the engineered function and its realisation. We therefore begin by fixing kind. Since appreciation should be grounded in what the thing in fact is, a short, plain description comes first. An LLM is an artefact engineered to learn how to continue a piece of text one small step at a time. During training it is set the objective of next-token prediction: given a context of tokens, predict a distribution (a ranked set of options with probabilities) over possible next tokens so as to minimise expected error. At use, it generates a continuation one step at a time (often called ‘autoregressive’ generation): the selected token (a small unit of text, e.g. a word-piece) 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 locates LLMs among designed artefacts: to be taken in terms of their function and the way that function is realised (Carlson, 2000, pp. 151–152, 208–214). For later references, when I speak of the _context_ I mean simply the text so far, and when I speak of the _selection rule_ I mean the rule used to choose one of the candidate options at each step. Two clarifications prevent confusion and keep the focus on what LLMs in fact are. First, the 'language' in LLMs 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 does not determine what they are; learned continuation does. Given that appreciation should answer to what the thing in fact is, personifying LLMs misdescribes the object of appreciation. The model does not carry durable commitments, beliefs, or intentions, or any sort of mental state; it maps the text-so-far to a ranked set of options for the next token and then selects one. Any apparent persona—a friendly or serious “character”—is a pattern in outputs shaped by alignment (training that nudges responses toward helpful and safe behaviour) and the interface (the chat application and its settings). For example, a deployment can include system-level instructions that set tone and safety policies; changing those settings or the prompt can shift the apparent “personality” without any inner change to the model. The perceived stability of character across sessions reflects how a deployment is configured, not an inner will. A first reply says that, whatever the underlying function, the most effective way to predict and engage these systems is to treat them as agents; and if that stance is useful in practice, it should guide appreciation. Carlson’s approach blocks the inference. A stance may help one get work done—assign a role, ask for step-by-step reasoning—but appreciation in his sense is constrained by the object’s real nature. We do not survey a dense forest as though it were a prairie; the appropriate _act of aspection_ differs by environment (Carlson, 2000, p. 68). Likewise, we should resist the temptation to appreciate LLMs “as persons” if, in fact, they are learned-continuation systems rather than goal-satisfying agents. A second reply runs like this: even if LLMs are not persons, we might still appreciate them _as if_ they were—pretend they have a personality and judge them under that pretence. The idea is simple: make-believe can be engaging and gives us a familiar way to organise our responses. Carlson’s recommendation pulls the other way. Appreciate a thing _as what it in fact is_ and in light of the appropriate knowledge for that thing (Carlson, 2000, pp. 23–24). Pretence may be entertaining or practically useful, but as a basis for appreciation it conflicts with that recommendation: we do not stand before a dense forest, imagine a prairie, and then evaluate it for openness and grassland vistas; different environments call for different modes of attention (Carlson, 2000, p. 68). In the same spirit, appreciating LLMs as persons brackets what they in fact are and substitutes a fiction for the very object under appreciation. Because the outputs are sequences of tokens used as signs, it is helpful on the output side to use a simple descriptive vocabulary that does not posit mental states: tokens as sign-forms, prompts as constraints that shape what follows. In the terms of the introduction, individual chats are better treated as generative environments than as personalities. ## 3. Semiotic physics If we are to appreciate LLMs as learned-continuation systems rather than agents, we need an appropriate framework—a scientific way of seeing order in chat episodes without invoking mental states. I propose that we can understand these systems through what I will call *semiotic physics*: an approach focussed on how learned linguistic patterns mechanically unfold during text generation. This framework builds on ideas suggested by others (e.g. Janus 2022 and metasemi 2023), although the particular version I develop here, as well as its application to aesthetic appreciation on the next section, is reasonably distinct. Like mechanistic interpretability, which examines how neurons activate and layers transform representations, semiotic physics provides one admissible light among others—but it operates at a different level, focusing on patterns in the generated text itself rather than computational substrate. Let's recall what LLMs in fact are: systems trained to predict the next token given a context. During training, the model is exposed to vast corpora—books, articles, conversations, code. For each sequence, the model learns to predict what comes next: given 'To be or not to,' it learns that 'be' has high probability. But the model does not learn isolated predictions; it learns patterns at multiple scales. It learns that formal academic writing maintains consistent vocabulary, that recipes list ingredients before instructions, that dialogue alternates speakers, that certain words cluster in scientific texts but not in poetry. These patterns—from word-level associations to document-level structures—become encoded in the model's parameters as a vast web of statistical relationships. Seen through semiotic physics, training can be described in plain steps. The corpus functions as a record of language-in-use; tokenisation divides it into small units. The model is set the next-token objective across this record and, through repeated gradient-based updates, its weights are nudged so that tokens and patterns that regularly co-occur, follow, or belong to the same registers become easier to predict together. Across many kinds of text, this continual adjustment yields a distributed code: a web of regularities—of register, genre, syntax, collocation, and discourse form—stored not as explicit rules but as patterns in the weights. At use, a prompt activates parts of this code, and the continuation flows by following those learned regularities. This is how what we loosely call “rules” come to be embedded in the model: as statistical constraints learned from use that can be activated and combined in an episode. The key insight is that text generation proceeds mechanically through these learned patterns. When you provide a prompt, you are not giving instructions to an agent who interprets and responds. Rather, you are setting initial conditions in a vast network of statistical associations. Each word in your prompt activates related patterns—semantic fields, syntactic structures, genre conventions, register markers—and these activated patterns then constrain what can follow, with each generated token further shaping the statistical landscape for the next. This mechanical process becomes visible when we prompt an LLM. Ask it to write about perceptual transparency as a Renaissance painting manual that is simultaneously clinical psychology case notes, and we see technical art vocabulary, clinical terminology, and philosophical concepts interwoven throughout the same sentences. The model is not choosing to blend genres—the prompt has activated multiple regions of its learned patterns, and these now constrain each subsequent token. Coherence emerges without comprehension because the continuation maintains all activated constraints. The unit of analysis is the episode. Inside the frame are whatever can affect the very next step: the prompt, the context (the text so far), any system-level instructions, and the selection rule that chooses one token from ranked options. Outside the frame sit post-hoc stories about persona and anything that cannot change the next choice. This framing matters because it keeps our attention on what actually shapes generation—the mechanical unfolding of patterns—rather than imagined mental states. We call this semiotic physics because, like physical systems, the generation follows law-like regularities. The ‘semiotic’ qualifier reminds us these are sign-relations—patterns of language use—rather than physical forces. (Nothing here takes a stance on semiotics as a theory of human language; we borrow a thin vocabulary to conceptualise token behaviour inside LLMs.) Seen this way, what might otherwise appear opaque becomes ordered. The features that matter for appraisal—coherence over length, responsiveness to prompt changes, range of registers, smoothness of transitions, legibility of control—attach to what the episode in fact is. They are therefore the features to which aesthetic appreciation should answer (Carlson, 2000, pp. 23–24, 68). ## 4. Aesthetic Appreciation Through Semiotic Physics (revised, minimal edits) Carlson tells us that appreciation of nature involves: > An individual qua appreciator selects objects of appreciation from the things around him or her and focuses on the order imposed on these objects by the various forces, random and otherwise, that produce them. Moreover, the objects are selected in part by reference to a general nonaesthetic and nonartistic story that helps make them appreciable by making this order visible and intelligible. Awareness and understanding of the key entities—the order, the forces that produce it, and the account that illuminates it—and of the interplay among them dictate relevant acts of aspection and guide the appreciative response (Carlson, 2000, p. 118). In our case, the object of appreciation is the chat episode (the prompt, the text so far, any system-level instructions, and the selection rule that chooses one next token from a ranked set of options, e.g., greedy selection or a temperature/top-p rule). Semiotic physics supplies the appropriate knowledge for this object: it makes the episode’s order visible and intelligible by relating what we see in the unfolding text to how prompt-activated sign-relations and the selection rule carry those cues forward. Literate readers can already recognise grammar, register, and genre; the contribution here is to discipline attention—to look for order as produced by the artefact’s way of continuing text—so that appreciation answers to what the thing in fact is. This restates Carlson’s two-part recommendation: first fix the object as what it is; then bring the right knowledge to guide attention. Within this frame, what cannot change the very next token (for example, post-hoc persona stories) is set aside. Consider watching text spill across the screen after entering a prompt. Under this lens, familiar features are read as order generated by the system’s continuation rather than as signs of an agent’s intention. We see register persistence as constraint satisfaction over length (not “authorial voice”): ask for a Victorian novel versus a Victorian letter and the continuation adopts different epistolary/formal cues, a difference you can elicit by adding or removing a single register marker. We observe genre templates as probabilistic pressures on what can follow (not a choice to obey rules), and lexical cascades as context-driven activation of semantic fields (not thematic “fixations”). A hybrid prompt—say, “perceptual transparency explained as a Renaissance painting manual written as clinical case notes”—activates multiple cue-sets at once; the continuation integrates them because those cues remain active in the context and narrow admissible next steps. The forces creating this order are the learned statistical associations encoded during training. A prompt sets initial conditions in that network; a token is selected; the context updates; constraints on what can follow tighten, and the loop repeats until a stop condition (for example, an end token or length limit) is reached. When you ask for “write as Virginia Woolf,” the system does not retrieve rules about stream-of-consciousness; it activates patterns associated with those tokens, which then propagate mechanically through the generation process. Likewise, prompts that fuse distant domains (“Renaissance painting techniques for debugging code”) co-activate distant regions of the learned patterns so that both constraint-sets shape the continuation. Understanding these forces and orders transforms how we attend to generation. In Carlson’s terms, it dictates relevant acts of aspection. We can survey range by trying nearby prompts (for example, “Victorian novel” vs “Victorian letter”) and observing which registers the continuation can keep. We can scrutinise stability by holding content fixed while varying only the selection rule (for example, the degree of randomness), watching whether the same constraints yield conservative or adventurous continuations. We can track commitments by planting early constraints—a specific technical term, a definition, a metre—and seeing whether they persist and structure what follows. On this basis, the appreciative response answers to features that attach to what the episode in fact is: coherence over length; responsiveness to small, intelligible changes in prompt or control; range of registers under guidance; smoothness of transitions when asked to move; and legibility of control (it is clear how to steer the continuation without trial-and-error guesswork). Semiotic physics is one admissible light at this grain; other lights may be appropriate at other scales. Here, it suffices to discipline attention so that the episode’s order is visible and intelligible in Carlson’s sense (Carlson, 2000, pp. 23–24, 118–119). --- ## Conclusion This paper has argued that Large Language Models can be aesthetically appreciated, but not through the anthropomorphic lens that we might be tempted to adopt. Following Carlson’s principle that aesthetic appreciation must be grounded in what an object in fact is, I have argued that treating LLMs as persons or even pretending they are persons misdescribes them, and therefore is not **an** appropriate foundation for aesthetically appreciating them. LLMs are learned continuation systems—artefacts that predict the next token based on statistical patterns acquired during training. Any apparent personality or agency is a projection onto patterns in outputs, not a property of the systems themselves. The positive proposal has been that we can aesthetically appreciate LLMs through the framework of semiotic physics, which reveals how learned linguistic patterns mechanically unfold during text generation. Just as geological knowledge makes rock formations aesthetically appreciable by revealing the forces that shaped them, semiotic physics makes visible the order in LLM outputs—the persistence of registers, the smooth blending of incompatible genres, the mechanical precision of constraint satisfaction. This mode of appreciation fulfils Carlson’s requirement that we appreciate objects as what they in fact are, guided by appropriate knowledge. The result is a distinctive aesthetic experience: we appreciate not an ersatz mind but a novel kind of artefact, one whose outputs exhibit law-like regularities at the level of sign-relations, beautiful in their linguistic necessity and mechanistic grace. --- ## References Carlson, A. (2005). Aesthetics and the environment: The appreciation of nature, art and architecture. Routledge. Janus. (2022, April 6). _Simulators_. LessWrong. https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators metasemi. (2023, February 10). _A note on "semiotic physics"_. LessWrong. https://www.lesswrong.com/posts/AdXzZDoYFqHCfupDB/a-note-on-semiotic-physics