You are to finish an analytic‑philosophy paper on [[the environmental aesthetics of generative AI]] (LLMs). Follow these instructions exactly. Goal and audience - Produce a submission‑ready paper suitable for venues like [[Philosophical Studies]] or Noûs. - Tone: analytic, clear, straightforward, unpretentious. Confirm understanding first - Before any plan, briefly confirm you understand these instructions in their entirety by paraphrasing and listing any open questions or missing files you need. Be thorough with the open questions. it is much better to sort things out later than for you to go off on the wrong track. Materials - Use and cite the attached files: papers in the project/files folder (including, but not limited to “Not minds but signs,” “Simulators,” and Carlson 2000), the old draft, the new draft, and the prior chat on [[semiotic physics]]. Also consult and adhere to the project [[Style Guide]] (the document referenced as “under a letter” in the project instructions). - Do not invent quotes or page numbers. If a required source is missing, ask for it; otherwise paraphrase cautiously and mark \[CITE/VERIFY\]. - Use author–date citations with page numbers for direct quotes. Formatting - Keep the Introduction heading as in the draft. For the numbered sections, use only top‑level numbered headings with very succinct, pithy titles: 1., 2., 3., … - No subsections (e.g., no 2.2, 3.4) and no sub‑heads within sections. - Length target: 8.000–10.000 words. Provide a per‑section budget and stay within ±10%. The detail in this plan is not itself an indicator of length; use judgment. Nonnegotiables for existing text - Keep the Introduction and Sections 1, 2, and 3 as in the draft, with only light editorial cleanup (typos/grammar). Do not alter their content or [[argumentative structure]]. - In the Introduction, replace the placeholder with one succinct paragraph summarising the paper’s structure. - Preserve Section 3’s Carlson quotes; verify wording and pagination; retain the part that generalises Carlson’s blueprint to works/artefacts (e.g., architecture), as already present in the draft. New [[content requirements]] 4. What LLMs in [[fact are]] - Argue that, on Carlson’s blueprint for artefacts (function + mode of realisation), LLMs are best understood as token predictors (or, if you think the detail is necessary ‘token‑sequence continuation engines’) and their mode of realisation is to do with their training, the corpora that they are trained on, the architecture produced by [[the training]] etc. etc. (you should decide exactly how things are described here, think [[not just]] about what has preceded this section but what has succeeded it) - This argument should be achieved through a dry, clear technical account of LLM training and functioning, and should also serve as a primer to those readers unfamiliar with how these things are made (although the readers of this article analytic philosophers will be able to follow complex arguments and ideas, they might lack technical knowledge of llms. the balance that is needed here is between making llms understandable to the readership, without dumbing down so much that the details of the philosophical ideas are lost. (the following quote from chalmers is a good starting point for detail + straightforwardness: "[[Language models]] are systems that assign probabilities to sequences of text. When given some initial text, they use these probabilities to generate new text.”) - Avoid “environment”, “naturans”, “[[semiotic physics]]” terminology in this section. the description should be neutral. - In the [[second half]] of the section (remember NO subsections) show how this presents a problem for agentive views. They are not appreciating LLMs for what they in [[fact are]]. make sure to give enough to detail so that this argument is clear to the reader, there is nothing person like about token prediction so it is hard to see how participation, personality views on aesthetics can really work 5. Nature understood (Carlson + Spinoza) \[NOTE, a lot of the text in this section can be taken, pretty much verbatim, from Section 1 (not the introduction) of the old draft that you will find in the file folder. draw on this where appropriate.\] - This section should make clear carlson’s [[environmental aesthetics]] (remember, section 3 primarily discusses the principle of fix object + understand in the light of relevant knowledge, here we are getting to how this idea is applied to the environment specifically.) - Repeat the block quote at the beginning of the paper to remind the readers of what carlson says: > 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) - Exposit Carlson’s two‑part schema for [[appreciating nature]]: appreciate as what it is (natural environment) and in light of the appropriate sciences. - Introduce Spinozist framing: natura naturans (generative, productive) vs natura naturata (the produced). Explain how this illuminates Carlson’s view: sciences as lenses tracking relations within and between naturans/naturata across levels. Make sure to add enough detail so that this idea is adequately clear to the reader. Another phrase ‘generative environment’ should be quickly, succinctly introduced, as shorthand for the naturans/naturata aspect. (this phrase should be returned to in the next section) - Very succinctly introduce and use ‘generative’, ‘generative environment’ as a way to speak about the natura/naturata aspect of nature where apt. 6. Can agent views be rescued? - We saw earlier that agency doesn’t really capture what LLMs in fact are, but perhaps it can be a ‘light’ in which to understand LLMs (analogous to a natural science)? - First half: consider an “agent‑light” appreciation informed by cognitive/psychological/neuroscientific lenses, analogising to the person‑level within natural environments (as we already do for the people within nature, drawing on biology, cognitive science, and neuroscience to understand that level). - Second half: critique this move for LLMs. Use the files “Not minds but signs” and “Simulators” papers if available. - Conclude that psychology‑style lights are generally ill‑suited bases for LLM appreciation, given their training objective and realisation. 7. Chats as generative environments - Argue that individual LLM chats instantiate generative environments. Introduce and consistently use the terms machina naturans (the ongoing generative process of token continuation under constraints) and machina naturata (the produced text stream/state), make it clear how they are analogous to natura naturata and naturans. - Make it clear: not that LLMs themselves are environments, but that chat episodes are instances of generative environments, LLM models, the weights and the prior tokens, \*determine\* how LLM chats evolve over time – that is, the 8. Semiotic physics \[remember this section and section 9 are the cumulation of all the argument, they need to be substantial and robust and follow elegantly on from what has preceded them\] - Introduce the idea of semiotic physics You can help get an idea of the level of abstractness, and/or the level of detail required here by looking at the old llm chat in the files folder. - Tie it to machina naturans/machina naturata: how sign dynamics, constraints, and transformation laws govern generative processes and products in chats. 9. An aesthetics of LLMs - Show how semiotic physics + machina framing yields concrete acts of aspection and criteria of appreciation (e.g., affordance legibility, responsiveness to contextual framing, stability/variety under perturbations, semiotic coherence/phase transitions). - Situate this among other possible lights (explicitly mention mechanistic interpretability as another lens); clarify complementarity and limits, and show how the approach fits the environmental‑aesthetics framing developed earlier. Conclusion - Provide a succinct conclusion (two paragraph max\] that restates the thesis and contributions. Process 1) Confirm understanding (return first; await “OK”): - Paraphrase the instructions and list open questions/missing files. 2) Preflight plan (return; await “OK”): - Section‑by‑section word budget (8.000–10.000 total); key claims; required citations (list precise quotes to verify); unresolved terminology choices (e.g., final policy for natura vs machina usage). 3) Drafting in stages: - Fill the Introduction placeholder paragraph. - Deliver Sections 4–6; pause for feedback. - Deliver Sections 7–9; pause for feedback. - Deliver the Conclusion and a formatted references list from used sources. 4) Verification and safety: - Verify quotes/page numbers against attachments; otherwise paraphrase and mark \[CITE/VERIFY\]. - Remove or bracket \[VERIFY\] any post‑2024 news or product claims unless supported by attached sources. - Standardise names (Carlson; Anscombe), fix typos, remove stray footnote markers only if they are orphaned and not required by the draft’s citation apparatus. Success criteria - Fidelity to Sections 1–3 and the Introduction (save placeholder fill). - Coherent, non‑agentive foundation in Section 4; clear Carlson+Spinoza in Section 5; robust chat‑as‑environment case with teleological finish state in Section 6; fair rescue+critique in Section 7; precise semiotic physics in Section 8; actionable appreciation criteria in Section 9. - Consistent terminology; accurate citations; no fabricated facts; 8.000–10.000 words; no subsections; very succinct section titles. 6. Parameter & tool recommendations • reasoning\_effort: high • agentic\_eagerness: medium (autonomous drafting with pauses for confirmations/verification) • verbosity: medium overall; higher within drafted sections and citation notes • max\_output\_tokens: ensure capacity for staged delivery (plan → 4–6 → 7–9 → conclusion) • Responses API: chain with previous\_response\_id if available to avoid re‑pasting long context • Tool budgets: use retrieval only for attached PDFs/notes; avoid open‑web browsing without explicit user consent 7. Stop conditions and safety gates • Stop and ask before: • Proceeding without the project/files folder, the style guide, or key texts (“Not minds but signs,” “Simulators,” Carlson 2000). • Quoting Carlson or others when PDFs are not attached or page numbers are unknown. • Including any post‑2024 factual/news claims (mark \[VERIFY\] or remove). • Exceeding 10.000 words or deviating from “no subsections”/title succinctness. • Flag and seek confirmation on: • Final terminology policy for natura vs machina usage and any other naming conventions. • Any substantial edits to Sections 1–3 beyond light cleanup. 8. Self‑check and residual risks • Consistency: Edits applied; “semiotic physics” fixed; Section 4 environment‑terminology ban reiterated; Introduction heading preservation clarified; pithy titles enforced. • Completeness: All required sections and constraints included; teleological finish state added; generative environment clarified. • Safety: Quote/news verification gates; no fabrication; \[CITE/VERIFY\] markers allowed. • Scope: Focus maintained on Carlson, Spinoza, machina framing, semiotic physics, mechanistic interpretability as comparative lens. DRAFT: the environmental aesthetics of generative AI Introduction 1. AI as (quasi) Agents 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 (we shall see some examples of this in the next section). This section will sketch three different flavours of this view. 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. Cross (2024), writing on AI art, exemplifies this approach when he argues that prompting, iteration, and sampling structure a dialogue where value lies in the interaction itself. "By adjusting inputs, iterating, and sampling," he writes, "an AI artist is engaged in a process of mapping – and perhaps interrogating – the way that the algorithm sees and understands" (Cross 2024, 7–8), although he concedes that "the analogy with performance art isn't a perfect one" (Cross 2024, 9). Anscomb could also be thought of as endorsing a pseudo-agent view. She denies literal mentality and creativity in present systems – "we are not yet at the stage where an AI can formulate intentions … I argue that AI agents cannot be artistically creative" – and adopts a procedural label for "AI agent" as "a self-contained ('autonomous') procedure". Yet she also holds that an AI "may work iteratively without human intervention to non-accidentally generate the formal features of an image" and that it can merit "some share of production credit", which positions her as treating AI as a pseudo-agent, an as-if collaborator rather than a mind-bearing agent. Nonetheless, her vocabulary pressures toward minimal literalism: AICAN is said to be "able to self-assess these products", and "an AI agent arguably deserves credit for its contribution to the production of a work qua art" – locutions typically reserved for bearers of credit rather than mere instruments. 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. This position has been developed by Mallory (2023), Krueger & Roberts (2024), and Krueger & Osler (2022), whilst Friend & Goffin (2025) discuss the view without endorsing it. As Friend & Goffin observe, prop-oriented make-believe explains ordinary exchanges with Alexa and ChatGPT, where users converse "as when we say 'thank you' … while knowing that there is no real agent producing the replies" (Friend & Goffin 2025, 9). By contrast, content-oriented make-believe clarifies richer, empathetic uses where "it is the interaction itself that matters", with users imaginatively treating the chatbot as a person within the ongoing exchange. %%Succinct one paragraph summary of the structure of the paper. %% 2. Appreciating LLMs as agents People, and perhaps other conscious beings (see Marchetti?), seem like a sort of thing that can be aesthetically appreciated. Most importantly for our interests here, this appreciation is not limited to physical beauty. If LLMs are a type of agent or quasi-agent, this might well allow for them to be aesthetically appreciated in a somewhat similar manner. In this section, we look at two ways in which this idea might be elaborated. The first route treats the system as a participant within an artist-structured interaction. On this approach, the evaluative focus shifts away from freestanding outputs toward what the duo does together through prompting, iterating, and mapping the system’s way of seeing. As Cross observes: By way of their selection of prompts, they elicit a sort of "participation" on the part of the AI … This participation allows them to interrogate the algorithm's latent space. (Cross 2024, 8) According to Cross, the interactional performance is primary; the images or videos are often documentation of that practice rather than the locus of value. He argues that "what is centrally important … is that it is less centrally focused on the output of the AI image generator. Instead, what matters is the interaction between the artist and the algorithm" (Cross 2024, 8). The product can thus function as documentation of the interactional process. Cross calls this the “exploration paradigm”: artists “relate to AI as a participant” and “create a space for interaction … by way of their prompts,” shifting appreciation “towards the artist’s interaction with the AI and the way in which the artist structures this interaction.” The generated images “function as a kind of documentation or evidence” of that exploration rather than the locus of value (Cross 2024, abstract; 9). The agent‑based framing is explicit: the work lies in a temporally extended dialogue in which the artist elicits “participation” from the model to “interrogate the algorithm’s latent space,” making the algorithm’s ways of seeing and representing the object of appreciation (Cross 2024, 8–9). Cross presents this as an as‑if stance and concedes that the performance‑art analogy “isn’t a perfect one” (Cross 2024, 10). The second route would be to appreciate the AI’s 'personality'. In ordinary life, we can enjoy the personality traits of others, things like quick wit, poise, intellect, modesty. There seems little reason to deny that such enjoyment is aesthetic. Indeed it seems to be a good example of everyday aesthetics. It is plausible that this sort of personal aesthetics is the foundation of our appreciation of what we might call performance personalities. Appreciating the improvisational agility of a comedian or the gravitas of a great orator seems like an extension of our more personal aesthetic appreciation of each other. It is easy to see how these ideas can be applied to LLMs and appreciation. Rather than treating AI as participant, we meet them on a more level playing field, as interlocutor. It is also tempting to say that, if one has interacted with many models, they have different personalities. OpenAI’s GPT‑4.5, for example, was sold as having a better personality than its predecessor. Indeed, after OpenAI replaced GPT‑4o with GPT‑5 in August 2025, user backlash over GPT‑5’s tone and 4o’s perceived ‘feel’ led OpenAI to reinstate GPT‑4o for Plus users and to adjust its model‑retirement policy. Reuters, The Verge. Also in August, fans organised a mock ‘funeral’ after Anthropic retired Claude 3 Sonnet on 21 July 2025 (WIRED). 3 Appreciating something for what it in fact is In the rest of this paper we will set about showing why treating LLMs as (quasi) agents is not a satisfactory way of aesthetically appreciating them, and propose an alternative account on which appreciation of LLMs is modelled on 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: 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 a work of art 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. This mirrors the thermometer contrast: mercury‑in‑glass versus electronic sensing realise the same function under different mechanisms. In §4, we apply this two‑step schema to large language models: first fix what they are under the artefact reading; then use purpose‑and‑mechanism knowledge to discipline how they are to be perceived under that category. 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). In what follows, the same method is applied without deviation: the kind is fixed, the relevant function and realisation are stated, the boundary is delimited, and the acts of aspection are derived accordingly. 4. What LLMs in fact are The following passage from Carlson provides three ideas that will be useful as this paper progresses: \[add Something more robust here. The idea on the table is that we are going to draw heavily on Carlson’s “aesthetics to the environment” for the rest of this paper. Not only are we going to say that some of his ideas about agent motivation and related approaches are perhaps fundamentally flawed; it will also provide us with material for our positive account. That is, we should apply an environmental aesthetics to the aesthetics of LLMs, specifically\] 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) In this section, we focus on a general principle that Carlson thinks grounds all types of aesthetic appreciation: to appreciate something appropriately, we should appreciate it as what it in fact is. aesthetic appreciation of anything, be it people or pets, farmyards or neighborhoods, shoes or shopping malls, appreciation must be centered on and driven by the real nature of the object of appreciation itself.1 Carlson considers this “a blueprint for aesthetic appreciation in general” (Carlson 2000, 12. Emphasis ours). At its core, the position rejects both formalism and radical subjectivism. Against the former, which restricts appreciation to sensory properties abstracted from context and knowledge, Carlson holds that even basic forms are not properly appreciated without understanding what they are. Against the latter, he argues that aesthetic response is constrained by what the object is – by its real nature – rather than by personal preference (Carlson 2000, 12). The principle operates via what Carlson – following Paul Ziff – calls “acts of aspection”, the different ways we attend to and appreciate objects. With artworks, we must know not only that something is a painting but what kind of painting it is. Carlson cites Ziff’s examples: “I survey a Tintoretto, while I scan an H. Bosch… look for light in a Claude, for colour in a Bonnard, for contoured volume in a Signorelli.” The thought is that, in knowing the type, we know what and how to appreciate (Carlson 2000). For artworks, this knowledge is readily available because, as Carlson notes, “Works of art are our own creations; it is for this reason that we know what is and what is not a part of a work, which of its aspects are of aesthetic significance, and how to appreciate them.” We attend to the piano’s sound rather than the coughing that interrupts it, recognise where a painting ends at its frame, and look at paintings rather than listen to them. This is built into what it is to be a painting, a symphony, or a sculpture (Carlson 2000). With nature and other non‑art objects, matters are different. Natural environments, unlike artworks, “typically are not the products of designers and typically have no design. Rather they come about ‘naturally’; they change, grow, and develop by means of natural processes” (Carlson 2000). Appropriate appreciation therefore draws not on art‑historical knowledge but on knowledge of natural processes and environmental systems: The fact that nature is natural – not our creation – does not mean, however, that we must be without knowledge of it. Natural objects are such that we can discover things about them that are independent of any involvement by us in their creation… This knowledge, essentially common‑sense/scientific knowledge, seems to me the only viable candidate for playing the role concerning the appreciation of nature that our knowledge of types of art, artistic traditions, and the like plays concerning the appreciation of art. (Carlson 2000) Carlson contrasts modes of attention appropriate to different environments: “We must survey a prairie environment, looking at the subtle contours of the land, feeling the wind across the open space, and smelling the mix of prairie grasses and flowers; but such an act of aspection has little place in a dense forest environment. There we examine and scrutinise, inspecting the detail of the forest floor, listening for the sounds of birds, and smelling for the scent of spruce and pine” (Carlson 2000, 119). The principle that we ought to appreciate things as what they are thereby rejects both restriction to form and unconstrained relativism. Different kinds – natural environments, architectural works, agricultural landscapes, artefacts – make different demands on attention, and appropriate appreciation employs knowledge relevant to the kind in view. On this understanding, the appreciator’s role is active yet disciplined by the object’s nature (Carlson 2000). Carlson’s approach begins with kind‑fixing. Appropriate appreciation requires that we attend to an object “as what it in fact is” and “in light of our knowledge of what it is” (Carlson 2000, 5). In the middle domain between pristine nature and pure art, that knowledge is function‑centred: designed 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” (Carlson 2000, 133–134). In short, for designed artefacts, their natures are fixed by \*function and the mode of its realisation\* (Carlson 2000, 133–134; cf. 189). Applying this to large language models: on Carlson’s approach, the kind is fixed by what the artefact is made to do \*and\* how that purpose is realised. An LLM is made to continue token sequences learned from tokenised corpora; at use it selects the next token given the preceding context. Nothing in this essence requires that the tokens be human words: the relevant “language” is any learned code; English is a frequent but non‑essential instance of the token‑sequence the artefact continues.  How this purpose is realised belongs to what the object is. Text is tokenised into subword units and mapped to vectors with positional information; during training the model reduces next-token error on sequences, thereby learning a conditional continuation rule; at use it applies self-attention over a bounded context window to compute a next-token distribution, a decoding policy selects one token, and the loop repeats; post-training alignment and interface choices bias which continuations are chosen without altering the underlying continuation function. A common objection says: the function is to produce human‑like dialogue. Carlson’s framework rebuts this. Dialogue is a \*use‑case‑level\* aim set by alignment and interface conventions; it is neither necessary nor sufficient for the artefact’s identity. Some deployments never present dialogue; some non‑LLM systems produce dialogue. By contrast, learned token continuation is both necessary to and characteristic of the kind across deployments. On Carlson’s account, that is the tighter essence claim: the purpose that organises correct appreciation is the continuation of token sequences under a learned predictor realised at inference, while dialogue is a contingent manifestation shaped by external aims. Following Carlson's principle that aesthetic appreciation must be informed by knowledge of what the object in fact is, we turn to fixing the nature of LLMs so that our attention can be disciplined by kind. In making this determination, we accept a shift in our descriptive framework from the natural sciences – which Carlson employs for natural environments – to computer science. This shift represents not a departure from Carlson's method but rather its direct application to a different domain of objects. We begin by establishing the broad category: an LLM is an engineered artefact. It is neither a person nor a natural object, neither a freestanding artwork nor an autonomous agent. Whilst its outputs can constitute artworks under certain conditions, the model itself remains a functional system designed for specific computational tasks. Correct appreciation should therefore track its function and origin rather than any projected persona or imagined interiority. At its core, the technical objective of an LLM is \*next-token prediction\*. The model learns to continue sequences by minimising predictive error across large text corpora. Through training, its parameters – or "weights" – are adjusted so that each subsequent token becomes less surprising given the preceding context. This objective fundamentally shapes both what the model learns and how it generalises to new inputs. Before the learning process begins, the training corpus undergoes \*tokenisation\*, whereby text is segmented into discrete units that may be complete words or sub-word fragments such as "un-" or "-tion". The model never manipulates ideas or concepts directly; rather, it operates on these token indices. Tokenisation thus establishes the grain of what the system can notice and reproduce – a constraint that matters for any appreciation of its handling of style, rhythm, or phrasing. The training process itself is \*self-supervised\* because the data supply their own supervisory signal: the correct answer for any prediction task is simply the following token in the sequence. Loss reduction emerges not from memorising specific sentences but from discovering regularities that effectively compress the data. These regularities encompass grammatical structures, collocational patterns, semantic associations, and narrative conventions – the full range of statistical dependencies present in natural language. Modern LLMs employ \*attention-based architectures\* that model dependencies across variable spans of context. The active computational state exists entirely within the context window presented at inference time. There are no persistent goals, no diachronic self, no memory beyond what fits in this window – unless external memory systems are explicitly added. Whilst increases in model capacity and window length affect performance characteristics, they do not alter the fundamental kind of system we are considering. After training completes, the model's weights become fixed, and it applies its learned predictive function through an \*autoregressive\* process. At each step, the model generates a probability distribution over possible next tokens; \*decoding policies\* then convert this probability mass into actual text, with different policies striking different balances between determinacy and variety. The precise behaviour of the system depends heavily on prompt design and the ordering of contextual information. Variation across multiple runs with identical inputs often reflects these sampling choices rather than any underlying intention or creative agency. Contemporary LLMs undergo additional \*post-training alignment\* through techniques such as instruction tuning and preference optimisation. These processes reshape the model's responses toward greater follow-ability and policy compliance. What users encounter as guardrails, refusals, or safety measures are learned behaviours induced through this alignment rather than hard-coded rules. Similarly, the interfaces through which we interact with these models – including system prompts, external tools, and retrieval augmentation – mediate our inputs and the model's outputs without fundamentally altering the trained conditional model beneath. Phenomena such as hallucination, sycophancy, and extreme sensitivity to prompt phrasing emerge as expected outcomes of the training objective and data distribution rather than as bugs or personality quirks. When contexts depart substantially from training regimes – a phenomenon known as \*distribution shift\* – performance degrades in predictable ways. These limitations should be understood as properties of a conditional probability model rather than as character traits of an agent. To mis-categorise the system encourages over-reading of its outputs as expressions of intention, emotion, or belief. With the nature of LLMs thus established, we can identify appropriate \*acts of aspection\* – to use Carlson's term – suited to their appreciation. Where one might survey a prairie environment or scrutinise a forest floor, with LLMs we might \*survey\* their affordance ranges by systematically varying tasks, prompts, and decoding policies. We might \*scrutinise\* their local behaviour under fixed conditions to assess calibration, robustness, and consistency. These modes of attention replace folk-psychological projection with observation disciplined by the system's actual nature. On this understanding, the essence of an LLM for appreciative purposes is fixed by its computational role, training history, inference regime, and alignment constraints. Vendor identity, exact parameter count within a broad range, and conversational persona are not essential properties. What appears as "personality" represents a stable response profile induced by the interaction of training data, alignment procedures, and system prompts – nothing more, nothing less. Whilst agent language may persist as convenient shorthand in everyday interaction, aesthetic evaluation should answer to what the system in fact is: a sophisticated conditional model trained to predict text. This technical primer equips us to appreciate LLMs according to Carlson's blueprint, employing categories and criteria appropriate to engineered artefacts rather than those borrowed from interpersonal aesthetics. In the following section, we shall apply these acts of aspection to concrete cases, developing criteria that might include legibility of affordances, responsiveness to contextual framing, and stability under perturbation. The goal throughout remains disciplined appreciation grounded in accurate categorisation rather than anthropomorphic projection. Reading documents I’ve got several uploaded PDFs, including Carlson’s book and others like "No minds but signs" and "A note on 'semiotic physics'." The file search tool tells me to use citations but not necessarily deep ones right now. I might cite the Cross document for general content if needed. I’ll check for the "Style Guide" referenced in the instructions. It’s clear now that the "under a letter" reference likely points to the style guide in the project instructions. I can ask for confirmation on that. As for the documents, I have everything I need, including drafts and PDFs on semiotic physics. The user’s instructions avoid open-web browsing, so I’ll avoid that. Since this is a drafting task, I should avoid web browsing and rely on file search. I need to keep Introduction and Sections 1–3 largely as they are with light edits, and add a succinct paragraph summarizing the paper’s structure. I’ll focus on confirming my understanding, listing open questions, and proposing author-date citations with pages for quotes, and verifying terminology usage of natura vs machina. Plus, I’ll confirm the word count target and section budgets. For now, I just need to confirm my understanding and list any open questions or missing files. I'll paraphrase the task and be thorough with the questions, without adding unnecessary details or emotional language. I don’t need citations in the confirmation step, but I’ll note the files available, like "Cross 2024" and "Simulators by Janus." To confirm, we’ll avoid quoting and citations for now. The steps are clear: maintain British spelling, confirm file names, placeholders, and preferred quoting style (with comma for year and page). We’ll ask about section numbering, quotes in Section 5, and terminology for "machina naturans/naturata." We’ll clarify structure for Sections 6 and 7 before proceeding. I need to confirm the phrasing for an incomplete section about chat episodes and generative environments. We'll also ask clarification about several elements like how to treat references in the "Agentive sections" and how to mark items pre-2024 news. Additionally, we need to check Chalmers' work, verify Carlson's quotes, and confirm the final word count target. Done Understood. You want a submission-ready analytic paper for Philosophical Studies/Noûs, keeping your Introduction and §§1–3 with only light copy-editing, inserting one structural-summary paragraph in the Introduction, and then writing new §§4–9 and a short Conclusion. Section headings must be top-level numerals only, no subsections, 8,000–10,000 words total, British spelling, author–date citations with page numbers for quotes, strict verification from the attached files only, and full adherence to your Style Guide. Process: confirm → preflight plan with word budget and citations to verify → draft §§4–6 → draft §§7–9 → conclusion and references. Open questions and missing items Materials and mapping - Confirm file mapping: is “draft of environment paper 25 Jul 2025” the “new draft”? Is “OLD DRAFT.txt” the “old draft” to mine for §5? - Carlson 2000: I will verify all quoted passages and page numbers from the attached PDF. Any preferred edition details for references? - Chalmers quote (“Language models are systems that assign probabilities…”) is not in the attached set as far as I can see. Provide the source PDF, or permit cautious paraphrase with \[CITE/VERIFY\] tag. - Do you want Schwitzgebel & Shevlin (2023) and Sebo (2023) kept as bare references without quotes, or should I replace with neutral paraphrase plus \[CITE/VERIFY\] given no PDFs attached? Citation style and formatting - Your preference is author–date with commas before pages, e.g., (2000, 6). Should I standardise all existing in-text citations in §§1–3 to this format? - References: Chicago author–date? Any journal-specific quirks to match? Edits within §§1–3 - The text includes editorial placeholders and bracketed notes (\[a\]–\[o\], \[VERIFY\], %%…%%, etc.). Keep as footnotes, convert to endnotes, or remove once resolved? - News claims about 2025 model retirements: delete, or retain with \[VERIFY\] as per your gate? Section-specific scope and terminology - §4 ban: confirm I should avoid “environment”, “naturans”, and “semiotic physics” terms entirely in §4. Acceptable neutral terms: “model”, “training objective”, “autoregressive generation”, “decoding”, “alignment”. - §4 second half: target is a clear problem for agentive views grounded in the token-prediction account. Any specific interlocutors to engage here beyond Cross, Anscomb, and fictionalism already in §§1–2? - §5: You want the Carlson p. 6 block quote repeated at the start of §5. Confirm repetition verbatim with corrected citation formatting. Spinoza: any preferred translations or editions to cite, or treat as standard philosophical shorthand without edition detail? - §6: “Agent-light” lens as an analogue to person-level sciences in nature. Scope limits: keep to psychology/cognitive science metaphors only, then critique with “Not minds but signs” and “Simulators”. Any other attached sources you want included here? - §7: Instruction line breaks mid-sentence (“determine how LLM chats evolve over time – that is, the …”). Supply the intended completion, or authorise me to complete it (“… the trajectory/dynamics of the exchange”). - §8: Semiotic physics depth. Use “No minds but signs”, “A note on ‘semiotic physics’”, and the prior chat as the backbone. Confirm no equations, no real-physics commitments, architecture-grounded only. - §9: Criteria list. Confirm inclusion of: affordance legibility, responsiveness to framing, stability/variety under perturbations, semiotic coherence and phase transitions; and an explicit comparison with mechanistic interpretability as a complementary lens. Terminology policy - Final policy for Latinisms and new terms: use *natura naturans*, *natura naturata*, *machina naturans*, *machina naturata* in italics at first mention, then plain thereafter. Accept? - Consistent use of “generative environment” from §5 onward. In §7 make explicit: models are not environments; chat episodes instantiate *generative environments*. Accept? - Capitalisation and hyphenation: “next-token prediction”, “post-training alignment”, “self-supervised”, “context window”. Accept these as house forms? Structure and budgets - Target 8,000–10,000 words. Any preference for the weight of §§8–9 relative to earlier sections? I will propose a per-section budget in the preflight plan. Examples and case material - Do you want worked micro-studies in §§7–9 (e.g., short controlled prompt perturbations) to exemplify the criteria, or keep it fully theoretical? Output artefacts - References list: include only works actually cited in the final text, drawn from attachments unless \[CITE/VERIFY\] is marked. Accept? If all of the above is correct, reply “OK” with any clarifications. Then I will send the preflight plan with word budgets, key claims, verification targets, and final terminology policy. Chalmers and LLMs.pdf PDF 1\. "is “draft of environment paper 25 Jul 2025” the “new draft”?" –NO New Draft was copied in at the bottom of the previous message. I have copied it into this message as well to avoid any confusion. This is the version which i want you to finish. OLD DRAFT.txt is "old draft" It begins as follows: Introduction Part of the reason paintings, novels, or films merit aesthetic appreciation is that they are the result of their makers’ efforts. A painter, writer, or director may spend months or years honing a work through sustained attention and revision; viewers or readers often experience these works precisely as the pr...." 2. "Chalmers quote (“Language models are systems that assign probabilities…”) is not in the attached set as far as I can see. Provide the source PDF, or permit cautious paraphrase with \[CITE/VERIFY\] tag." – I have attached a pdf. 3. "Do you want Schwitzgebel & Shevlin (2023) and Sebo (2023) kept as bare references without quotes, or should I replace with neutral paraphrase plus \[CITE/VERIFY\] given no PDFs attached?" they should be kept as bare references. the same goes for any others like this you find. new draft: the environmental aesthetics of generative AI Introduction 1. AI as (quasi) Agents 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 (we shall see some examples of this in the next section). This section will sketch three different flavours of this view. 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. Cross (2024), writing on AI art, exemplifies this approach when he argues that prompting, iteration, and sampling structure a dialogue where value lies in the interaction itself. "By adjusting inputs, iterating, and sampling," he writes, "an AI artist is engaged in a process of mapping – and perhaps interrogating – the way that the algorithm sees and understands" (Cross 2024, 7–8), although he concedes that "the analogy with performance art isn't a perfect one" (Cross 2024, 9). Anscomb could also be thought of as endorsing a pseudo-agent view. She denies literal mentality and creativity in present systems – "we are not yet at the stage where an AI can formulate intentions … I argue that AI agents cannot be artistically creative" – and adopts a procedural label for "AI agent" as "a self-contained ('autonomous') procedure". Yet she also holds that an AI "may work iteratively without human intervention to non-accidentally generate the formal features of an image" and that it can merit "some share of production credit", which positions her as treating AI as a pseudo-agent, an as-if collaborator rather than a mind-bearing agent. Nonetheless, her vocabulary pressures toward minimal literalism: AICAN is said to be "able to self-assess these products", and "an AI agent arguably deserves credit for its contribution to the production of a work qua art" – locutions typically reserved for bearers of credit rather than mere instruments. 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. This position has been developed by Mallory (2023), Krueger & Roberts (2024), and Krueger & Osler (2022), whilst Friend & Goffin (2025) discuss the view without endorsing it. As Friend & Goffin observe, prop-oriented make-believe explains ordinary exchanges with Alexa and ChatGPT, where users converse "as when we say 'thank you' … while knowing that there is no real agent producing the replies" (Friend & Goffin 2025, 9). By contrast, content-oriented make-believe clarifies richer, empathetic uses where "it is the interaction itself that matters", with users imaginatively treating the chatbot as a person within the ongoing exchange. %%Succinct one paragraph summary of the structure of the paper. %% 2. Appreciating LLMs as agents People, and perhaps other conscious beings (see Marchetti?), seem like a sort of thing that can be aesthetically appreciated. Most importantly for our interests here, this appreciation is not limited to physical beauty. If LLMs are a type of agent or quasi-agent, this might well allow for them to be aesthetically appreciated in a somewhat similar manner. In this section, we look at two ways in which this idea might be elaborated. The first route treats the system as a participant within an artist-structured interaction. On this approach, the evaluative focus shifts away from freestanding outputs toward what the duo does together through prompting, iterating, and mapping the system’s way of seeing. As Cross observes: By way of their selection of prompts, they elicit a sort of "participation" on the part of the AI … This participation allows them to interrogate the algorithm's latent space. (Cross 2024, 8) According to Cross, the interactional performance is primary; the images or videos are often documentation of that practice rather than the locus of value. He argues that "what is centrally important … is that it is less centrally focused on the output of the AI image generator. Instead, what matters is the interaction between the artist and the algorithm" (Cross 2024, 8). The product can thus function as documentation of the interactional process. Cross calls this the “exploration paradigm”: artists “relate to AI as a participant” and “create a space for interaction … by way of their prompts,” shifting appreciation “towards the artist’s interaction with the AI and the way in which the artist structures this interaction.” The generated images “function as a kind of documentation or evidence” of that exploration rather than the locus of value (Cross 2024, abstract; 9). The agent‑based framing is explicit: the work lies in a temporally extended dialogue in which the artist elicits “participation” from the model to “interrogate the algorithm’s latent space,” making the algorithm’s ways of seeing and representing the object of appreciation (Cross 2024, 8–9). Cross presents this as an as‑if stance and concedes that the performance‑art analogy “isn’t a perfect one” (Cross 2024, 10). The second route would be to appreciate the AI’s 'personality'. In ordinary life, we can enjoy the personality traits of others, things like quick wit, poise, intellect, modesty. There seems little reason to deny that such enjoyment is aesthetic. Indeed it seems to be a good example of everyday aesthetics. It is plausible that this sort of personal aesthetics is the foundation of our appreciation of what we might call performance personalities. Appreciating the improvisational agility of a comedian or the gravitas of a great orator seems like an extension of our more personal aesthetic appreciation of each other. It is easy to see how these ideas can be applied to LLMs and appreciation. Rather than treating AI as participant, we meet them on a more level playing field, as interlocutor. It is also tempting to say that, if one has interacted with many models, they have different personalities. OpenAI’s GPT‑4.5, for example, was sold as having a better personality than its predecessor. Indeed, after OpenAI replaced GPT‑4o with GPT‑5 in August 2025, user backlash over GPT‑5’s tone and 4o’s perceived ‘feel’ led OpenAI to reinstate GPT‑4o for Plus users and to adjust its model‑retirement policy. Reuters, The Verge. Also in August, fans organised a mock ‘funeral’ after Anthropic retired Claude 3 Sonnet on 21 July 2025 (WIRED). 3 Appreciating something for what it in fact is In the rest of this paper we will set about showing why treating LLMs as (quasi) agents is not a satisfactory way of aesthetically appreciating them, and propose an alternative account on which appreciation of LLMs is modelled on 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: 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 a work of art 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. This mirrors the thermometer contrast: mercury‑in‑glass versus electronic sensing realise the same function under different mechanisms. In §4, we apply this two‑step schema to large language models: first fix what they are under the artefact reading; then use purpose‑and‑mechanism knowledge to discipline how they are to be perceived under that category. 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). In what follows, the same method is applied without deviation: the kind is fixed, the relevant function and realisation are stated, the boundary is delimited, and the acts of aspection are derived accordingly. 4. What LLMs in fact are The following passage from Carlson provides three ideas that will be useful as this paper progresses: \[add Something more robust here. The idea on the table is that we are going to draw heavily on Carlson’s “aesthetics to the environment” for the rest of this paper. Not only are we going to say that some of his ideas about agent motivation and related approaches are perhaps fundamentally flawed; it will also provide us with material for our positive account. That is, we should apply an environmental aesthetics to the aesthetics of LLMs, specifically\] 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) In this section, we focus on a general principle that Carlson thinks grounds all types of aesthetic appreciation: to appreciate something appropriately, we should appreciate it as what it in fact is. aesthetic appreciation of anything, be it people or pets, farmyards or neighborhoods, shoes or shopping malls, appreciation must be centered on and driven by the real nature of the object of appreciation itself.1 Carlson considers this “a blueprint for aesthetic appreciation in general” (Carlson 2000, 12. Emphasis ours). At its core, the position rejects both formalism and radical subjectivism. Against the former, which restricts appreciation to sensory properties abstracted from context and knowledge, Carlson holds that even basic forms are not properly appreciated without understanding what they are. Against the latter, he argues that aesthetic response is constrained by what the object is – by its real nature – rather than by personal preference (Carlson 2000, 12). The principle operates via what Carlson – following Paul Ziff – calls “acts of aspection”, the different ways we attend to and appreciate objects. With artworks, we must know not only that something is a painting but what kind of painting it is. Carlson cites Ziff’s examples: “I survey a Tintoretto, while I scan an H. Bosch… look for light in a Claude, for colour in a Bonnard, for contoured volume in a Signorelli.” The thought is that, in knowing the type, we know what and how to appreciate (Carlson 2000). For artworks, this knowledge is readily available because, as Carlson notes, “Works of art are our own creations; it is for this reason that we know what is and what is not a part of a work, which of its aspects are of aesthetic significance, and how to appreciate them.” We attend to the piano’s sound rather than the coughing that interrupts it, recognise where a painting ends at its frame, and look at paintings rather than listen to them. This is built into what it is to be a painting, a symphony, or a sculpture (Carlson 2000). With nature and other non‑art objects, matters are different. Natural environments, unlike artworks, “typically are not the products of designers and typically have no design. Rather they come about ‘naturally’; they change, grow, and develop by means of natural processes” (Carlson 2000). Appropriate appreciation therefore draws not on art‑historical knowledge but on knowledge of natural processes and environmental systems: The fact that nature is natural – not our creation – does not mean, however, that we must be without knowledge of it. Natural objects are such that we can discover things about them that are independent of any involvement by us in their creation… This knowledge, essentially common‑sense/scientific knowledge, seems to me the only viable candidate for playing the role concerning the appreciation of nature that our knowledge of types of art, artistic traditions, and the like plays concerning the appreciation of art. (Carlson 2000) Carlson contrasts modes of attention appropriate to different environments: “We must survey a prairie environment, looking at the subtle contours of the land, feeling the wind across the open space, and smelling the mix of prairie grasses and flowers; but such an act of aspection has little place in a dense forest environment. There we examine and scrutinise, inspecting the detail of the forest floor, listening for the sounds of birds, and smelling for the scent of spruce and pine” (Carlson 2000, 119). The principle that we ought to appreciate things as what they are thereby rejects both restriction to form and unconstrained relativism. Different kinds – natural environments, architectural works, agricultural landscapes, artefacts – make different demands on attention, and appropriate appreciation employs knowledge relevant to the kind in view. On this understanding, the appreciator’s role is active yet disciplined by the object’s nature (Carlson 2000). Carlson’s approach begins with kind‑fixing. Appropriate appreciation requires that we attend to an object “as what it in fact is” and “in light of our knowledge of what it is” (Carlson 2000, 5). In the middle domain between pristine nature and pure art, that knowledge is function‑centred: designed 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” (Carlson 2000, 133–134). In short, for designed artefacts, their natures are fixed by \*function and the mode of its realisation\* (Carlson 2000, 133–134; cf. 189). Applying this to large language models: on Carlson’s approach, the kind is fixed by what the artefact is made to do \*and\* how that purpose is realised. An LLM is made to continue token sequences learned from tokenised corpora; at use it selects the next token given the preceding context. Nothing in this essence requires that the tokens be human words: the relevant “language” is any learned code; English is a frequent but non‑essential instance of the token‑sequence the artefact continues.  How this purpose is realised belongs to what the object is. Text is tokenised into subword units and mapped to vectors with positional information; during training the model reduces next-token error on sequences, thereby learning a conditional continuation rule; at use it applies self-attention over a bounded context window to compute a next-token distribution, a decoding policy selects one token, and the loop repeats; post-training alignment and interface choices bias which continuations are chosen without altering the underlying continuation function. A common objection says: the function is to produce human‑like dialogue. Carlson’s framework rebuts this. Dialogue is a \*use‑case‑level\* aim set by alignment and interface conventions; it is neither necessary nor sufficient for the artefact’s identity. Some deployments never present dialogue; some non‑LLM systems produce dialogue. By contrast, learned token continuation is both necessary to and characteristic of the kind across deployments. On Carlson’s account, that is the tighter essence claim: the purpose that organises correct appreciation is the continuation of token sequences under a learned predictor realised at inference, while dialogue is a contingent manifestation shaped by external aims. Following Carlson's principle that aesthetic appreciation must be informed by knowledge of what the object in fact is, we turn to fixing the nature of LLMs so that our attention can be disciplined by kind. In making this determination, we accept a shift in our descriptive framework from the natural sciences – which Carlson employs for natural environments – to computer science. This shift represents not a departure from Carlson's method but rather its direct application to a different domain of objects. We begin by establishing the broad category: an LLM is an engineered artefact. It is neither a person nor a natural object, neither a freestanding artwork nor an autonomous agent. Whilst its outputs can constitute artworks under certain conditions, the model itself remains a functional system designed for specific computational tasks. Correct appreciation should therefore track its function and origin rather than any projected persona or imagined interiority. At its core, the technical objective of an LLM is \*next-token prediction\*. The model learns to continue sequences by minimising predictive error across large text corpora. Through training, its parameters – or "weights" – are adjusted so that each subsequent token becomes less surprising given the preceding context. This objective fundamentally shapes both what the model learns and how it generalises to new inputs. Before the learning process begins, the training corpus undergoes \*tokenisation\*, whereby text is segmented into discrete units that may be complete words or sub-word fragments such as "un-" or "-tion". The model never manipulates ideas or concepts directly; rather, it operates on these token indices. Tokenisation thus establishes the grain of what the system can notice and reproduce – a constraint that matters for any appreciation of its handling of style, rhythm, or phrasing. The training process itself is \*self-supervised\* because the data supply their own supervisory signal: the correct answer for any prediction task is simply the following token in the sequence. Loss reduction emerges not from memorising specific sentences but from discovering regularities that effectively compress the data. These regularities encompass grammatical structures, collocational patterns, semantic associations, and narrative conventions – the full range of statistical dependencies present in natural language. Modern LLMs employ \*attention-based architectures\* that model dependencies across variable spans of context. The active computational state exists entirely within the context window presented at inference time. There are no persistent goals, no diachronic self, no memory beyond what fits in this window – unless external memory systems are explicitly added. Whilst increases in model capacity and window length affect performance characteristics, they do not alter the fundamental kind of system we are considering. After training completes, the model's weights become fixed, and it applies its learned predictive function through an \*autoregressive\* process. At each step, the model generates a probability distribution over possible next tokens; \*decoding policies\* then convert this probability mass into actual text, with different policies striking different balances between determinacy and variety. The precise behaviour of the system depends heavily on prompt design and the ordering of contextual information. Variation across multiple runs with identical inputs often reflects these sampling choices rather than any underlying intention or creative agency. Contemporary LLMs undergo additional \*post-training alignment\* through techniques such as instruction tuning and preference optimisation. These processes reshape the model's responses toward greater follow-ability and policy compliance. What users encounter as guardrails, refusals, or safety measures are learned behaviours induced through this alignment rather than hard-coded rules. Similarly, the interfaces through which we interact with these models – including system prompts, external tools, and retrieval augmentation – mediate our inputs and the model's outputs without fundamentally altering the trained conditional model beneath. Phenomena such as hallucination, sycophancy, and extreme sensitivity to prompt phrasing emerge as expected outcomes of the training objective and data distribution rather than as bugs or personality quirks. When contexts depart substantially from training regimes – a phenomenon known as \*distribution shift\* – performance degrades in predictable ways. These limitations should be understood as properties of a conditional probability model rather than as character traits of an agent. To mis-categorise the system encourages over-reading of its outputs as expressions of intention, emotion, or belief. With the nature of LLMs thus established, we can identify appropriate \*acts of aspection\* – to use Carlson's term – suited to their appreciation. Where one might survey a prairie environment or scrutinise a forest floor, with LLMs we might \*survey\* their affordance ranges by systematically varying tasks, prompts, and decoding policies. We might \*scrutinise\* their local behaviour under fixed conditions to assess calibration, robustness, and consistency. These modes of attention replace folk-psychological projection with observation disciplined by the system's actual nature. On this understanding, the essence of an LLM for appreciative purposes is fixed by its computational role, training history, inference regime, and alignment constraints. Vendor identity, exact parameter count within a broad range, and conversational persona are not essential properties. What appears as "personality" represents a stable response profile induced by the interaction of training data, alignment procedures, and system prompts – nothing more, nothing less. Whilst agent language may persist as convenient shorthand in everyday interaction, aesthetic evaluation should answer to what the system in fact is: a sophisticated conditional model trained to predict text. This technical primer equips us to appreciate LLMs according to Carlson's blueprint, employing categories and criteria appropriate to engineered artefacts rather than those borrowed from interpersonal aesthetics. In the following section, we shall apply these acts of aspection to concrete cases, developing criteria that might include legibility of affordances, responsiveness to contextual framing, and stability under perturbation. The goal throughout remains disciplined appreciation grounded in accurate categorisation rather than anthropomorphic projection. 4. "Your preference is author–date with commas before pages, e.g., (2000, 6). Should I standardise all existing in-text citations in §§1–3 to this format?" No. The style should this(Author's surname YEAR, p. x) or (Author's surname YEAR) as required. 5. "The text includes editorial placeholders and bracketed notes (\[a\]–\[o\], \[VERIFY\], %%…%%, etc.). Keep as footnotes, convert to endnotes, or remove once resolved?" replace all place holders and bracketed notes with text where appropriate, or footnotes where appropriate, if unclear leave as is. 6. "§4 ban: confirm I should avoid “environment”, “naturans”, and “semiotic physics” terms entirely in §4. Acceptable neutral terms: “model”, “training objective”, “autoregressive generation”, “decoding”, “alignment”." yep. good. 7. "§4 second half: target is a clear problem for agentive views grounded in the token-prediction account. Any specific interlocutors to engage here beyond Cross, Anscomb, and fictionalism already in §§1–2?" Nope, and you don't need to go into too much detail on these individuals. the argument to be made at this point seems general enough to cause problems for all of these views –LLMs are not literally agent like, so against literalism, and fictionalism and psudo-agent are by their very nature not appreciating LLMs for what they in fact are (note that this does NOT conflict or overlap with suggestion dealt with in a later section as to whether agent views make good 'lights' for appreciating llms. 8. "§5: You want the Carlson p. 6 block quote repeated at the start of §5. Confirm repetition verbatim with corrected citation formatting. "I confirm" 9. Spinoza: any preferred translations or editions to cite, or treat as standard philosophical shorthand without edition detail?" Don’t worry too much about any missing citations here—just put placeholders in. I’ll deal with that at the end. Thanks. 10. "§6: “Agent-light” lens as an analogue to person-level sciences in nature. Scope limits: keep to psychology/cognitive science metaphors only, then critique with “Not minds but signs” and “Simulators”. Any other attached sources you want included here?" they are not metaphors, they are a potential light, a potential means of understanding the thing itself, remember to keep to this framework from which this idea comes from closely and throughout the paper. mainly those two of the sources you have been given. I remember 'Simulators' in particular has some interesting arguments against (I consider what Janus calls oracle, genie etc. views just to be species of agent view. The point is not just to mention these arguments against agent views, but frame them as reasons why agent views are not good lights for understanding LLMs, in that they mischaracterise the particular sorts of machina naturans at work. there is no level or aspects of an LLMs systems machina that is properly captured by the agentive frame. –i think janus is best for this, double check that the criticisms in the the signs not minds paper could be used here in the same sort of way. 11. "§8: Semiotic physics depth. Use “No minds but signs”, “A note on ‘semiotic physics’”, and the prior chat as the backbone. Confirm no equations, no real-physics commitments, architecture-grounded only." yep. the important thing is that it has to capture the complexities of LLMs in such a way as to show how they strike us as agentive or agent-like (refer back to the relevant idea in section 1 when the agent views are introduced). as mentioned, this should be a reasonably substantial section. "§9: Criteria list. Confirm inclusion of: affordance legibility, responsiveness to framing, stability/variety under perturbations, semiotic coherence and phase transitions; and an explicit comparison with mechanistic interpretability as a complementary lens." I don’t know what affordance legibility means. Responsiveness to framing is clearly couched in the terminology used in the relevant part of the paper. I quite like stability and variety under perturbations. I like their semiotic coherence and phase transitions. Yes, I would like an explicit comparison with mechanistic interpretability as a complementary lens, but it should only really be footnote-length. We should say, “There’s this other discipline over here,” and then explain why mechanistic interpretability can be considered a light in the Carlson sense, but leave it at that—more or less. It should be a fairly brief note somewhere, a little footnote or endnote.