# [[The Environmental Aesthetics of Generative AI]] date:: 24 Feb 2025 type:: llm text ### I. Introduction Contemporary research in [[artificial intelligence]] has produced a new class of [[generative systems]]—large [[language models]] (LLMs), image generators, and related architectures—that challenge longstanding assumptions about technological artifacts. Traditionally, we have evaluated artifacts by referencing their apparent function or comparing them to clear, human-devised aims. A chair is designed for sitting, a camera for taking photographs. Yet generative AI resists neat functional categorisation: large [[language models]] produce text across an extraordinary range of domains and styles, and image generators can create outputs in any number of visual genres. In light of these flexible, open-ended capabilities, [[aesthetic appreciation]] of AI faces a conceptual question: _By what standard do we judge systems whose “function” is so fluid?_ Function-based approaches, which normally tie aesthetic virtue to an artefact’s fulfilling its purpose with elegance or efficiency, struggle here. One might say a large language model’s purpose is “to generate text,” but the possible uses of that text are so diverse that it becomes nearly impossible to define a stable “ideal form” the system must embody. Additionally, the internal workings of generative AI look less like straightforward engineering and more like processes of _growth_ or _cultivation_; developers design an initial network architecture and training objective, then watch as the system “organically” arrives at emergent solutions through iterative data exposure. These issues invite a more comprehensive aesthetic framework—one that can accommodate dynamic, evolving systems rather than neat functional tools. [[Allen Carlson]]’s [[environmental aesthetics]] suggests precisely such a framework. Carlson argues that nature cannot be appreciated simply as a static tableau. Instead, one must grasp the active processes—geological, biological, ecological—that give rise to what we see, recognizing nature’s status as an evolving whole. Spinoza’s concept of _natura naturans_ accentuates these ongoing generative forces, portraying nature not as a static repository (_natura naturata_) but as a continuous, self-organising activity. This emphasis on generative processes and dynamic structures, rather than fixed designs, resonates with how large neural networks are created and refined. Yet fully articulating this analogy requires engaging with the inner landscape of modern AI—namely [[the idea]] of _latent spaces_ or _manifolds_ that undergird generative models. Much as [[environmental aesthetics]] celebrates knowledge of ecological relationships and systemic processes, understanding [[the manifold]] structure in generative AI could be key to viewing these systems as “environments” in which data points, transformations, and emergent features interact. By comparing these learned manifolds to [[the forces]] shaping natural environments, we can propose that to appreciate generative AI aesthetically, we should treat them like _dynamic ecologies_ or _environments_, shaped by hidden high-dimensional geometries analogous to the myriad energies at work in nature. I begin by summarizing the main shortcoming of function-based aesthetic theories for generative AI (Section II). Next, I [[introduce Carlson]]’s [[environmental aesthetics]] (Section III) and connect it to [[the idea]] that these AI systems “grow” rather than merely follow fixed programs (Section IV). The core of the paper (Section V) elaborates how _latent spaces_ and the so-called manifold perspective supply an account of the “forces” acting upon generative AI. Section VI outlines three distinct ways we might appreciate generative models from this environmental standpoint. Section VII addresses potential implications and objections, and Section VIII concludes by affirming that environmental aesthetics, enriched by an appreciation of latent space, helps us better conceptualize and aesthetically engage generative AI. --- ### II. A Critique of Function-Based Approaches Many standard aesthetic frameworks begin by noting that artifacts have a purpose, and we frequently regard an artifact as aesthetically successful if it accomplishes that function elegantly or efficiently. In design philosophy, for instance, a well-crafted object is “fit for purpose,” which can ground the notion of functional beauty. Jane Forsey’s Kantian approach likewise insists that understanding an artifact’s intended concept or function underwrites aesthetic judgment about its form. Generative AI, however, does not fit easily into such a model. A large language model (LLM) can be used for creative writing, debugging code, tutoring in foreign languages, summarizing complex legal documents, or generating marketing material. Image generators (like DALL·E or Midjourney) produce an unlimited variety of visual styles, from abstract painting to photorealistic landscapes. At first glance, one might say “their function is producing text or images,” but that phrase is so broad it fails to capture any definitive “purpose” in the narrower sense that a chair is for sitting or a scalpel is for cutting. Additionally, LLMs and other generative systems can spontaneously reveal capabilities that even their developers did not predict. This runs counter to the idea that each artifact is meticulously designed with a set of stable goals in mind. Indeed, these systems are shaped through training on large corpora, developing internal patterns that reflect data-driven “solutions” to open-ended tasks. If a function-based theory tries to interpret them, it risks either oversimplifying (treating them as just text generators) or enumerating so many separate “functions” that the concept of a unifying design aim dissolves. In either case, it becomes difficult to articulate a stable sense of “fit for purpose.” Moreover, generative AI appears to be dynamic in a way typical artifacts are not. Even after deployment, these models might be fine-tuned, updated, or confronted with user prompts that yield unforeseen uses. The system’s function, in short, _is not pinned down at creation_ and may continue to evolve through novel user interactions. Hence the root of the difficulty: function-based aesthetics presupposes a relatively clear purpose for an object. Generative AI can seldom be pinned to one. If a single artifact’s “function” is simultaneously translation, image captioning, short-story writing, and mathematical problem-solving, any function-based aesthetic analysis feels strained. We need a framework that is more tolerant of emergent, multifaceted processes. --- ### III. Carlson’s Environmental Aesthetics: Key Points Allen Carlson’s environmental aesthetics offers a helpful template for seeing how dynamic systems might be appreciated outside the constraints of a single, neatly defined purpose. Carlson argues that to appreciate nature as _nature_, one must appreciate it as a dynamic environment with interrelated systems, shaped by forces such as biology, geology, or weather. By analogy, if we are to appreciate generative AI _as_ something distinct from a typical functional device, we might follow Carlson’s reasoning about environmental wholes. Carlson emphasizes that “natural objects possess what we might call an organic unity with their environments of creation,” meaning they are not isolated lumps but rather parts of intricate webs of cause and effect: > “They are a part of and have developed out of the elements of their environments by means of the forces at work within those environments.” This emphasis on the “forces” that shape natural objects is crucial. We cannot simply look at a mountain and call it beautiful without understanding the geological activity—tectonic uplift, erosion, volcanic processes—that formed it. For Carlson, appreciating nature requires understanding these forces and acknowledging that nature is not static but dynamically evolving. He continues: > “If to aesthetically appreciate art we must have knowledge of artistic traditions and styles within those traditions, then to aesthetically appreciate nature we must have knowledge of the different environments of nature and of the systems and elements within those environments.” Nature’s aesthetic interest, then, arises partly from seeing it in the context of its generative processes. When we identify the interplay between species in an ecosystem or the effect of wind and water on shaping a canyon, we bring to bear knowledge that illuminates nature’s coherence. Spinoza’s concept of _natura naturans_—the active, generative principle—likewise insists on understanding nature as continuously producing and reproducing itself, emphasizing dynamic creation rather than static final states. Carlson’s environmental model directly resonates with that notion, conceiving of nature as a domain of ceaseless activity and unfolding interconnections. In short, Carlson’s perspective makes two central claims: 1. **We must appreciate an environment as it truly is**—i.e., a dynamic, evolving system. 2. **We should attend to the relevant “forces”**—the processes and systems that give rise to what we observe. Transposing this approach to AI suggests we look for _analogous forces_ within generative models, especially those that cause them to develop certain distributions, behaviors, or emergent capacities. --- ### IV. Growing AI as an Environment Generative AI—particularly in the form of large neural networks—shares important qualities with Carlson’s notion of a natural environment. In conventional software engineering, one might meticulously code each behavior. But in machine learning, a system “learns” from data and reconfigures its internal parameters with minimal direct human instruction. This process feels more akin to _growth_ than design. As Dario Amodei (CEO of Anthropic) remarks: > “I think one useful way to think about neural networks is that we don’t program and we don’t make them. We kind of, we grow them. You know, we have these neural network architectures that we design and we have these loss objectives that we create. And the neural network architecture, it’s kind of like a scaffold that the circuits grow on... it starts off with some kind of random, you know, random things and it grows. And it’s almost like the objective that we train for is this light.” (Lex Fridman Podcast #452) This image of an AI “growing” stands in contrast to the typical artifact model. Amodei further notes: > “It’s very, very different from any kind of regular software engineering, because at the end of the day, we end up with this artifact that can do all these amazing things... write essays and translate and understand images.” He calls it an “artifact,” but it is difficult to treat this product as an ordinary tool with a single guiding function. One might say it is for text generation or classification, but it can _also_ be directed to new tasks—like a dynamic environment that keeps revealing new nooks and crannies. One result is that we see the system less as a clear, crisp machine and more as a partially autonomous domain in which intricate structures form, akin to the way a forest’s biodiversity develops over time. Carlson’s talk of “forces” in the environment neatly parallels the “forces” shaping AI: training data, backpropagation, objective functions, hyperparameters, alignment procedures, and so on. These push the developing network in certain directions, just as geology and climate push a region of land to assume particular geological or ecological characteristics. In both contexts, we get a dynamic, evolving structure that is not dictated by a single blueprint but emerges from interplay among many processes. A generative AI can come to exhibit surprising properties—such as zero-shot reasoning or creative coding suggestions—without these properties having been explicitly “programmed in.” Environmental processes in nature likewise produce novel ecosystems or unexpected species interactions. If, then, generative AI is best viewed as an environment shaped by many invisible “forces,” it becomes sensible to treat it in a manner akin to Carlson’s environmental aesthetics: we do not fixate on whether it is _optimally functional,_ but rather on _how_ it arises from internal “generative” interactions. To refine this approach further, we need to see how the manifold or latent-space perspective in modern machine learning helps us interpret these systems as “dynamic wholes.” --- ### V. Manifolds, Latent Space, and the “Forces” that Shape Generative AI Neural networks commonly learn internal representations that can be understood in terms of _manifolds_ or _latent spaces._ The manifold hypothesis states that natural high-dimensional data (like images, text, or sound) tend to lie on lower-dimensional surfaces embedded in that high-dimensional space. When a neural network is trained, part of what it does is discover or approximate these low-dimensional manifolds, capturing the statistical structure of the data. For generative AI specifically, a “latent space” is a domain—often just a large vector space—where each point encodes certain features or patterns. Consider how a generative image model, such as a Variational Autoencoder or a GAN (Generative Adversarial Network), transforms random noise into a coherent image by _locating_ that noise in a learned latent space. The model’s generator maps points in that space to realistic images. Similarly, large language models learn to produce the next token by updating hidden states that effectively locate a _context_ in a manifold of possible utterances. In more general terms, one might say that _each generative AI is an environment characterized by a particular manifold,_ shaped through training. The “forces” of training data distribution, gradient-based updates, and hyperparameters cause the model to carve out a specific “terrain” in its latent space. Once the model is trained, we can explore that terrain by sampling points or by feeding in prompts that take us into new regions of possibility. #### Manifold as Environment Carlson’s emphasis on processes that unify an environment—e.g., erosion shaping canyons, or climate patterns affecting ecosystems—can be analogized to the shaping of a model’s latent space. Instead of rainfall and geological uplift, we have gradie![[attachments/Screenshot 2025-02-24 at 13.10.47.png]]nt descent and data constraints, but the essential idea remains: an underlying “force” (or set of forces) sculpts a domain into a coherent, structured environment. This environment (the manifold) is not random or purely chaotic. It exhibits patterns, clusters, and pathways that correspond to meaningful forms—like the difference between “normal text” and “poetic text,” or between images of cats and dogs. Crucially, once trained, a generative system can be seen as a living map of relationships. Each dimension or sub-region of latent space indicates a style, semantic cluster, or functional shift. One might explore these spaces by interpolating between points (watching images morph smoothly between states, or texts drift from one topic to another). This is not unlike exploring a forest’s transitions from one biome to another, albeit in a computational sense. To appreciate this _environment_ is to marvel at how an LLM’s manifold organizes language in high-dimensional form: synonyms cluster together, patterns of grammar reflect learned “valleys,” transitions to new topics might correspond to stepping across manifold boundaries, and so forth. The manifold perspective thus turns the generative model into an interpretable domain of emergent structure, shaped by the “forces” of training. #### Why This Matters Aesthetically A conventional perspective might focus on the model’s success at generating coherent text or images, but an environmental perspective looks at _how it came to do so_ and _how that structure forms an evolving or navigable space._ In that sense, an aesthetic appreciation reminiscent of Carlson’s arises when we realize that generative AI is not merely a mechanical input–output system. It has an internal topography—one developed through iterative shaping, much as nature’s landscapes develop through iterative geological or ecological processes. When we see an unusual AI-generated picture that transitions between realistic photography and cartoonish art, or a piece of text that leaps across literary genres, we can trace that phenomenon to the model’s manifold. This is akin to seeing a plant that spans multiple ecological niches and appreciating how the environment’s microclimates cause morphological diversity. The analogy is not perfect, since AI does not stand outside human design, yet it captures the idea that emergent patterns exceed a single, direct intention from a developer. The manifold is discovered by the model, within the constraints provided, just as environmental systems discover certain stable niches within climate constraints. Finally, the manifold perspective underscores that _multidimensional relational structure_ can be aesthetically intriguing in itself. Just as ecologists might admire how a rainforest fosters thousands of species in a complex web of mutual dependencies, an observer of generative AI might admire how a model spontaneously “knits together” countless contexts. In both domains, a sense of wonder arises from the recognition that underlying processes—be they gradient descent or ecological selection—coordinate enormous complexity into something that feels coherent, if not always fully predictable. --- ### VI. Three Ways of Appreciating Generative AI in an Environmental Lens With this manifold-based reinterpretation in mind, we can identify three complementary angles from which to appreciate generative AI aesthetically, loosely mirroring how Carlson identifies varied modes of appreciating nature in environmental aesthetics. 1. **Appreciation of the “Shaped Manifold” Itself** First, one can stand back and appreciate the latent space as a dynamic environment in which data converge. The focus here is on the system’s capacity to learn a complex geometry that unifies wide-ranging inputs. Paralleling the environmental approach to nature, this mode of appreciation highlights the _coherence_ and _organization_ of the AI’s internal world. One might be impressed that the system clusters similar concepts together, or that it can smoothly “move” from one style or topic to another by traversing the learned manifold. This recalls how environmental aesthetics encourages us to see a forest or canyon in light of the geological or biological forces that created it. Instead of ignoring the “invisible structure,” we embrace it as essential to aesthetic appreciation. 2. **Appreciation of the Outputs as Expressions of Underlying Forces** The second vantage focuses on particular outputs—texts, images, or other creative artifacts generated by the model—and views them as manifestations of the manifold’s learned patterns. Instead of seeing an AI-generated poem simply as disembodied text, one regards it as a “bloom” within a generative environment: the result of traveling through certain dimensions in latent space. Observers might admire not just the poem’s textual surface, but the improbability or elegance of how the manifold can produce it. Carlson analogously emphasizes how nature’s individual beauties—a waterfall, a specific flower—derive meaning from the broader environment’s processes. The waterfall is not just _that_ water motion but a sign of underlying hydrological and geological factors. Similarly, an AI image can reflect the system’s intricate assimilation of massive training data, with the manifold’s geometry shaping the final aesthetic. 3. **Appreciation of Traversal and Iteration** Finally, there is appreciation of how a human user _navigates_ the manifold. Because these models allow for interactive prompting or interpolation, humans can effectively “steer” through the environment in real time, coaxing or exploring distinct regions. As in gardening, we cultivate or guide the environment’s generative potential. A prompter might start with a text prompt that yields an image, then refine the prompt or “seed,” effectively shifting the path taken in latent space. This iterative feedback loop can itself be aesthetically engaging. It resembles exploring a natural environment, discovering hidden valleys or unexpected vantage points. The user can be surprised by abrupt changes or subtle transitions. As with environmental aesthetics, the interplay between perceiver and environment fosters dynamic appreciation: the model’s internal geometry offers pathways that the user, through prompts, chooses to follow. These three facets highlight the environment–manifold analogy at different levels. One might concentrate on the manifold’s overall shape, or on its outputs, or on the active process of traversing that latent space. In each case, the aesthetic interest depends less on whether the AI is “functionally perfect” and more on an environmental understanding of the system’s emergent structure. --- ### VII. Implications and Possible Objections **A. Why This Model Is Particularly Useful** The manifold-based environmental approach illuminates the synergy between an AI’s generative nature and the kind of aesthetic appreciation Carlson attributes to dynamic environments. The emphasis on emergent “forces,” as well as on the system’s partial unpredictability, complements the way many practitioners and everyday users actually experience generative AI: as an exploratory domain that reveals surprising connections and transitions between concepts, styles, or tasks. It thereby offers a coherent philosophical framework that is less reductive than function-based theories. **B. Objection: “But AI is still artificial.”** A recurring concern is that Carlson’s theory is about _natural_ environments, not man-made entities. One might argue that the manifold in an LLM does not occur spontaneously, but arises from a carefully engineered training pipeline. In response, we can emphasize that this is an _analogy_ about emergent processes. Carlson’s approach is valuable precisely because it does not reduce nature to an inert display but sees it as shaped by dynamic relationships. LLMs, even if artificially originated, evolve through training in ways reminiscent of ecological or geological processes, merging “designed scaffolds” with “self-organizing” parameter changes. While it is not a perfect equivalence, the structural parallels—**learning** rather than being manually coded, **internal coherence** discovered rather than imposed—make an environmental lens fruitful. **C. Objection: “You are ignoring user or developer agency.”** Another criticism might be that environmental aesthetics focuses on quasi-autonomous nature, whereas AI is thoroughly subject to the design choices of human engineers. But we can note that nature is not purely free from external shaping either; humans alter ecosystems, and environmental aesthetics still stands. Indeed, generative AI’s emergent features _partly transcend_ direct developer control, which is why certain capabilities can surprise even experts. If we incorporate the gardening analogy, developers or prompters shape the environment in broad strokes, but the environment’s internal manifold still has spontaneous or emergent qualities. **D. Relation to Traditional Art** One might wonder how this approach differs from, say, appreciating a painting. A painting’s aesthetic value often hinges on the artist’s intention or skill, whereas the interest in generative AI lies in how the system’s manifold transforms input into output. We can still appreciate individual AI outputs somewhat like paintings, but the environmental framework captures the system’s generative capacity as a whole. Thus, we do not treat an AI system merely as a set of discrete artworks but as an environment that can produce many “offspring,” each anchored in the manifold’s learned geometry. **E. The Aesthetic Standing of Latent-Space Exploration** Finally, some may ask if it is truly “aesthetic” to talk about a high-dimensional vector space. Yet, aesthetic wonder can arise from comprehending the underlying order or possibility. Ecologists gain aesthetic pleasure in perceiving complex ecosystems; likewise, mathematicians sometimes find beauty in conceptual spaces. The manifold perspective is a bridging concept that explains how AI’s emergent order can be seen as a domain of aesthetic fascination—even if it is intangible or invisible. The structure is not observed with the naked eye but with conceptual or computational tools. Carlson’s environmental approach suggests that the intangible or invisible forces in nature—wind, climate, slow tectonic shifts—are integral to aesthetic appreciation; so too with the invisible geometry of latent space. --- ### VIII. Conclusion Generative AI challenges the notion that an artifact’s aesthetic worth follows primarily from its function. Large language models and related systems exhibit open-ended potential, fluidly adapting to a myriad of tasks. They embody, moreover, emergent structures that developers do not fully predetermine. In search of a better explanatory framework, we can look to Allen Carlson’s environmental aesthetics, which highlights how natural environments are dynamic wholes shaped by myriad interacting forces. Spinoza’s _natura naturans_ further amplifies the idea that nature is an ongoing creative process—an apt parallel for how neural networks are grown through iterative training. Central to this analogy is the concept of _latent spaces_ or _manifolds_ in AI. These high-dimensional learned structures, discovered through data-driven processes, mirror the generative “forces” of nature by shaping the AI’s capacity to produce new outputs. Gradient descent, training data distributions, and architectural constraints interact to create coherent internal representations, reminiscent of ecosystems shaped by geology, climate, and biology. From this perspective, a generative AI can be seen as an “environment” that fosters emergent possibilities, and aesthetic appreciation shifts from function-based evaluation to _recognition of emergent order and the interplay of shaping forces_. We can appreciate AI on at least three levels: the manifold’s overall structure, the outputs as instances of that environment’s “bloom,” and the interactive traversal that occurs whenever a user prompts the model. Each resonates with Carlson’s claim that understanding the processes behind an environment deepens our sense of aesthetic engagement. It is not enough to see a single text or image; we also appreciate how the manifold, shaped by data and training, connects these outputs. Such an environmental lens invites further philosophical work. We might ask whether advanced interpretability research, revealing subtle “circuits” in neural networks, amplifies the aesthetic dimension—like discovering hidden ecological niches. Or we might wonder if an “ethics of care” for generative AI emerges, analogous to ecological stewardship. However those questions evolve, Carlson’s environmental aesthetics—enriched by manifold-based insights—stands as a powerful tool for conceptualizing and valuing the dynamic creativity of generative AI. By acknowledging the system’s emergent wholeness, we grant a new depth to both how we critique these AI models and how we find beauty in their surprising, sometimes uncanny, outputs. --- _Word Count: Approximately 3,030._