# Draft 27 Feb 2025 - [type:: draft] - [date:: 26 Feb 2025] - [paper:: [[the environmental aesthetics of generative ai]]] ## Introduction Here, I argue that there is an instructive parallel between how we come to appreciate the [[natural environment]] and how we might come to appreciate the output of generative AI (from here on 'ai art') such as Midjourney, ChatGPT, and Sora. It is clear that the outputs of generative ai systems are a quite different sort of thing to traditional artworks: a painting or a poem can take months and be a direct result of an individual's labour, attention, and effort; but an AI painting or AI poem can be be produced in seconds, and is [[the result]] of (broadly speaking) **a vast corpus of [[experiential artifacts]] (images, text) encoded in [[latent space]], and a user's prompt.** This means that when it comes to appreciation we have [[two options]]: 1. AI art is not the sort of thing that can be aesthetically appreciated. 2. we appreciate ai art in a different way than we do art art In this paper, we put forward a way of making sense of the second. We start by considering a related question: can generative ai systems *themselves* be aesthetically appreciated? One possibility is that we appreciate them as [[designed objects]], and can therefore adopt one of the recent accounts of [[design aesthetics]] that have been proposed. **After showing why this is unlikely to work**, we put forward our own suggestion: we can make sense of the [[aesthetic appreciation]] of AI systems by adapting Carlson's *[[environmental aesthetics]]*. According to Carlson's [[Natural Environmental Model]] appreciation of nature rests on understanding of *what it is* –particularly, but not exclusively, from a scientific perspective. Similarly, we suggest that appreciating the outputs of generative AI systems such as midjourney or chatgpt can be grounded in our knowledge of what those systems are. –particularly, but not exclusively, from a scientific point of view. **Finally, we return to [[the aesthetics]] of the outputs of generative AI, arguing that.** ## 1. The Failure of Function - Generative AI is sometimes described as a 'new type of tool' (REF). - If this were correct then existing accounts of [[design aesthetics]] could be extended to cover these such systems. - This would at least be a step towards understanding how ai art can be appreciated. - Two recent accounts of [[the aesthetics of design]] suggest that our appreciating a [[designed object]] rests in how well it fulfils its function. - For Parsons and Carlson's Functional Beauty approach, understanding an object's function shapes the way it looks. - If an object appears well-suited to its function it will look aesthetically pleasing. - In similar vein, Forsey's approach assesses how an object fulfils its intended purpose, judging it as a specimen of its kind and considering how it realises its function. However, this reliance on function does not seem well-suited to generative AI systems. Unlike hammers, or even digital applications like Microsoft Word, LLMs at least do not have a single, fixed function. What is ChatGPT *for*? ### More needs to go here - one more sentence trying to make this persuasive - Move onto the form/function argument, which might be stronger. - Indeed, an interesting thing about LLMs is that uses for them have been discovered by users after they have been released. ' llm link to other article about novel discoveries a) is this sentence true? b) if it is search the web for cases in which users have discovered novel uses for LLMs. - One system might generate images, produce text for study aids, draft code, compose short stories, or support a range of unforeseen applications. - Due to this open-endedness, it is unclear which function-based standard would allow us to judge whether a system’s physical form or underlying architecture adequately fulfils one purpose. - Moreover, many designers of generative AI cannot entirely explain how these systems accomplish their various uses. - %% Add stuff about form and function being part of perfection seeing or judging%% - Deep learning involves complex, sometimes opaque processes in which a model “learns” from large-scale data rather than being explicitly programmed for a singular goal. - The resulting “function” is open to user innovation, rather than being fixed in advance. - With no single specification to anchor an evaluation, conventional function-based theories do not neatly capture how we might aesthetically engage with or evaluate such models. %%the above can be cut down fairly substantially i think.%% ## 2. The Natural Environmental Model Carlson's *Natural Environmental Model* for environmental aesthetics rests on two ideas: >"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) I will consider each in turn. We can take "natural" in "natural environment" simply to mean something which, unlike art or artifacts, is not man made; "environment" needs spelling out in a little more detail. Carlson does not believe that the environment should be thought of as a collection of objects or scenic views, but a system of interconnected elements in varying types of relationship, and subject to various types of processes and forces. (ibid. p. 44) and that even environments themselves "come about “naturally,” [in that] they change, grow, and develop by means of natural processes.". The elements, processes, and forces that Carlson has in mind here relate to the second part of the natural environmental model: that appreciation of nature takes place "in light of knowledge provided by the natural sciences" (REF). Understanding the elements and processes postulated by biologists, geologists, etc. are what Carlson believes are essential for appreciation of the natural environment. Indeed, he compares understanding of these systems to the understanding of artistic traditions, practices, etc. that required to properly appreciate artworks. Consider examining a coastal cliff composed of sedimentary rock layers. When understanding their formation through sediment deposition and compaction, one's attention is drawn to variations in texture, colour, and composition in a way that a novice's would not. As Carlson puts it: >When nature is aesthetically appreciated in virtue of the natural and environmental sciences, positive aesthetic appreciation is singularly appropriate, for, on the one hand, pristine nature—nature in its natural state—is an aesthetic ideal and, on the other, as science increasingly finds, or at least appears to find, unity, order, and harmony in nature, nature itself, when appreciated in light of such knowledge, appears more fully beautiful. (p. 11) This way of conceptualising the processive nature fits well with an older idea: Spinoza's concept of _natura naturans_. Natura naturans refers to the active, self-causing dimension of nature, which he equates with "God or Nature".[^1] - **one more evocative sentence** - A further analogy can be drawn to Spinoza’s idea of natura naturans (nature as active, self-producing). - Carlson’s discussion of environments as dynamic resonates with Spinoza’s depiction of nature as an ongoing, generative force, in contrast to natura naturata (nature as created or already-produced). ## 3 Generative Environments We suggest that Carlson's recommendations also hold for the aesthetic appreciation of generative AI: we should appreciate these system for what they are, and we should appreciate them in light of our scientific knowledge of what they are. Moreover, we shall suggest in a moment that Carlson's Environmental Model (the naturans model) can in fact be adapted so as to provide at least the starting point for an aesthetics of generative AI. We can begin to see this by considering the following remarks from Dario Amodei, the CEO of Anthropic (my italics). > 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, and they sort of, you know, 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. > > And so we create the scaffold that it grows on and we create the, you know, the light that it grows towards. But the thing that we actually create, it's this almost biological, you know, entity or organism that we're studying. And so 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. > > (Lex Fridman, Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452) my emphasis.  Our contention is that generative AI can be approached in a way that parallels Carlson’s environmental model. - That is, generative AI can be understood as another type of _naturans_ system, *machina naturans*, allowing for generative ai systems to be appreciated in a similar way to Carlson suggests for the environment. - When Amodei talks about continuous, partly autonomous production captures something about how generative AI operates. - - In neural network training, an architecture is designed, a loss objective is specified, and then the system “develops” through iterative exposure to data. - This leads to emergent capabilities that are not individually coded by designers. - - This “growth” is reminiscent of a partly autonomous or emergent process. Designers do not individually code every capability; rather, capabilities arise from iterative exposure to large-scale data. - The analogy to a “growing” system highlights how these models evolve in complexity and capability over time, as though they were living entities shaped by training forces. - In generative models, outputs often emerge from learned patterns rather than explicit programming that pre-specifies every aspect. - The system “generates” rather than merely retrieves. - This sense of an ongoing, partly self-directed process is reminiscent of how Spinoza depicts nature’s productive activity. - With an AI system, user prompts supply new contexts or “seeds,” but the internal generative routines autonomously create novel configurations. - Hence, just as natura naturans emphasises nature’s active production of phenomena, generative AI might be regarded as a “machina naturans,” a machine that continually produces outputs shaped by internal processes and interactions with users. - **3.1 An Evolving, Grown Process** - Some researchers describe large models as “grown” rather than strictly programmed, invoking a biological analogy. - Rather than having a single function, generative AI systems can serve multiple roles. - One user might prompt a text-based model for language tutoring, another for creative writing or code debugging, and yet another for summarising research articles. - This open-ended flexibility is better captured by thinking of the system as an evolving environment than as a finite tool with a single purpose. - **3.2 Machina Naturans** - Drawing from the notion of natura naturans, we can refer to generative AI as machina naturans. - This term signals that these systems display a self-directed or emergent aspect; they produce content in ways that cannot be reduced simply to direct instructions. - Internally, generative AI relies on latent spaces: high-dimensional, learned representations in which data are mapped and associated. - These spaces resemble a conceptual “terrain” shaped by the model’s training set. - A text-to-image system, for instance, learns an extensive “manifold” of possible images. - Depending on the prompt and the underlying latent space, different regions of this manifold are activated, producing varied, sometimes surprising outputs. - One might compare the training data to “soil,” whose composition influences what “blooms” in the model’s outputs. - If certain art historical styles are prevalent in the dataset, the model will “grow” images reflecting those styles more readily. - If an entire region of possible imagery was scarcely included in the dataset, prompting for those styles or subjects might lead to incongruous or incomplete results. - Even adjusting parameters like temperature or sampling strategies can be likened to modifying the “soil conditions,” altering how easily certain outputs are “encouraged.” - In short, the generative system behaves as a dynamic environment whose emergent forms reflect the interplay among data distributions, model architecture, and user prompts. - **3.3 Navigating an Internal Terrain** - Once trained, a generative model can be explored through prompts, akin to navigating a digital environment. - A user might discover that requests for certain artistic styles or textual formulations produce outcomes that cluster around particular “regions” of the latent space. - Some of these regions might be more easily reached; others might require intricate prompting or model fine-tuning. - The system’s “terrain” can therefore appear surprisingly vast and varied. - Seeing generative AI as an evolving system—rather than a static artefact or single-function tool—underscores the importance of knowledge for aesthetic appreciation. - One can certainly enjoy an AI-generated image or poem on first glance, but deeper engagement benefits from some sense of how the “soil” was formed and how a user’s input interacts with the system’s latent space. - This conceptual move parallels Carlson’s argument that appreciating a forest or reef is enriched by understanding the ecological forces driving its structure. - In Carlson’s words, environments are shaped by “various forces, random and otherwise,” and appreciation requires an account that “makes this order visible and intelligible.” - Generative AI is likewise formed by “forces,” such as training data distributions, model architectures, and parameter tuning. - While those forces are cultural, algorithmic, or mathematical rather than natural, they shape what outputs the model can produce. - To consider a generative model as an “environment” is to see it not as a static artefact or a single-purpose machine, but as a dynamic system whose “terrain” can be navigated by users through prompts or queries. - Much like a natural environment that allows multiple pathways of exploration, an AI system—once trained—can generate a multitude of textual or visual creations, driven partly by user interaction and partly by its learned internal structures. --- In neural network training, an architecture is designed, a loss objective is specified, and then the system “develops” through iterative exposure to data. - This leads to emergent capabilities that are not individually coded by designers. - The analogy to a “growing” system highlights how these models evolve in complexity and capability over time, as though they were living entities shaped by training forces. - Rather than having a single function, generative AI systems can serve multiple roles. - One user might prompt a text-based model for language tutoring, another for creative writing or code debugging, and yet another for summarising research articles. - This open-ended flexibility is better captured by thinking of the system as an evolving environment than as a finite tool with a single purpose. - 3.2 Machina Naturans - Drawing from the notion of natura naturans, we can refer to generative AI as machina naturans. - This term signals that these systems display a self-directed or emergent aspect; they produce content in ways that cannot be reduced simply to direct instructions. - Internally, generative AI relies on latent spaces: high-dimensional, learned representations in which data are mapped and associated. - These spaces resemble a conceptual “terrain” shaped by the model’s training set. - A text-to-image system, for instance, learns an extensive “manifold” of possible images. - Depending on the prompt and the underlying latent space, different regions of this manifold are activated, producing varied, sometimes surprising outputs. - One might compare the training data to “soil,” whose composition influences what “blooms” in the model’s outputs. - If certain art historical styles are prevalent in the dataset, the model will “grow” images reflecting those styles more readily. - If an entire region of possible imagery was scarcely included in the dataset, prompting for those styles or subjects might lead to incongruous or incomplete results. - Even adjusting parameters like temperature or sampling strategies can be likened to modifying the “soil conditions,” altering how easily certain outputs are “encouraged.” - In short, the generative system behaves as a dynamic environment whose emergent forms reflect the interplay among data distributions, model architecture, and user prompts. - 3.3 Navigating an Internal Terrain - Once trained, a generative model can be explored through prompts, akin to navigating a digital environment. - A user might discover that requests for certain artistic styles or textual formulations produce outcomes that cluster around particular “regions” of the latent space. - Some of these regions might be more easily reached; others might require intricate prompting or model fine-tuning. - The system’s “terrain” can therefore appear surprisingly vast and varied. - Seeing generative AI as an evolving system—rather than a static artefact or single-function tool—underscores the importance of knowledge for aesthetic appreciation. - One can certainly enjoy an AI-generated image or poem on first glance, but deeper engagement benefits from some sense of how the “soil” was formed and how a user’s input interacts with the system’s latent space. - This conceptual move parallels Carlson’s argument that appreciating a forest or reef is enriched by understanding the ecological forces driving its structure. - shaped by evolving forces and conditions. - In Carlson’s words, environments are shaped by “various forces, random and otherwise,” and appreciation requires an account that “makes this order visible and intelligible.” - Generative AI is likewise formed by “forces,” such as training data distributions, model architectures, and parameter tuning. - While those forces are cultural, algorithmic, or mathematical rather than natural, they shape what outputs the model can produce. - To consider a generative model as an “environment” is to see it not as a static artefact or a single-purpose machine, but as a dynamic system whose “terrain” can be navigated by users through prompts or queries. - Much like a natural environment that allows multiple pathways of exploration, an AI system—once trained—can generate a multitude of textual or visual creations, driven partly by user interaction and partly by its learned internal structures. - 2.3 Spinoza’s Notion of Natura Naturans - - Although Carlson is concerned with natural rather than artificial systems, the philosophical image of continuous, partly autonomous production captures something about how generative AI operates. - In generative models, outputs often emerge from learned patterns rather than explicit programming that pre-specifies every aspect. - The system “generates” rather than merely retrieves. - This sense of an ongoing, partly self-directed process is reminiscent of how Spinoza depicts nature’s productive activity. - With an AI system, user prompts supply new contexts or “seeds,” but the internal generative routines autonomously create novel configurations. - Hence, just as natura naturans emphasises nature’s active production of phenomena, generative AI might be regarded as a “machina naturans,” a machine that continually produces outputs shaped by internal processes and interactions with users. - 3. What Generative Systems Are - 3.1 An Evolving, Grown Process - Some researchers describe large models as “grown” rather than strictly programmed, invoking a biological analogy. - ## 4. Appreciating Generative Systems and their Outputs - 4.1 Beyond Purely Human-Made Artworks - When confronted with a text or image generated by an AI system such as Midjourney, one might judge it as if it were a human-made artwork. - However, this approach often misrepresents the partly autonomous way these outputs arise. - Alternatively, seeing the output as “the fruit of an evolving, partly self-directed process” aligns with Carlson’s emphasis on the interplay of shaping forces. - The horticultural analogy is apt here: just as a gardener cultivates certain conditions but cannot wholly dictate how plants will grow, a user can direct an AI model but not entirely foresee or control every outcome. - In classical design aesthetics, we often look for a product’s function and judge whether its form fits that end. - In the environmental approach, by contrast, we notice how multiple elements are shaped over time by various forces. - A generative AI system similarly blends design (model architecture) and emergent growth (training results). - Its “outputs” stand in partial continuity with the data patterns and the user’s instructions. - This suggests a new lens for aesthetic appreciation, focusing on how a system’s generative process evolves and how that process is manifested in any particular work. - 4.2 A Horticultural Metaphor - In horticulture, appreciating a cultivated yet semi-wild landscape involves acknowledging not only the gardener’s design but also natural processes like soil composition, seed dispersal, and weather patterns. - Similarly, with AI, the user interacts with a machine whose outputs depend on complex internal processes. - Users guide the model through prompts, while the underlying network architecture, data distribution, and training constraints function like the soil and climate in which creative “plants” grow. - This metaphor underscores that knowledge of the system’s “ecology”—the interplay of data, parameters, user inputs, and internal representations—enables a more informed aesthetic appreciation of AI outputs. - One sees the final text, image, or music track as emerging from the system’s dynamic environment rather than from a purely mechanical or fully intentional creation. - 4.3 Carlson’s Focus on Informed Attention - Carlson often emphasises that knowledge guides how we pay attention, preventing aesthetic experience from devolving into aimless gazing. - In the context of AI, practical or even technical knowledge helps direct our attention to relevant features of the output and the processes behind it. - For instance, realising that the model has absorbed certain biases or stylistic preferences might prompt us to look for telling signs in the images it generates. - Knowing that it was trained on predominantly Western art can direct our attention to the specific forms, palettes, and compositional tendencies that recur. - None of this demands that everyday users master the intricacies of backpropagation or neural network topology. - Carlson points out that scientific understanding is related to everyday knowledge by degree rather than by a categorical divide. - A home gardener need not be a trained botanist to grasp something of soil acidity, sun exposure, and pollination. - Likewise, one might cultivate an intuitive familiarity with an AI system by experimenting with prompts and observing patterns in its responses. - This experiential, hands-on engagement supports an aesthetic appreciation that remains grounded in a practical sense of the system’s generative character. - 5. Objections - 5.1 The Technical Knowledge Objection - One objection is that an environmental approach to generative AI may demand substantial technical or specialist knowledge. - If we accept Carlson’s view that knowledge of an environment’s forces underwrites aesthetic appreciation, it might seem that only machine learning experts can fully appreciate AI. - This appears to contradict the relative accessibility of AI systems to non-specialist users. - However, Carlson’s continuum argument offers a reply: although a professional ecologist has a deep grasp of biological systems, a gardener can achieve meaningful environmental knowledge through experience. - The same is true of generative AI. - Knowledge can be layered. - Some users might be experts with detailed understanding of model architectures, while others gain familiarity by interacting with the system, noticing patterns, and sharing anecdotes. - Both levels of knowledge help transform otherwise inscrutable AI outputs into more meaningful, aesthetically appreciable phenomena. - 5.2 The Objection of Rapid Technological Change - Another concern is that generative AI evolves quickly. - Once a user achieves a basic understanding of how one model behaves, a new version or competitor may appear, potentially overturning old insights. - A sceptic might argue this constant flux undermines the stable base of knowledge required for aesthetic appreciation. - Yet one might respond that this mirrors the ongoing change in natural environments themselves—ecosystems evolve, species migrate, and climates shift. - In Carlson’s analysis, knowledge remains an evolving, provisional tool for guiding attention. - Similarly, as AI systems update or diverge, we can maintain an informed stance by continuing our engagement, much like an environmentalist keeps abreast of ecosystem changes. - The appreciation process becomes an ongoing dialogue with a system in flux, rather than a once-and-for-all acquisition of facts. - 5.3 Specialist Credentials and Public Engagement - A final objection is that only a limited community has the resources to become skilled in “AI gardening.” - Yet we already see that generative AI has permeated popular culture, with hobbyists, visual artists, writers, and lay enthusiasts actively exploring these systems. - They exchange prompts, share discoveries about unexpected outputs, and comment on emerging trends. - This cultural practice itself fosters a collective, practically oriented knowledge base. - Indeed, many participants in these communities have no formal machine learning training, but still develop a sense of which prompts generate intriguing results, what stylistic or narrative shapes might emerge, or how a model’s “personality” differs from earlier versions. - Such grassroots familiarity can suffice to underwrite the kind of informed aesthetic attention that Carlson extols. - Concluding Remarks - Generative AI resists neat classification as a functionally defined artefact or as purely human-made work. - Instead, these systems can be seen as dynamic environments whose outputs result from complex interactions among data, model architectures, training regimes, and user inputs. - Drawing on Carlson’s environmental aesthetics, we have suggested that users might appreciate generative AI by attending to these formative “forces,” whether through formal expertise or everyday experimentation. - The parallel with natural environments, while imperfect, highlights how appreciation can arise from seeing AI as a partly autonomous, evolving system, rather than merely a tool to be judged by a single function. - By understanding how training data, prompts, and algorithmic constraints collectively “shape” the system, we can develop a richer aesthetic engagement with its outputs. - Carlson’s approach challenges us to look beyond surface properties and identify an object or environment’s formative structure, whether in a forest or a neural network. - While appreciating a forest might draw on knowledge of ecology or geology, appreciating AI systems can draw on the interplay of training data, model design, and emergent capabilities. - Both cases involve knowledge that orients our attention toward deeper patterns. - Ultimately, the environmental model offers a promising framework for understanding how generative AI can become an object of aesthetic appreciation, one that honours the open-ended nature of these systems while maintaining a structured, knowledge-inflected mode of engagement. [^1]: 1. The twin of natura naturans is _natura naturata_, the stative aspect of nature, produced by naturans.