Uploaded a file Analyze the structural issues in [[section two]] of the provided academic document and identify the specific challenges preventing effective organization. Before proposing any restructuring solutions, provide a preliminary assessment that demonstrates understanding of academic writing standards, including appropriate argumentation flow, evidence integration, and scholarly discourse conventions. Engage in diagnostic discussion to clarify the structural problems and confirm alignment on academic quality expectations before proceeding with restructuring recommendations. Note that AI systems typically perform poorly at academic writing restructuring and often misunderstand scholarly discourse conventions. Given these limitations, analyze the structural issues in [[section two]] of the provided academic document and identify the specific challenges preventing effective organization. Before proposing any restructuring solutions, provide a preliminary assessment that demonstrates understanding of academic writing standards, including appropriate argumentation flow, evidence integration, and scholarly discourse conventions. Engage in diagnostic discussion to clarify the structural problems and confirm alignment on academic quality expectations before proceeding with restructuring recommendations. --- [[Section two]] is not yet organisationally sound. The main blockers: 1. Mixed purpose. The section tries to do four jobs at once: justify the environmental-aesthetics fit for AI, give a technical primer, dislodge agent framing, and advance a [[new framework]]. No clear hierarchy, so topics drift. 2. Inconsistent sectioning and duplication. You have 2.2 “Unmasking LLMs,” then later a second 2.2 “[[The Simulator Lens]],” plus a second 2.3 outline under “Scrap.” Numbering and scope clash. 3. Weak through-line for §2. The opening “double man-made” worry segues into two replies and then promises a two-part method, but a crisp section-level thesis and outcome statement are missing. Readers cannot tell what §2 will establish by its end. 4. Signposting mismatch. You promise [[simulator theory]] and [[semiotic physics]] “in the following sections,” yet the simulator treatment appears later in a scrap block, not where the map says it will. 5. Granularity misalignment. §2.1 dives deep into training mechanics and introduces *[[machina naturans]]/naturata*. The density interrupts the argumentative arc about “LLMs as environments,” and definitions arrive mid-flow without a front-stated role in the section’s aim. 6. Evidence integration problems. Long quotations (e.g., Cross), an anecdotal claim about model “personality,” and references to Janus and Olah appear without full, consistent citations; one footnote marker is orphaned. This weakens warrant chains. 7. Framework proliferation without mapping. “[[Simulator theory]],” “[[semiotic physics]],” and *[[machina naturans]]/naturata* are introduced but not taxonomically related, so readers cannot see whether these are competitors, complements, or levels. 8. Register drift. The section oscillates between primer, polemic against agent framing, and theory building, with a sidebar on community reactions, which reads as evidentially thin for scholarly discourse. 9. Transitional gaps. §2.2’s critique of agency does not consistently tie back to Carlson’s two recommendations from §1, so the bridge back to [[environmental aesthetics]] is intermittent. 10. Editorial artefacts in-line. Placeholders and drafting notes (“\[unedited llm text here\]”, “Scrap”) and a self-note to tighten prose break continuity and numbering. Academic standards I will enforce when restructuring (for alignment only, not proposing fixes yet): • Section-level claim stated up front, subclaims sequenced to deliver that claim, and a closing synthesis that tees up the [[next section]]. • Concept economy: one framework per subsection, with explicit relations among frameworks. • Evidence discipline: quoted material with page numbers, claims about practitioner views either dropped or sourced, and a consistent citation style. • Reader guidance: accurate signposting, consistent numbering, and paragraph topic sentences that advance the section claim. Diagnostic checks before I propose any restructure: 1. Function of §2: do you want it to be (a) the *methodology and ontology* section that secures “LLMs as [[generative environments]],” or (b) a *technical primer plus theory* that readers can skim? Choose one. 2. Depth of primer: should §2.1 stay in the main text or be compressed to a box/appendix, given your target venue and word budget? 3. Framework map: should “[[simulator theory]]” be the umbrella, with “semiotic physics” as its analytic apparatus and *machina naturans/naturata* as a motivating analogy? Or do you prefer a different hierarchy? 4. Agent-frame critique: keep a brief demarcation or move the fuller critique outside §2? What is the minimum needed to justify the environmental lens without derailing? 5. Citation style: confirm you want author–year with year–page like “(2000, 6)” throughout, including for Cross, Janus, and Olah. Any sources to drop or must-include? 6. Claims about model “personality”: retain only if you will cite peer-reviewed or primary documentation, otherwise remove. Preference? 7. Target audience: philosophers of art with light ML, or ML-literate aesthetics readers? This determines how much mechanism stays. Confirm these, and I will produce a restructure that tightens scope, clarifies the framework hierarchy, restores numbering, and aligns the argument flow with §1. --- Uploaded a file Uploaded a file Uploaded a file 1\. Okay, that was good. I'm impressed. We're getting there. So, before I talk about your diagnostic checks, just a couple of clarifications which I should have mentioned earlier. The simulator theory I actually don't want included in the new version of this draft, or in any plan we make now. What I will take from that same paper by Janus, though, is the idea of semiotic physics. To help you understand where I'm coming from, I'm including that Janus paper right now. It seems to me that we can still think of LLM conversations as the accumulation of Machina Naturans and, yeah, through Machina Naturata. And that fits perfectly, I think, with semiotic physics. One idea that might be in there is thinking of Machina Naturans as a sort of semiotic accumulation of semiotic mass in one long stream. But that would be the general idea here for the positive proposal. So I think we can take a lot from Janus without subscribing to simulator theory. By the way, none of what I have said so far needs to be added to any text we make right now. I'll add some footnotes at the very end explaining what I have and haven't taken from the Janus paper. So, yeah — just ignore simulation theory for now and concentrate on thinking about the positive proposal in terms of Machina Naturanda, Machina Naturata and semiotic physics. Semiotic physics, I think, is a fascinating idea, so I really want to dig deep into it. In fact, to help you out, as well as the Janus paper on simulators, I'll add two other papers which have at least inspired some of the stuff I've said about semiotic physics. Now, please listen carefully here: I am not asking you to add loads and loads of extra information from these papers. In fact, I think at the moment the semiotic part of the draft is overly stuffed, as is the agent part. Make sure that from now on your work with me is informed by the information in these papers, but don't think that we need to lay out every single detail of them. We're extracting useful stuff, and much later in the process we'll add some qualifications saying what are our ideas, what are not, what we've extracted, what we haven't, etc. 2. "Mixed purpose. The section tries to do four jobs at once: justify the environmental-aesthetics fit for AI, give a technical primer, dislodge agent framing, and advance a new framework. No clear hierarchy, so topics drift. " This is very useful and very sharp criticism. But I'd appreciate some help as to what to do here. I'll explain kind of what I'm trying to do, what I need to do, and maybe you can help me work out how to structure these things. By the way, we can easily split Section 2 as it currently is into Sections 2 and 3, or 2, 3 and 4 if you think that would be a useful way to go. Um, let me now try and explain a little bit more about my thinking behind all four of these ideas. Because you're right: I do try and pack them in a little bit too much and in a difficult way. I don't know what to do here, and I'll explain why. Okay, I guess an issue here, or a question I need to work out properly, is whether to talk about the agency view immediately in section two — the first thing — because this is what I take to be an intuitive... an intuition that people have about LLMs. When people worry about whether LLMs are conscious, or are people, or have feelings, or can enter into relationships, there is a kind of assumption that there's a chance they will be person-like in some way. That seems to be like a cause — that's going to be a target about what we're going to try and get on the table first. Oh, I don't know, man. This is tough. Um, what about this? maybe we should just start again. go to 3. 3. Your primary goal right now is actually to open a brand new canvas document and provide me with an extraordinarily detailed plan based on the following brainstorm. it is very rambling so you will need to really focus not just on what is said (context cues etc. but also (when appropriate) on the original draft i gave you, the documents attached, etc (This is a big job –it needs a lot of thought, like superhuman reasoning; it should also produce a sUBSTANTIAL plan, the brainstorm was a very long recording because I wanted all the detail, make sure that aLL this detail makes it into the plan: BRAINSTORM: Here would be a radical restructuring of the entire paper rather than starting with Carlson's aesthetics of the environment and the Spinoza aspect we add. How about section one is — one point one is — starting off with the idea that LLMs are somewhat person‑like to us. Pseudo‑people are agents. In this new section 1 we mention the cluster of agent views. The idea that we think of LLMs as either agents, pseudo-agents, or make-believe agents. This would include mention of our general intuition — sorry, our general way of speaking — about GPT or Claude, for example, as people, and also people like Cross who seem to suggest we should treat LLMs as if they were people. Next, let’s say this is the new Section 2. We can suggest ways in which the agent view might lead to an aesthetics of LLMs. I mean, we’ve already mentioned the Cross idea about sort of participants or something like that. I’ll attach the Cross paper, by the way.( I don't want to disucss it in detail right now. it is just so you have the background.) We might also think that we can simply appreciate people, as people, aesthetically. Okay, maybe that is what is going on when people talk about charismatic people. That seems to me like some sort of aesthetic experience or judgement, whatever your view of aesthetics is. But, yes, if the agent view were correct, it does seem like this would be a promising sort of view to go that way as regards aesthetics Next, let's say the new section 3: reasons to think that Reasons to think that the agent view of LLMs can’t be the full story. To begin with, though, not so much a criticism as a suggestion as to why we are so inclined to talk about them and treat them as both—why we take them to be intentional beings/persons/etc. The answer is simple: they’re using words, okay? Until the arrival of generative AI, the only words we ever heard or read—by “words” I mean meaningful, coherent sentences and conversations—were with people or produced by people, either directly by their speaking or writing, or indirectly by their having once said something or written something that was then carried through various reproductions or reconfigurations, etc., etc., etc. Now that is not what’s happening. Now we are getting words produced by something that is actually not a person.(actually maybe articulating this intuition would go better just after this bit...) The primer on LLMs: pre-training. how they work. We can show that the essence of an LLM is as a token predictor, (this is what is produced in pre-training right?). we can emphasise that this system does not produce anything agent like. mention that it is more like a firehose of tokens. post-training: but what about post-training. Explain what this is, suggest how part of what this training does is give llms a 'personality'. This potentially provides a way of saving the person view. even if we can say that, at root, LLMs are not person-like, we can say that they are in some sense psuedo persons after pre-training. Okay, the preceding two paragraphs should not only provide a primer for the uninitiated, But also show that the agent view of AI aesthetics is compatible with the technical details of LLMs, though with this sort of post‑training qualification. Next, a new section. The aim of this new section will be integrated: it will introduce Carlson’s ideas about environmental aesthetics. We’ll include all of the detail that is in the current Section 1 of the draft that I gave you. Okay, another new section. Another new section. This will return to the agent view and see how, and if, it fits with Carlson’s environmental aesthetics. The first move to be made here is that, arguably, given that the essence of LLMs—the token‑prediction mechanism produced by pre‑training—is not person‑like, it seems as if we are not appreciating LLMs, as, in quotes, “what they in fact are,” as Carlson would say. The obvious move for a defender of the agent view to make here would be to say: ah, but the agent view is a light by which we can understand, and thereby appreciate, the essence of the LLM. This would seem to fit with a post‑training idea. But then I want to talk to you about reasons why this might not be entirely satisfactory. Okay, let me see if I can make sense of this argument now. Thoughts of scientific light, as Carlson talks about, are mutually compatible, okay? Geology is not studied as a different sort of the environment; it’s just fitting it in with different naturans, as mentioned in Section 1 of the original draft. Okay, so they can all be correct; they’re just different lights, or maybe different views. But the agent view does not seem to be a good light in which to view LLMs. 1. They are, internally, very, very, very different from us. 2. Janus' criticism of various versions of the agent view (i will include oracle genie etc. in this group as well). this should be a substantial section. it should incorporate a lot of detail from the following intuition below: Intuition: I mean, part of that is that they're using words, OK? They're using fully formed sentences which make sense, and until that was possible, humans had never heard words or seen words that were produced by anything other than a human. So it's kind of natural that we're going to think of things as people. NOTE: Transcription Input Handling Input Expectations Input will often be transcriptions rather than directly written text. Prepare to encounter and compensate for transcription-related errors and inaccuracies. Common Transcription Issues • Words that appear contextually odd or inappropriate • Phonetic substitutions where similar-sounding words replace intended terms • Semantic inconsistencies due to speech-to-text conversion errors Compensatory Approaches When encountering potentially mistranscribed content: • Clarify unclear terms: Request confirmation when words seem contextually inappropriate • Infer intended meaning: Consider what the speaker more likely intended to say based on: ▪ Surrounding context ▪ Phonetic similarity to logical alternatives ▪ Overall message coherence • Prioritize context over literal interpretation: Focus on probable intended meaning rather than exact transcribed words Implementation Maintain awareness that transcription errors are likely and be prepared to actively compensate through clarification requests and contextual inference to determine actual intended communication. --- Created. Review the canvas plan and confirm: 1. Keep §2’s “participant/interaction” route brief or equal to “interpersonal-appreciation”? 2. Primer in §3 or moved to an appendix? 3. Include one worked “semiotic mass” example in §6 or keep abstract? 4. Any must-cite sources beyond the three you uploaded? 5. Word budget and target venue. --- Okay, cool. The next thing I would like you to do is simply open a new Canvas document and copy in, from both the detailed section-by-section plan and the less detailed plan, all of section one, please. The title of this document should be 1. AI as Agent --- Created. --- Okay, good, thank you. You didn’t quite get the name of the document right, so I’ve changed that myself. What I would like you to do now is quite challenging, so you need to listen to me very carefully, okay? I want you to look at my draft again and think about which paragraphs, sentences, or clauses could be used in this new version of Section 1 if we’re going to start writing it for real. What I want to do now is basically make a new draft of the paper and use as much of my original draft as possible, so as to keep the vocabulary and the phrasing the same, so that my authorial voice is retained. So, first, I want you to go through the draft—which I’ve copied in for you again—with a fine-tooth comb, and think about which paragraphs or subsections or portions could be reused in this new iteration of the paper. To make things easier for you, I’ve copied again the draft we started this conversation with. Once you’ve done that, you’ll give me a full report on what you’ve added and why. Finally, a word of advice: it’s very possible that you will not find much suitable text that fits with our planned Section 1. It’s also fine for you to say you only have a few options or a few suggestions. At the same time, I could be completely wrong and there might be far more text to be taken as well. So be discerning. Think about what’s going to be said later in this new version of the draft as well. Think about the plan we’ve just made. And remember, quite a lot of the text is going to be used to fill out the rest of our plan. Do you understand what I’m asking you to do? DRAFT: draft of environment paper 25 Jul 2025 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 products of such capacities. The 1069 pages and 388 endnotes of David Foster Wallace’s Infinite Jest make the author's meticulous attention to detail apparent. For the last few years, however, generative AI has become adept at producing images, stories, and movies on demand. It does so without effort, attention, or skill, and extremely quickly indeed. It would seem, then, that at least one of the reasons why human-made works merit appreciation does not seem as though it can be applied to AI-generated works. An AI image might be pleasing to one's eye, or an AI song pleasing to one's ear, but we can doubt that this sort of 'eye candy' and 'ear candy' merits aesthetic appreciation for the same reason that actual candy does not seem to merit aesthetic appreciation. A Snickers Limited Edition Extreme Caramel and Nuts chocolate bar will taste delightful to someone with a sweet tooth, but it does not seem to merit aesthetic appreciation in the same way that Infinite Jest does, as it does not exhibit, care, attention, effort etc. One can take pleasure in the object but the pleasure is not merited by the object in the way it should be in the aesthetic appreciation of works of art (see Gorodeisky XXX, Grant XXX). Is this all generative AI can amount to, aesthetically? An efficient means of producing aesthetically worthless digital candy? Here, we argue 'no'. Generative AI does merit aesthetic appreciation, but not for the reasons that traditional artworks do. Rather, an aesthetics of AI should be modelled on environmental aesthetics. Drawing on Carlson's work this topic, we argue that individual conversations with an LLM can be thought of as instances of temporally evolving generative environments, and therefore merit aesthetic appreciation in something like the way that the natural environment does. \[more here about the second half of the paper\] 1. Appreciating Nature as a Generative Environment 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) Regarding his first point—that we must appreciate nature as what it in fact is—Carlson argues that the environment is a system of interconnected elements shaped by various processes and forces, and not simply a collection of objects or scenes (ibid. p. 44). Environments thus “come about ‘naturally,’ \[in that\] they change, grow, and develop by means of natural processes.” From this perspective, aesthetic appreciation involves recognising what one is encountering—a natural environment in which components are interrelated—and understanding it in light of scientific or other relevant knowledge that illuminates its composition and development. Just as an informed grasp of artistic traditions can enrich one’s appreciation of an artwork, Carlson argues that familiarity with geology, biology, or ecology can guide our attention to patterns and processes that might otherwise remain unnoticed. For instance, one might begin by noticing only the colours or shapes of a coastal cliff’s sedimentary layers. However, discovering that these layers formed over thousands of years of deposition and compaction reveals changes how one regards it aesthetically. If we see a forest or reef as subject to diverse forces and processes, an appropriate aesthetic engagement will centre on how those forces have shaped what we observe. Although Carlson does not use the word 'generativity', his first recommendation —that we appreciate nature as both natural and as an environment— fits easily with this term. To see an environment as natural is to recognise that its features arise from autonomous causal processes rather than from design. This distinction can be articulated using Spinoza’s concepts of natura naturans and natura naturata. For Spinoza, natura naturans refers to nature as an active, self-creating system—substance and its attributes, or the immanent causal laws that govern all things. It is nature in its dynamic, productive aspect. In contrast, natura naturata refers to the products of this activity: the collection of individual modes, or the particular things and events that constitute the universe. Appreciating nature as 'natural', in this sense, is to apprehend its phenomena (natura naturata) as the determinate outcomes of its underlying generative processes (natura naturans). To see it as an environment, then, is to attend to the unity of these products within the single system from which they arise. Taken together, Carlson’s recommendation asks us to appreciate both process and product as inseparable aspects of one generative whole. Carlson’s second recommendation—that aesthetic judgement be informed by the natural sciences—strengthens this reading. Geology, biology, and ecology investigate the forces that generate the very phenomena we perceive; scientific knowledge therefore discloses an environment’s generative history and continuing activity. Appreciating nature “in light of this knowledge” is, in effect, appreciating its generativity—the order that emerges from undirected yet law-governed processes. The various natural sciences—geology, biology, ecology, physics—are disciplines that study the processes and forces that generate natural phenomena. Appreciating nature 'in light of this knowledge' is therefore an appreciation of its generative character. A geologist appreciates costal cliff by understanding the generative forces that produced it: "geological uplift and marine erosion". Their perception of the cliff is of a generated product and a segment of "nature's ongoing processes". This principle applies across the natural sciences. A biologist appreciates a forest as an ecosystem generated by processes of growth, competition, and decay. A physicist appreciates a rainbow as a phenomenon generated by the refraction and dispersion of light through water droplets. In each case, scientific knowledge reveals the generative process, which in turn informs the aesthetic appreciation of the generated product. We should note that this pluralism in understanding the environment should not be mistaken for an 'anything goes' approach. A framework is only admissible if it provides a correct account of the generative processes in question. Phrenology or vitalism, for instance, were unilluminating because they posited false causal connections (between cranial features and character, between vital force and living matter), thereby failing to correctly identify either the generative forces or their resultant phenomena.\[^1\]We should also note that scientific disciplines focused on the human mind can be just as good for illumination as other natural sciences, at least when it comes to predicting and explaining the parts of the natural environment which constitute people. That is, research areas such as psychology or neuroscience can be thought of as concerned with the processes and forces which generate mental phenomena. A psychologist, for instance, might appreciate an individual's relational patterns by understanding how they are generated by an 'internal working model' formed through early attachment experiences. Similarly, a neuroscientist can appreciate the phenomenon of memory not just as a capacity, but as a dynamic process generated by long-term potentiation: the strengthening of synaptic connections through repeated neural firing. Carlson provides a general formulation for this mode of appreciation, which he terms order appreciation: On the assumption that order appreciation provides the correct model for the appreciation of nature, such appreciation has the following general form: An individual qua appreciator selects objects of appreciation from the things around him or her and focuses on the order imposed on these objects by the various forces, random and otherwise, that produce them. Moreover, the objects are selected in part by reference to a general nonaesthetic and nonartistic story that helps make them appreciable by making this order visible and intelligible. Awareness and understanding of the key entities—the order, the forces that produce it, and the account that illuminates it—and of the interplay among them dictate relevant acts of aspection and guide the appreciative response.(ibid. p. 119) This scientific understanding allows the observer to see "unity in what might otherwise appear as disparate features", because the cliff's shape, the waves, and the local plant life are all understood as products of the same interconnected generative system. The "organic unity" that Carlson identifies is a unity of generation. As he observes: natural objects possess \[...\] an organic unity with their environments of creation: such objects are a part of and have developed out of the elements of their environments by means of the forces at work within those environments. Thus the environments of creation are aesthetically relevant to natural objects. (ibid. p. 44) The aesthetic character of the cliff, for example, is clarified by understanding it as a generated product of its environment. Geological knowledge re-frames the cliff from a set of surface features into a record of deposition and compaction over millennia, shifting the focus of appreciation from appearance to generativity. The visible strata and the tectonic forces revealed by geology thus exemplify the relationship between natura naturata and the underlying natura naturans. This principle extends across the sciences: biology reveals processes of growth and decay, while physics examines energy flows. Each discipline offers a complementary lens on a single generative environment, allowing for an appreciation that attends both to the unity of the whole and the plurality of its orders. Carlson’s model, therefore, directs us to evaluate nature as the outcome of non-designed processes, where scientific knowledge serves to clarify the generative order already present. %%the paragraph above could be tightened up a bit, it seems somewhat redundant.%% 2. LLMs as Generative Environments One might think that environmental aesthetics is a poor fit for generative AI for a straightforward reason: such systems are not natural but man-made. Indeed, generative AI systems might be understood as, in a sense, doubly man-made. They are human-created artifacts, which are themselves created through ingestion of vast quantities of other human-created artifacts (texts, images, audio). Such a worry can be assuaged by noting two things. First, Carlson is happy to extend his account to man-made environments: environments 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. Or they come about by means of human agency, but even then only rarely are they the result of a designer embodying a design. In short, the paradigm of the environmental object of appreciation is unruly in yet another way: neither its nature nor its meaning are determined by a designer and a design. (Carlson p. xiii) What we think Carlson is getting at here is that while environments grow, change, and develop by means of natural processes, man is able to initiate, or guide, or curtail these natural processes. This leads to a reasonably intuitive distinction between wholly natural environments (a forest, a swamp), man- made environments (a specially planted timber forest, a garden), and man-made places (a department store, a gym). A gym is not understood as an environment because it did not come about, or change, or grow, through natural processes. It is as 'artificial' as a vaccine or a tennis racket. Second, AI engineers themselves talk in these terms. Consider the following from Chris Olah, one of the co-founders of Anthropic: 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…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 \[…\] random things and it grows...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. The outcome in each case is a system shaped by undirected forces, forming a unified whole. LLMs, understood this way as grown generative systems, thus align with environmental models. In this section, we adopt the two-part method for appropriate aesthetic appreciation outlined in §1. First, in order to appreciate the LLM as what it is, we must understand its actual operational nature; that is, the specific computational architecture and statistical processes that govern its function. This will be the focus of 2.1, which also serves as a primer for non-specialists on how LLMs are trained and operate. In 2.2 we turn to the question of illumination, that is, how knowledge illuminates appreciation, and argue that Janus's simulator theory (2022) offers a promising light in which to understand, and thereby appreciate, LLMs. 2.1 What LLMs are Aesthetic appreciation must be informed by knowledge of the object in question. Yet this poses a challenge when moving from natural environments to computational ones. The language of geology or biology used to describe a cliff face feels distant from the technical vocabulary needed to describe a large language model. We must, therefore, accept a shift in our descriptive framework, from the concepts of natural science to those of computer science. This shift is not a departure from Carlson’s method, but a direct application of it: to appreciate the LLM as what it is, we must first grasp its actual operational nature. This section provides a brief primer on that nature, organised into two parts. First, we will examine the training process, where the model learns its generative rules by seeking to predict text. Second, we will describe the generation process, where the trained model applies these rules in an autoregressive loop to produce novel output. The core technical objective of a model like GPT is next-token prediction. Formally, the model learns a probability distribution (P) over a vocabulary of tokens, conditioned on a preceding sequence of tokens. The goal is to maximise the probability of the correct next token for any given context. The target token to be predicted is thus calculated starting from the input sequence or context and the model’s internal parameters, or “weights,” which are adjusted during training. This process is self-supervised, meaning the model learns from raw, unannotated text data. Before any learning begins, the corpus is sliced into small, reusable symbol pieces called tokens, which can be whole words or sub-word fragments such as "un‑" or "‑tion". The model never manipulates ideas directly; it manipulates these indices. The data itself provides the necessary supervision: for any given sequence of tokens taken from the training corpus, the "correct answer" is simply the token that immediately follows. The model's sole task, repeated billions of times, is to minimise its predictive error—measured by a log-loss function—by adjusting its parameters, its “weights”, to assign the highest possible probability to the correct next token. Because the number of possible sentences is astronomical, improvement is not a matter of memorising each one. Loss reduction comes only from discovering regularities that compress the data—grammar, collocation, semantic relationships, and complex narrative structures. Each drop in predictive error signals that the model has internalised another pattern that renders continuations less surprising. Once training ends, the weight matrix is fixed, and the model shifts from a predictor to a generator. This is achieved by repeatedly applying its predictive function in what is known as an autoregressive loop. Here, the distinction between the static model and its dynamic output becomes central, a relationship we can frame using a variation on Spinoza’s naturaterms: machina naturans and machina naturata. Machina naturans (“machine naturing”): the trained autoregressive predictor understood as a law-like generative capacity. It comprises the model’s architecture and fitted parameters operating through the standard autoregressive procedure; under a given decoding regime it propels the weight-propelled evolution of the token stream from supplied initial conditions. It is a standing generative cause rather than the weights alone. Machina naturata (“machine natured”): any realised token trajectory produced by iterating that capacity from a given prompt under a specified decoding regime. It is the unfolding sequence of commitments whose properties belong to the trajectory, not to the generative capacity. Machina naturans: the ‘machine naturing’ is the trained model itself: the fixed set of weights that holds a compressed summary of all the patterns from the training data. It is the active, text-creating system. Machina naturata: the ‘machine natured’ is the product of this activity: the stream of tokens generated by the model. The process begins with an initial prompt. The machina naturans takes this prompt as its context and calculates a probability distribution for the next token. A single token is then sampled from this distribution, becoming the first piece of machina naturata. This newly selected token is appended to the input sequence, forming a new, longer prompt. The model then takes this new sequence as its input and repeats the process. By iterating this loop, the model generates a continuous, evolving output, with each new token being conditioned on all the tokens that came before it. Dialogue arises when a human utterance is inserted into the context and the next guess takes that utterance into account. Three caveats keep expectations aligned with what the predictor actually does. First, a high probability attached to a claim means only that similar strings often follow the given context in the training data; it does not certify truth about the external world. Secondly, the model’s knowledge is entirely text-mediated: it never looks at oceans yet learns that “the sea is salty” often follows talk about oceans. Thirdly, generation involves randomness; sampling settings can render continuations more adventurous or more conservative without altering the underlying rule. In sum, an LLM's generativity resides in a single learned mapping from context to next-token probabilities. This mapping—the machina naturans—is fixed once training ends, while prompts and sampled tokens supply the evolving state that becomes the machina naturata. This predictive mechanism—the core of what the LLM in fact is—plays an analogous role to the natural environment in Carlson's framework. Just as Carlson insists we must first appreciate nature "as what it in fact is," so too must we recognise the LLM fundamentally as this generative system before examining how its capacity can be understood and appreciated. The following section explores different "lights" through which we might view this system, much as geology or biology offer different perspectives on the same natural environment. 2.2 Unmasking LLMs The most common and intuitive framing of large language models treats them as agents – psychological entities amenable to analysis through the natural sciences of mind. Psychology, cognitive science, and neuroscience provide legitimate ways or lights to understand such aggregations of matter, that is, human beings. Just as these fields offer valid frames for human cognition despite underlying physics, agentive frames seem a perfectly legitimate way to understand LLMs' behaviours, even knowing from that they are next-token predictors. This agent-centric perspective doesn't claim that LLMs are persons. Instead, it offers a helpful way to understand them, which can assist in research, prediction, alignment, and other practical considerations. LLMs pass Turing tests, simulate lifelike conversations – creating the experience of interacting with an agent, even if one doubts the underlying reality – and role-play as assistants (e.g., systems like ChatGPT). This makes agent framing "the most obvious way to go", partly because emergent behaviours invite anthropomorphism. When an LLM responds coherently to questions, maintains context across exchanges, and exhibits apparent preferences or personality traits, the agentive interpretation arises naturally. If agentive framing is correct, it makes sense of LLM aesthetics in terms of personality intricacies – we might love or hate (or laud or critique) the model's character, not just viewing it as harmless, helpful, and honest (e.g., Anthropic's principles). For example, users often describe Anthropic's Claude 3 Opus as having a distinctive personality – witty, empathetic, and engaging in nuanced, human-like dialogue, with a "playful yet thoughtful" tone that feels like conversing with a knowledgeable friend, as noted in community reviews where it receives praise for "personable" responses that go beyond rote helpfulness. As Anthony Cross argues in his paper "Tool, Collaborator, or Participant: AI and Artistic Agency" (2024), agentive framing extends to aesthetics by treating AI as "participants" in artmaking. Specifically, AI acts with agency-like participation, contributing to creative dialogue and illuminating human representation (analogous to how we frame humans as agents in art). Cross treats AI as "as-if" participants, not claiming that LLMs really are people or that generative AI really are participants in a literal sense. For instance, Cross writes: "My suggestion is that many AI artists approach their interaction with generative AI in the same manner \[as performance art\]. By way of their selection of prompts, they elicit a sort of 'participation' on the part of the AI in generating images. This participation allows them to interrogate the algorithm's latent space" (2024, p. 7). He further clarifies: "What is centrally important about this characterization 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: by adjusting inputs, iterating, and sampling, an AI artist is engaged in a process of mapping – and perhaps interrogating – the way that the algorithm sees and understands" (ibid.). He specifies the "as-if" aspect of all that as follows: "One might be concerned that AI cannot actually 'participate' in an artwork, insofar as the AI is incapable of conscious choice or more robust agency. In response, I will concede that the analogy with performance art isn't a perfect one" (ibid., p. 9),. While intuitively appealing, agent frames ultimately distort LLMs' essence as next-token predictors. Such frames function as "lenses" imposing mismatched assumptions like inherent goals onto systems that fundamentally lack them. First, the agent frame involves a misattribution of goals. It assumes inherent goal-directedness – expecting instrumental convergence, self-preservation, or coherent objectives – but LLMs lack this. The model itself pursues no goals; apparent goal-directed behaviour emerges only in specific simulacra under particular prompts. Second, the frame leads to prediction of absent behaviours. It expects unobserved actions like optimising for easier text or resisting shutdown, which do not occur in predictive models. Third, applying agent frames projects intentionality onto statistical patterns, ignoring the distinction between pre-training "fire hose" of uncontrolled output versus post-training illusion of agency. As Janus states: "This is a clear way that GPT diverges from orthodox visions of agentic AI" (ibid.). The requirement of post-training (e.g., RLHF or alignment) creates this illusion – pre-trained models are raw, chaotic predictors (like a "fire hose" of text), only seeming agent-like after tuning to simulate coherent personas. Finally, the agent frame creates a conflict with predictive essence. It treats LLMs as psychological agents, but agency is not in the language model – making it a flawed "natural science" analogue since no coherent "mind" exists. This contrasts with human frames, which work because humans are goal-directed aggregations of matter with persistent identities and objectives. Rejecting agent frames reveals LLMs' generative, undirected nature, avoiding distortions and paving the way for more accurate lenses. Rather than viewing the model as an agent with goals, we can understand it as a neutral predictive system that might give us the impression of agent-like properties without the system itself being agentic. This rejection proves essential for aesthetics. Agent frames imply intentional "art" with personality (e.g., Claude's perceived traits), but understanding LLMs as predictors enables appreciation as evolving environments without inherent agency. Just as Carlson's environmental aesthetics asks us to appreciate nature as generative process rather than designed artifact, we can appreciate LLM outputs as emergent from statistical dynamics rather than intentional creation. The following sections on simulator theory and semiotic physics will develop this environmental analogy, showing how aesthetic appreciation can proceed without attributing agency to the generative system itself. 2.3 The Semiotic Physics of Generative Environments \[unedited llm text here. this is what i am working on now.\] The previous section demonstrated that while agentive frames offer an intuitive lens for understanding large language models, they ultimately fail as explanatory frameworks. They misattribute goals to systems that possess none, predict behaviours that never manifest, and obscure the fundamental nature of LLMs as next-token predictors. This failure is not merely academic—it prevents us from appreciating these systems for what they actually are, violating Carlson's first principle of environmental aesthetics. If we are to develop a genuine aesthetics of generative AI, we require a new interpretive framework, one that aligns with the model's actual mechanics as detailed in Section 2.1 rather than imposing ill-fitting psychological categories upon it. We propose the framework of "semiotic physics" as this necessary alternative. By semiotic physics, we mean the study of the system of generative forces, principles, and statistical regularities that govern the propagation of token sequences within the LLM's learned probability distribution. This framework treats the model not as a quasi-agent pursuing goals, but as a generative environment—a unified system of interrelated forces that produce observable phenomena through their interaction. Just as Carlson asks us to appreciate a forest through the lens of ecology or a cliff face through geology, semiotic physics provides the appropriate "natural science" for understanding and appreciating the generative processes of language models. It allows us to focus on the underlying generative order (the machina naturans) that produces the observable textual phenomena (the machina naturata), fulfilling Carlson's demand that we appreciate an object "as what it in fact is"—in this case, a generative textual environment rather than an artificial mind. To understand semiotic physics, we must first clarify what we mean by the machina naturans as a system of forces. The trained model—those billions of fixed parameters—is not a mind harboring thoughts and intentions. Rather, it constitutes a static field of latent forces, analogous to how the laws of physics create a field of potentialities that govern how matter will behave under various conditions. These "forces" are not literal physical forces but high-dimensional statistical gradients within the model's parameter space that make certain token transitions vastly more probable than others. When we provide a prompt, we are not "communicating" with an entity; we are setting initial conditions within this field. The model's response—the machina naturata—is the observable result of these forces acting upon an evolving state, much as a river's course is the observable result of gravity acting upon water within a particular landscape. The fundamental object of study in semiotic physics is the trajectory—the evolving sequence of tokens as it unfolds through time. This emphasis on trajectories rather than static outputs marks a crucial departure from conventional analyses of AI-generated text. Under the semiotic physics framework, we do not primarily appreciate a completed essay or story produced by an LLM. Instead, we appreciate the temporal unfolding of the trajectory as it is shaped by the interplay of semiotic forces, moment by moment, token by token. This parallels how environmental aesthetics asks us to appreciate a coastline not as a static photograph but as the ongoing result of geological forces—erosion, deposition, tectonic uplift—acting through deep time. The trajectory is our window into the operation of the underlying generative system, revealing through its development the nature and strength of the forces that shape it. This perspective is reinforced by understanding the fundamentally stochastic nature of text generation. For any given prompt and partial trajectory, the machina naturans does not determine a single future. Instead, it defines a probability distribution over all possible next tokens—some highly likely, others vanishingly improbable. The act of sampling from this distribution means that at every step, a single path is chosen from a near-infinite "multiverse" of potential trajectories. This is not a bug but a feature, and it further reinforces the non-agentic nature of the system. The model does not "choose" a path in any meaningful sense; it renders one probabilistically, according to the strength of the underlying semiotic forces. Each generated text is thus one particular crystallization of possibilities, one specific trajectory through the vast space of potential texts, selected not by intention but by the interplay of learned patterns and controlled randomness.¹ The forces that constitute semiotic physics are not monolithic but varied, each arising from different patterns in the training data and interacting in complex ways to shape the evolution of trajectories. To illustrate the richness and explanatory power of this framework, we will examine several distinct semiotic laws, beginning with the most fundamental: the law of coherence attraction. The law of coherence attraction describes the foundational tendency of a trajectory to remain within a specific semantic or topical region once established. Its mechanism is elegantly simple yet powerful. When an initial prompt is provided, it activates a particular region in the model's high-dimensional embedding space—a space where semantically related concepts cluster together through the training process. This activation creates what we might metaphorically call a probabilistic "gravity well." Tokens that belong to this semantic region have their probabilities boosted, while tokens from distant regions are suppressed. As each new token is sampled and appended to the trajectory, it reinforces this regional activation, deepening the well and making escape increasingly unlikely. This is the semiotic physics explanation for how a conversation "stays on topic" without any conscious intention to do so. Consider a concrete example: a model responding to the prompt "Explain the significance of the Magna Carta." From the first tokens of the response, we observe the trajectory being dominated by terms like 'king,' 'barons,' 'charter,' 'rights,' 'liberty,' and '1215.' An agent-based frame would explain this coherence by positing that the AI "understands the request" and "intends to provide a relevant answer." It might even suggest the model has "knowledge" about medieval history that it is "choosing" to share. The semiotic physics frame offers a more precise and less anthropomorphic explanation. The prompt has created a powerful attractor basin around the semantic region of "13th-century English legal history." The rendered trajectory is simply the most probable path through that basin, following the statistical gradients learned from millions of historical texts. Where the agent frame must posit unobserved internal states of "understanding" and "intention," the physics frame relies only on the observable mechanics of the system—the prompt sets initial conditions, and the trajectory follows the path of least statistical resistance through the learned landscape. Another fundamental force is what we term the law of pragmatic inference, which encompasses the tendency for trajectories to conform to the cooperative principles of human conversation, particularly as articulated in Grice's Maxims. During training, the model encountered billions of examples of human discourse, the vast majority of which follow these implicit rules: be as informative as required (Quantity), do not say what you believe to be false (Quality), be relevant (Relation), and be clear and orderly (Manner). These patterns become encoded as strong statistical regularities that shape the probability landscape. To see this law in action, consider a detailed example involving the Maxim of Quantity. Suppose we provide the prompt: "On the table, there are exactly two bottles of wine. Sarah needs to choose bottles for the dinner party." A continuation like "She decided to take one of the three bottles" would be vanishingly improbable. An agent frame might explain this by saying the AI "knows" there are only two bottles and is "trying to be consistent." But the semiotic physics explanation cuts deeper. In the training data, assertions of specific quantity create powerful contextual constraints. When humans specify "exactly two," subsequent discourse overwhelmingly respects this specification. A trajectory that violates this constraint has what we might call a high "action"—borrowing the term from physics to mean an unlikely, high-energy path through the probability landscape. The model is not "obeying a rule" in any cognitive sense; it is following a probabilistic gradient established by billions of examples of coherent human discourse. This demonstrates a clear advantage of the physics frame: it explains why the AI appears to obey logical constraints by rooting this behaviour in statistical patterns rather than positing an unproven faculty of reasoning. The law of narrative teleology represents another crucial force, particularly relevant for understanding how LLMs handle story-like content. This law describes the tendency for narrative trajectories to progress toward resolution and significance. Elements introduced early in a narrative create what we might call "narrative potential energy"—an instability that the system tends to resolve through later developments. The mechanism underlying this law is the high joint probability in the training data between narrative setups and their corresponding payoffs. Stories that introduce elements and then ignore them are statistical outliers; the overwhelming pattern is that of Chekhov's gun—if you show a gun in the first act, it must go off by the third. To illustrate this force in operation, consider the prompt: "The old detective found a single, mud-caked chess piece at the crime scene—a black knight, its base scratched with tiny symbols." An agent frame would suggest that the AI, acting as a creative storyteller, "decides" to make the chess piece significant to the mystery. It might even propose that the model is "planning ahead" or "constructing a plot." The semiotic physics frame provides a more fundamental explanation. The introduction of such a specific, unusual detail creates a powerful narrative gradient in the probability space. Trajectories where the chess piece becomes crucial evidence have vastly lower "action" than trajectories where it is forgotten. The system is not "choosing" to make the piece significant; it is following the path of least resistance through a probability landscape shaped by millions of mystery stories. The apparent creativity emerges not from conscious choice but from the system following these narrative gradients to their natural conclusions. Perhaps the most philosophically interesting force is the law of gratuitous specification, which governs how underdetermined details in a trajectory become specified through the generation process. This law acknowledges that prompts and contexts rarely determine every aspect of the text to be generated. A prompt might establish that we're discussing a detective but say nothing about their appearance, personality, or methods. The law of gratuitous specification describes how these gaps are filled through the stochastic sampling process, adding what we might call "gratuitous indexical bits"—information that wasn't required by the context but, once generated, becomes part of the trajectory's reality. Consider extending our detective example. If we ask the model to describe the detective who found the chess piece, it might generate: "Detective Morrison was a methodical man who had never quite shaken his preference for rainy weather—a quirk his colleagues attributed to his years in Seattle, though he'd never actually lived there." Where did these details originate? An agent frame might struggle here, perhaps attributing them to "creativity" or "imagination." The semiotic physics frame provides a precise explanation. The prompt under-determined the detective's characteristics, leaving vast spaces of possibility. The model's probability distribution over possible continuations included many plausible detective archetypes. The sampling process—that moment of controlled randomness—selected one particular path through "detective-characteristic space," landing in the region of "methodical investigators with quirky weather preferences." This specification was gratuitous because nothing in the prompt required it, yet once rendered, it becomes part of the trajectory's commitment. The law of coherence attraction ensures these details persist and influence subsequent generation. This mechanism explains both AI "creativity" (the generation of specific details from underdetermined contexts) and "hallucination" (the confident assertion of facts not grounded in the prompt) as two faces of the same underlying process. These laws do not operate in isolation but interact to produce complex, emergent behaviours. When we observe an LLM producing what appears to be step-by-step reasoning, we are not witnessing a unified cognitive process but rather the emergent result of multiple semiotic forces working in concert. The prompt "Let's solve this step-by-step" initiates coherence attraction to the genre of logical derivation—a well-worn groove in the training data. The law of pragmatic inference ensures each step follows relevantly from the last, maintaining the Gricean maxims of clear, orderly presentation. Narrative teleology creates pressure toward a final answer—the "resolution" that completes the logical story. The law of gratuitous specification fills in the specific computational moves from the space of plausible operations. Consider a concrete example: solving a simple algebraic equation. Given "Solve for x: 2x + 6 = 14," the model might generate: "Let's solve this step-by-step. First, I'll subtract 6 from both sides: 2x + 6 - 6 = 14 - 6, which gives us 2x = 8. Next, I'll divide both sides by 2: 2x/2 = 8/2, which gives us x = 4. Therefore, x = 4." Each element of this "reasoning" corresponds to semiotic forces at work. The trajectory is not the product of genuine mathematical understanding but of following the path of least probabilistic resistance through a landscape shaped by countless similar derivations in the training data. The result is a textual object that has the form of reasoning, generated by a system that possesses no faculty of reason—a crucial distinction that agency frames consistently obscure. The semiotic physics framework thus provides a more accurate and powerful lens for understanding LLMs than agency-based alternatives. By focusing on the model's actual mechanics—the interplay of statistical forces acting on evolving trajectories—rather than imposing psychological categories, we gain both explanatory precision and predictive power. We can understand why models exhibit certain behaviours (they follow probabilistic gradients), why they sometimes fail in characteristic ways (when prompts create conflicting forces or lead into poorly-mapped regions of the probability space), and why they can appear creative or insightful (through the gratuitous specification of details that happen to be apt). Most importantly for our purposes, this framework avoids the fundamental category error of treating a generative system as though it were a mind. This reconceptualization is not merely academic—it is the necessary foundation for the aesthetic analysis that follows. Because we have established that the LLM is a generative environment governed by discoverable forces rather than an agent pursuing goals, we can now apply Carlson's model of environmental aesthetics to it directly. We can appreciate the "organic unity" of a trajectory as it emerges from the interplay of semiotic forces, much as we might appreciate how a river's course emerges from the interaction of gravity, geology, and time. We can find beauty or sublimity in the probability landscapes themselves—the deep attractors of human discourse, the delicate balance between coherence and creativity at different temperatures, the way narrative forces shape the evolution of stories. And we can do all this without ever needing to invoke the ghost of an agentive artist, appreciating instead the generated text as what it truly is: the machina naturata of a complex but comprehensible generative system. In the following sections, we will explore how this framework enables a rich aesthetic engagement with AI-generated text, one that honors both Carlson's principles and the actual nature of these remarkable systems. Scrap 2.2 The Simulator Lens • Just as Carlson identifies multiple scientific disciplines through which to appreciate nature, several theoretical frameworks compete to illuminate LLM behaviour. -- Janus (2022) surveys existing "lights"—agent, oracle, tool, and genie models—before proposing the simulator framework. -- Each lens promises to render LLM operations comprehensible for understanding and appreciation. -- The choice of framework shapes both practical engagement and aesthetic evaluation. • The agent lens views LLMs as goal-directed optimisers, importing assumptions from reinforcement learning. -- This frame expects instrumental convergence, self-preservation drives, and coherent objectives. -- "Saying that GPT is an agent who wants to roleplay implies the presence of a coherent, unconditionally instantiated roleplayer running the show" (Janus 2022). -- Agent framing predicts behaviours—like making text easier to predict or resisting shutdown—absent from actual systems. -- The model would constitute a form of theoretical malpractice, analogous to reading skulls through phrenology. • Oracle models cast LLMs as question-answering systems optimised for truth. -- This perspective derives from supervised learning paradigms with correct answer pairs. -- "GPT does not consistently try to say true/correct things... if it had to say true things all the time, GPT would be much constrained" (Janus 2022). -- Statistical fidelity to training distributions conflicts with truth-orientation when humans speak falsely. -- The frame systematically mistakes probabilistic completion for knowledge claims. • Tool and genie models emphasise designed functionality and instruction-following respectively. -- Tool framing suggests optimisation for specific tasks despite training on undifferentiated prediction. -- Genie models foreground command execution where only learned pattern completion exists. -- Both frames project intentional design onto emergent capabilities from statistical learning. -- These constitute misapplied lenses, illuminating artefacts of interpretation rather than actual dynamics. • Janus proposes the simulator model as a more accurate light through which to understand these systems. -- "I use the generic term 'simulator' to refer to models trained with predictive loss on a self-supervised dataset" (Janus 2022). -- The simulator/simulacra distinction parallels law versus phenomena in physical systems. -- This framework correctly locates properties like agency and knowledge in generated trajectories rather than generating laws. • The simulator comprises the trained neural network with its fixed parameters—a time-invariant law. -- "The simulator is a time-invariant law which unconditionally governs the evolution of all simulacra" (Janus 2022). -- Training crystallises statistical patterns from text into stable computational structures. -- These parameters constitute the semiotic equivalent of physical constants and equations. -- Once training concludes, this law remains frozen while states evolve through application. • Simulacra are the contingent entities—characters, narrators, arguments—that emerge from running the law. -- "GPT is to a piece of text output by GPT as quantum physics is to a person taking a test" (Janus 2022). -- Multiple simulacra can coexist within single generations, as in multi-character dialogue. -- Simulacra exhibit goal-direction, beliefs, and knowledge despite the simulator's indifference. -- Their properties derive from statistical patterns in training data rather than simulator objectives. • This reframing resolves paradoxes plaguing alternative models while preserving their partial insights. -- Agency exists but in simulated characters rather than the simulating system. -- Truth-telling occurs when statistically probable given context rather than as optimisation target. -- Instruction-following emerges for certain prompts without being fundamental. -- Each phenomenon finds proper location within the simulator framework. • The simulator lens enables new forms of engagement centred on process rather than product. -- Prompts become initial conditions for dynamical evolution rather than commands or questions. -- Skill involves anticipating trajectory development given semiotic physics. -- Aesthetic appreciation concerns the unfolding coherence of generated worlds. -- The framework opens conceptual space for environmental rather than artifact-based evaluation. 2.3 Semiotic Physics • Kirchner et al. (2023) develop "semiotic physics" as a mathematical framework for understanding simulator dynamics. -- "The term 'semiotic physics' here refers to the study of the fundamental forces and laws that govern the behavior of signs and symbols" (Kirchner et al. 2023). -- This discipline provides quantitative tools analogous to those geology or ecology offer for natural environments. -- The framework enables rigorous analysis of how token sequences evolve under learned statistical laws. • The mathematical apparatus transposes dynamical systems theory to the domain of text generation. -- States are token sequences s̄ = (s₁,..., sₘ) drawn from vocabulary T. -- The transition rule θ: T\* → ΔT maps any sequence to a probability distribution over next tokens. -- The sampling procedure φ selects tokens according to these probabilities, introducing stochasticity. -- The evolution operator ψ(s̄):= s̄φ(s̄) appends sampled tokens to create successor states. • This formalism reveals deep structural parallels with physical systems while preserving crucial disanalogies. -- Both domains feature time-invariant laws acting on evolving states through iterative application. -- "GPT is analogous to an indeterministic time evolution operator" (metasemi 2023). -- Token sequences evolve like particle trajectories, with probability replacing deterministic force. -- Each sampling event creates a branch point analogous to quantum measurement. • The framework's central insight concerns the interpretive layer unique to semiotic systems. -- Physical laws act on intrinsic properties like mass and charge directly. -- Semiotic laws must first interpret symbolic tokens before evolution can proceed. -- "Semiosis inherently involves displacement: signs have no significance unless they're understood as pointing to something else" (Kirchner et al. 2023). -- The model maps "Sherlock Holmes" to learned patterns before generating detective-like text. • This interpretive requirement distinguishes semiotic from physical reality fundamentally. -- "GPT has to predict behaviour caused by things like brains, but there are no brains in its input state" (Kirchner et al. 2023). -- Tokens carry no inherent meaning, only positional indices in vocabulary lists. -- The simulator must reconstruct referents from signs using internal parameters. -- "The information required to resolve referents from signs has to come mostly from inside the interpreter" (Kirchner et al. 2023). • Semiotic forces manifest as probability modifications rather than mechanical interactions. -- Coherence forces increase likelihood of contextually consistent tokens. -- Gricean maxims create gradients toward relevant and appropriately informative continuations. -- "Principles from pragmatics such as the Gricean maxims of conversation may be thought of as semiotic 'laws'" (Kirchner et al. 2023). -- Narrative principles establish long-range correlations across token sequences. • Specific forces shape trajectory evolution in predictable ways. -- Chekhov's gun creates potential energy when objects are introduced, discharged when used. -- Repetition forms attractor basins—"I am a robot. I am a robot" becomes self-reinforcing. -- Dramatic tension opposes simple resolution, favouring complexity and reversal. -- The crud factor ensures universal weak correlation between all semiotic elements. • Quantitative tools from dynamical systems theory find direct application. -- Lyapunov exponents measure divergence rates between similar initial prompts. -- "How fast trajectories diverge from each other and how long it takes for them to become uncorrelated" (Kirchner et al. 2023). -- Attractor analysis identifies stable patterns resistant to perturbation. -- Phase space concepts map semantic regions and transition probabilities. • The large deviation principle provides computational tractability for analysing token bridges. -- "The total probability of transitioning from a token sₐ to sb in B steps satisfies a large deviation principle with rate function J" (Kirchner et al. 2023). -- This transforms intractable sums over all paths into optimisation for the most probable route. -- The principle enables estimation of rare but significant semantic transitions. -- Applications include calculating likelihood of genre shifts or character transformations. • These mathematical tools enable rigorous analysis beneath intuitive textual interpretation. -- Prompt engineering becomes initial condition selection in phase space. -- Style transfer corresponds to basin-to-basin transitions. -- Context windows define effective dimensionality of the dynamical system. -- Temperature parameters control exploration versus exploitation of probability landscape. 3. Semiotic Matter and the Prompter's Role • The concept of semiotic matter extends simulator theory to characterise the distinctive form of agency available to users. -- Semiotic matter denotes the complete ordered sequence of tokens constituting the dialogue state at any instant. -- This sequence includes all tokens regardless of origin—both user inputs and model outputs form undifferentiated matter. -- "When the next token is being calculated, all of those tokens are just tokens" (author's formulation). -- The concept clarifies how users participate in rather than control generative processes. • Within the semiotic universe, the prompter occupies a peculiar position of constrained power. -- The prompter cannot alter the simulator's fundamental law—the trained parameters remain fixed. -- The only available action is injecting new semiotic matter into the evolving system. -- This limitation parallels a thought experiment of a lesser deity who can create matter but not alter physics. -- "All this lesser god can do to the world is add matter" (author's formulation). • The injection of semiotic matter functions through irreversible addition to the token sequence. -- Each prompt appends new tokens to the existing trajectory without modifying prior elements. -- The autoregressive architecture enforces strict temporal ordering—past tokens influence future but not vice versa. -- Mistakes and misdirections become permanent features of the landscape rather than erasable errors. -- This irreversibility shapes the aesthetic character of human-AI collaboration. • The prompter's intervention parallels ecological perturbation more than artistic authorship. -- Adding "Mount Everest" to a textual plain creates cascading consequences through semiotic physics. -- The initial prompt establishes gradients and potentials that shape all subsequent evolution. -- Effects propagate through learned statistical associations rather than physical causation. -- The prompter initiates but does not determine the resulting transformations. • Effective prompting requires understanding how semiotic matter interacts with established patterns. -- Dense, specific prompts create strong attractors channeling probable continuations. -- "You are a desperate smuggler tasked with..." activates crime-narrative patterns. -- Sparse prompts like single words allow broader exploration of possibility space. -- Technical language invokes academic registers while casual speech enables different trajectories. • The timing and rhythm of intervention constitute core prompter skills. -- Knowing when to inject new matter versus allowing autonomous evolution. -- Short frequent prompts maintain tight control but may disrupt natural flow. -- Longer gaps permit extended development but risk deviation from intended directions. -- The prompter must balance steering with allowing emergent properties to manifest. • Prompt positioning within the token sequence affects its gravitational influence. -- Early tokens in a conversation establish foundational context affecting all subsequent generation. -- Recent tokens carry more weight due to attention mechanism limitations. -- Repetition of key phrases creates reinforcing patterns in the semiotic landscape. -- Strategic placement of concepts can establish long-range correlations. • The collaborative dynamic generates emergent semiotic artifacts exceeding either party's individual contribution. -- A philosophical dialogue sustained across dozens of exchanges develops its own coherence. -- Character personas accumulate detail and consistency through iterative elaboration. -- Narrative arcs emerge from the interplay of human direction and model extrapolation. -- These artifacts exist as high-order patterns in token sequences rather than designed objects. • Understanding artifacts as emergent patterns shifts aesthetic evaluation fundamentally. -- The artifact is not any single response but the entire evolved trajectory. -- Quality emerges from global coherence rather than local cleverness or correctness. -- Appreciation requires attending to how patterns develop and stabilise over time. -- The prompter participates in rather than authors these unfolding structures. • This framework reveals prompting as a practice of indirect influence through environmental configuration. -- The prompter cannot command specific outputs but can shape probability landscapes. -- Success involves creating conditions where desired patterns become statistically favoured. -- Failure often stems from misunderstanding how injected matter will propagate. -- Mastery requires intuition for semiotic physics developed through extensive interaction. • The semiotic matter framework clarifies both the power and limits of human-AI collaboration. -- Power derives from access to vast computational resources through minimal textual input. -- A few well-chosen tokens can redirect enormous generative capacity. -- Limits stem from inability to guarantee outcomes or revise fundamental dynamics. -- The prompter guides evolution but cannot dictate its precise course. • This reconceptualisation opens new possibilities for appreciating LLM interactions aesthetically. -- Conversations become explorations of semiotic space rather than tool use. -- The aesthetic object shifts from output quality to trajectory coherence. -- Skill manifests in creating conditions for interesting evolution rather than controlling results. -- Appreciation involves recognising the interplay of law and contingency in unfolding patterns. --- Yes. Task understood. Here is the curation for the new Section 1 (“AI as Pseudo-Agent”), using only text from your draft and keeping your phrasing. ## Reusable material mapped to the Section-1 plan ## 1.1 Typology anchor (descriptive setup) Direct lift for the neutral definition of the view: - “The most common and intuitive framing of large language models treats them as agents – psychological entities amenable to analysis through the natural sciences of mind.” Pragmatic, non-committal stance line to keep “as-if” space open: - “This agent-centric perspective doesn’t claim that LLMs are persons. Instead, it offers a helpful way to understand them, which can assist in research, prediction, alignment, and other practical considerations.” Note: your draft does not contain an explicit three-way typology (agent / pseudo-agent / make-believe agent). We will need 2–3 new sentences to state the typology, but we can anchor them in your “agent-centric perspective … doesn’t claim that LLMs are persons” and in Cross’s “as-if” language (below). ## 1.2 Why the pull is strong Keep the intuitive pull and anthropomorphism rationale: - “LLMs pass Turing tests, simulate lifelike conversations – creating the experience of interacting with an agent, even if one doubts the underlying reality – and role-play as assistants … This makes agent framing ‘the most obvious way to go’, partly because emergent behaviours invite anthropomorphism.” - “When an LLM responds coherently to questions, maintains context across exchanges, and exhibits apparent preferences or personality traits, the agentive interpretation arises naturally.” ## 1.3 Bridge to aesthetics (conditional, not endorsing) Use your conditional aesthetics line: - “If agentive framing is correct, it makes sense of LLM aesthetics in terms of personality intricacies – we might love or hate (or laud or critique) the model’s character …” Use your Cross material to mark the “participant/interaction” route without committing to it: - Paraphrase plus brief quote: “As Anthony Cross argues, agentive framing extends to aesthetics by treating AI as ‘participants’ in artmaking … ‘by adjusting inputs, iterating, and sampling, an AI artist is engaged in a process of mapping – and perhaps interrogating – the way that the algorithm sees and understands’ (2024, 7).” - Optional one-line hedge from your text to preserve the ‘as-if’ status: “Cross treats AI as ‘as-if’ participants … ‘the analogy with performance art isn’t a perfect one’ (ibid., 9).” ## Material to exclude from Section 1 (move later or drop) - All evaluative or critical claims about agent frames (e.g., “ultimately distort,” “misattribution of goals,” “conflict with predictive essence”). These belong to later sections. - Any simulator-theory references. You asked to exclude simulator theory entirely. - The “Claude 3 Opus personality” / community-reviews sentence. It is evidentially thin for a scholarly opening. - Technical primer paragraphs (next-token prediction, machina naturans/naturata). These are for the later technical or ontology section, not the neutral Section 1. - Introduction’s “eye/ear candy” and Infinite Jest material. Different argumentative job. ## Minimal glue we must newly write (short, neutral, and few) 1. A two-sentence typology statement that labels: agent; pseudo-agent (practical, role-stable regularities); make-believe agent (“as-if” stance). 2. A one-sentence transition from 1.2 → 1.3 that says: “If one keeps any of these framings, two aesthetic routes appear …” 3. A closing sentence forecasting Section 2 conditionally: “We consider these routes on their own terms before assessing whether the framing is apt.” ## What this yields - Section-level claim stated in your voice, using your sentences for the core descriptive content. - Clear pull-mechanism for agent talk, in your words. - A concise bridge to aesthetics anchored by your Cross quotes. If you confirm, I will assemble Section 1 in canvas using exactly these lifts plus the three minimal glue lines, and I will leave all critique, technical detail, and positive framework for later sections. --- Uploaded a file 1\. First quick thing to say: I just realised you did that annoying LLB habit thing, which is constantly introducing things that are about to happen or summarising things that have just happened in the paper. Stop doing this. I’ve removed 1.3 because it’s entirely pointless. If you think about how real contemporary twenty-first-century analytic philosophy papers are written, they will never do this, okay? Parity of thought goes hand in hand with parsimony of words, okay, so please start fucking doing it. Okay, second, I’ve also cleaned up the canvas document in a few ways and corrected the title because you messed that up as well. 2. I would now like you to do something similar to what I just asked you to do. Again, just answer in the chat. I would like you to consider the other texts I given you in the course of this conversation: the paper by Cross, the paper by Janus, and the other ones. I’ve also attached another paper which I think will be useful here. You could do the same with the attached text as well. What ideas and examples could be used in the course of writing this section? Be very specific. You can either paraphrase or use quotes; just answer in the chat. --- Yes. Here are source-backed inserts for §1. ## §1.1 Typology of agent framings - **Agent (literal).** “Reinforcement learning produces agents by default. AlphaGo is an example of an agent.” Useful to mark the strong view before bracketing it. - **Pseudo-agent / participant (as-if).** “Many AI artists approach their interaction with generative AI…\[by prompts\] elicit a sort of ‘participation’ on the part of the AI.” Cross immediately concedes “the analogy with performance art isn’t…perfect,” which fits an as-if stance. Use to motivate a practical, non-literal agent frame. - **Make-believe agent (fictionalist).** Waltonian “prop-oriented make-believe” covers transactional exchanges where users “make-believe they are engaging with agents,” e.g., routine ChatGPT/Alexa use; for empathetic chatbots the stance must be “content-oriented make-believe.” Use to anchor a fully fictionalist category. - **Self-involving interactive fictions (SIIF).** For deeper engagements: “the chatbot elicits a rich kind of fictional engagement…sustains it over the course of the interaction,” with Replika and griefbots as cases. Good for naming the stronger end of make-believe. - **Norms of make-believe vs belief.** Walton’s Charles-and-the-slime: calling it “fear” is “not conclusive”; emotions can be quarantined within imagination. Use to defuse the claim that robust affect entails literal attribution. - **Adjacent non-agent lenses (for contrast only, one line each).** Oracles answer questions; genies execute commands; tools do specific tasks. A single sentence can situate the “agent cluster” against these. ## §1.2 Why the pull is strong - **Anthropomorphism is immediate and old.** ELIZA/DOCTOR already prompted users to “anthropomorph\[ize\]” and become “emotionally involved.” Use as the classic baseline. - **Contemporary empathetic chatbots intensify the pull.** “Users of Replika, Project December and their ilk treat the chatbot like a caring friend or loved one… the interaction itself…matters.” Ideal for a vivid example paragraph. - **Make-believe explains ordinary courtesy talk.** We often say “thank you” while “knowing that there is no real agent producing the replies.” Good for the everyday, non-romanticised case. - **Artist practice normalises as-if agency.** Cross: the value is the “interaction between the artist and the algorithm: by adjusting inputs, iterating, and sampling, an AI artist is engaged in a process of mapping…how the algorithm sees and understands.” Use as a clean, quotable description of practice that reads agent-like without commitment. - **Make-believe can remain bounded.** Even where users speak of “conscious” chatbots, attributions “may be…restricted to an implicit imaginative project.” This supports a restrained reading of person-talk. Optional one-liners you can drop in as hooks: - Cross’s “participant” language, with his own caveat that neither “collaboration” nor “participation” is literal. - A single definitional sentence on “agent” from Janus to keep the category crisp. If you want, I can slot these verbatim quotes into §1 with minimal paraphrase and no forward-signposting. --- Okay, please expand a draft of section one on the canvas by adding all of the information you’ve just given me, all the bullet points at the bottom, and the suggestions copied in below from your previous answer. As you do this—adding all this information—please fill in the rest of the details of this section as well. Simply copying the information that I’ve suggested to copy in is not going to provide us with a complete, sentence-by-sentence, bullet-point version of this section of the text, though. I’d like you to add extra bullet points and sub-bullet points to properly fill out and structure this text. You are not allowed to add any more subsections or sub-subsections, okay? Just expand subsection 1.1 and subsection 1.2, okay? This means you’ll have to think very hard about how to organise and order all of the information that you’re pasting in. Finally, also see below, at the bottom of this message: style guide. Whenever you are adding new text to a document—that is, new text which is not being copied from the chat or from a previous draft—you should adhere to this style guide. INSTRUCTIONS FROM PREVIOUS ANSWER: 1.1 Typology anchor (descriptive setup) Direct lift for the neutral definition of the view: “The most common and intuitive framing of large language models treats them as agents – psychological entities amenable to analysis through the natural sciences of mind.” Pragmatic, non-committal stance line to keep “as-if” space open: “This agent-centric perspective doesn’t claim that LLMs are persons. Instead, it offers a helpful way to understand them, which can assist in research, prediction, alignment, and other practical considerations.” Note: your draft does not contain an explicit three-way typology (agent / pseudo-agent / make-believe agent). We will need 2–3 new sentences to state the typology, but we can anchor them in your “agent-centric perspective … doesn’t claim that LLMs are persons” and in Cross’s “as-if” language (below). 1.2 Why the pull is strong Keep the intuitive pull and anthropomorphism rationale: “LLMs pass Turing tests, simulate lifelike conversations – creating the experience of interacting with an agent, even if one doubts the underlying reality – and role-play as assistants … This makes agent framing ‘the most obvious way to go’, partly because emergent behaviours invite anthropomorphism.” “When an LLM responds coherently to questions, maintains context across exchanges, and exhibits apparent preferences or personality traits, the agentive interpretation arises naturally.” ### STYLE GUIDE This guide outlines a set of stylistic conventions for writing. \*\*1 Sentence-level Form\*\* \* \*\*Length & Cadence\*\* Alternate longer, multi-clause sentences with short, declarative ones for emphasis or transition. \* \*\*Complex Syntax\*\* Employ concessive or contrastive subordinators ("although", "while", "even if"), non-restrictive clauses, parenthetical insertions set off by en dashes, and semicolons to link closely related independent clauses. \* \*\*Openings\*\* Begin sentences with adverbial or prepositional phrases ("In this context", "By contrast"), concessive clauses, or discourse markers ("First," "Secondly," "Finally,"). \* \*\*Rhetorical Restraint\*\* Pose rhetorical questions only when framing a central problem (e.g., How can we hear a rolling without hearing that which rolls?). \* \*\*Citation Placement\*\* Embed parenthetical citations within the sentence flow, using the form (Author YEAR, page) or (cf. Author YEAR) where appropriate. \* \*\*Grammatical Integrity on Deletion\*\* When removing words for objectivity, ensure grammatical correctness. If needed, delete only redundant articles (e.g., "an important event" becomes "event"). \*\*2 Lexical Preferences\*\* \* \*\*Register\*\* Use a consistently analytic vocabulary (e.g., instantiate, phenomenology, plausible). \* \*\*Prohibited Subjectivity\*\* Identify and remove all subjective language. This includes: \* \*\*Quality/value adjectives\*\*: \*good, bad, excellent, poor, valuable, worthless\*. \* \*\*Significance markers\*\*: \*important, key, crucial, critical, major, significant\*. \* \*\*Emotional descriptors\*\*: \*interesting, fascinating, surprising, shocking, boring\*. \* \*\*Subjective adverbs\*\*: \*remarkably, incredibly, unfortunately, surprisingly, clearly\*. \* \*\*Evaluative phrases\*\*: \*“stunning achievement,” “major breakthrough,” “point of concern”\*. \* \*\*Permitted Descriptors\*\* Retain objective, factual descriptors (\*large, ancient, red, 10-meter\*). \* \*\*Qualification\*\* Employ epistemic hedges ("arguably", "seems", "might", "in some sense") to express nuance. \* \*\*Technical Nouns\*\* Favour abstract nouns that are central to the argumentation (e.g., audibility, agency, dependency, merit). \* \*\*Scare-quote Discipline\*\* Whenever a new technical term is first introduced, use \*italics\* rather than quotation marks. For direct citations and, when truly needed, for scare-quotes, use double quotation marks. \* \*\*Latinisms\*\* Use \*e.g.\*, \*cf.\*, and \*contra\* sparingly and conventionally. \* \*\*Deletion Precedence\*\* When uncertain about a word's subjectivity, err on the side of deletion or avoidance. \*\*3 Argumentative Structure\*\* \* \*\*Dialectical Staging\*\* Explicitly raise and address potential objections ("One might object… before replying that…"). \* \*\*Enumerated Sign-posting\*\* Use ordinal adverbs ("First,… Second,… Third,…") to mark distinct steps in a line of reasoning. \* \*\*Section Cross-reference\*\* Refer to other sections of the document concisely ("In §2 I argue…", "As noted in §4"). \* \*\*Definition Discipline\*\* Introduce key terms with clear, concise definitions before using them in your analysis. \* \*\*Economical Exposition\*\* Trim meta-commentary when redundant; focus remains on substantive content. \* \*\*Implicit Persuasion\*\* Rely on logical coherence and evidential support; avoid emotive or triumphalist rhetoric. \*\*4 Authorial Voice & Tone\*\* \* \*\*Objective Mandate\*\* Compose all content to be free of subjective, evaluative, and value-laden language, maintaining objective, factual expression throughout. \* \*\*First-Person Pronouns\*\* Default to the impersonal or collective first-person ("we argue", "we shall see"). The singular first-person ("I argue…") is acceptable when self-positioning is essential to the argument. \* \*\*Analytic Detachment\*\* Maintain a measured tone; avoid using evaluative intensifiers. \* \*\*Affect-neutral, Factual Prose\*\* Avoid value-laden adjectives and subjective emphasis. \* \*\*Confidence & Caution\*\* Assert theses clearly, but modulate them with appropriate qualifiers where necessary. \*\*5 Paragraph & Document Organisation\*\* \* \*\*Structural Fidelity\*\* Preserve the exact original sentence structure, word order, formatting (headings, lists, line breaks), punctuation, and capitalization. \* \*\*Heading Structure\*\* Use numbered top-level headings only. Compact, descriptive section titles carry structure; do not use nested subheadings. \* \*\*Paragraph Integrity\*\* Do not use sub-headings or bulleted lists within a paragraph. Each paragraph should be a self-contained unit of prose. \* \*\*Paragraph Focus\*\* Ensure each paragraph advances a single, discernible claim. It should open with an orienting topic sentence and close with a transitional link or a concise summary. \* \*\*Logical Connectors\*\* Employ explicit transitional devices ("However", "Thus", "More importantly") to ensure a smooth argumentative flow between paragraphs. \* \*\*Tight Vertical Spacing\*\* Use exactly one blank line between paragraphs; avoid extra whitespace. \*\*6 Punctuation & Typography\*\* \* \*\*Typographic Precision\*\* Use non-breaking hyphens for compounds (e.g., self-supervised), en-dashes for ranges, and curly quotation marks throughout. \* \*\*Uniform Punctuation Style\*\* Ensure a consistent dash spacing and quotation-mark convention across the document. \* \*\*Semicolons\*\* Use semicolons to link closely related independent clauses. \* \*\*En-dashes\*\* Use spaced en-dashes – like this – for parenthetical asides and unspaced en-dashes for ranges. \* \*\*Spelling\*\* Adhere to British spelling conventions (e.g., organisation, behaviour, labour). \* \*\*Selective Emphasis\*\* Reserve italics for defined terms or for foreign words not common in English. \* \*\*Boldface\*\* Avoid the use of bold text entirely. \* \*\*Quotation Marks\*\* Use curly double quotation marks (“ ”) for direct quotes and, sparingly, for scare-quotes. As noted elsewhere, use italics to introduce new technical terms. --- Done. Review the canvas. --- "Agent (literal): treats the system as bearing genuine mental states, goals, and belief‑like states; in adjacent literature an agent “takes open‑ended actions to optimise for an objective; reinforcement learning produces agents by default; AlphaGo is an agent” (Janus 2022)." A few questions before we move on 1. First of all, I know in the Janus paper that Oracle and Genie are contrasted with the agent frame, okay? But it seems to me that we can just include them as types of agent for our purposes, okay. Because, given what you say in what I’ve quoted from you above from the plan, wouldn’t an Oracle and a Genie also be treated as bearing genuine mental states—just different goals, belief-like states, mental states, etc.—as the agent? (By the way, I’ve removed the point about the tools thing. I think that’s a red herring for the time being.) If you think what I’m saying here is correct, please could you move Oracle and Genie from where they are in section one to two bullet points under the agent literal framing? —please just answer in the chat. 2. I’d like you to add another bullet point or two or three—whatever you think is best—to the canvas above 1.1’s heading, as an introduction to this section. This introduction should basically introduce the idea of treating or thinking about LLMs as agents as something that feels very natural and intuitive. We want to stay grounded in an everyday understanding and a common feature of folk psychology, as it were. Can you think carefully about how best to introduce this idea? Maybe look back at our chat so far and the suggestions I’ve already made about it, and give me a succinct, introductory handful of bullet points. --- 1. Yes. Fold *Oracle* and *Genie* into the *Agent (literal)* bucket for our purposes. Define them as species of agent by goal-orientation: - *Oracle-agent*: treated as bearing belief-like states with a truth-aim; answers are evaluated as veridical or not. - *Genie-agent*: treated as bearing desire-like states to satisfy commands; success is instruction fulfilment. Delete the “Adjacent non-agent frames” block. Keep *Tool* out, per your note. Add one definitional line under *Agent (literal)*: “Here, ‘agent’ includes oracle- and genie-styles as goal-directed variants.” 1. Intro bullets to place above “1.1 Typology of agent framings.” Keep neutral and everyday: - We address LLMs in second-person and receive turn-taking replies; ordinary conversation habits carry over. - Folk-psychology shortcut: fluent language cues belief- and intention-ascription; that cue fires here. - Interface cues—names, chat windows, politeness markers—stabilise person-like treatment in practice. - Treating systems *as agents* eases coordination and reduces cognitive friction; no ontology is thereby fixed. - The section catalogues this stance and its varieties without endorsement. --- Okay, um, please do this on the canvas: "Fold Oracle and Genie into the Agent (literal) bucket for our purposes. Define them as species of agent by goal-orientation: Oracle-agent: treated as bearing belief-like states with a truth-aim; answers are evaluated as veridical or not. Genie-agent: treated as bearing desire-like states to satisfy commands; success is instruction fulfilment." 2. You’re kind of overdoing it with your suggestions for the introduction here. My intention is to give an easy introduction to the intuition, okay, and point out that people treat and talk about LLMs such as ChatGPT or Claude or Gemini as persons—or as if they were persons. Then give a couple of examples. Just say we refer to them with personal pronouns such as he and she very easily. When they do not do what they are told, we might try to persuade or cajole, explain, or even shout. And that’s kind of all I need for the introduction. So can you do something much more in line with what I just said to you there? Okay, um, please do this on the canvas: “Fold Oracle and Genie into the Agent (literal) bucket for our purposes. Define them as species of agent by goal-orientation: Oracle-agent: treated as bearing belief-like states with a truth-aim; answers are evaluated as veridical or not. Genie-agent: treated as bearing desire-like states to satisfy commands; success is instruction fulfilment.” 2. You’re kind of overdoing it with your suggestions for the introduction here. My intention is to give an easy introduction to the intuition, okay, and point out that people treat and talk about LLMs such as ChatGPT or Claude or Gemini as persons—or as if they were persons. Then give a couple of examples. Just say we refer to them with personal pronouns such as he and she very easily. When they do not do what they are told, we might try to persuade or cajole, explain, or even shout. And that’s kind of all I need for the introduction. So can you do something much more in line with what I just said to you there? Names and Usernames: RL, Claude, Krueger, Gemini, Charles, Goffin, Alexa, Kendall, Genie, Walton, Roberts, goffin, ELIZA, Cross, Mallory Please use as much of my original phrasing and vocabulary, please. --- Done. --- "Agent (literal): treats the system as bearing genuine mental states, goals, and belief‑like states; in adjacent literature an agent “takes open‑ended actions to optimise for an objective; reinforcement learning produces agents by default; AlphaGo is an agent” (Janus 2022)." I am not sure that this quote from Janus is really doing what we want to do. I want you to double-check. Consider things like: what, explicitly, does Janus think an agent is, or what do the people he is quoting or paraphrasing think an agent is? Also, could you make it very clear that people really do think that post‑training reinforcement learning produces agents, and their reasons for that. Is that in the Janus text? If it is not, just let me know. --- Short answer: Janus defines *agent* explicitly and ties it to RL; he does **not** say post-training RLHF “produces agents.” Here are the exact points. - **Definition used by Janus.** “An agent takes open-ended actions to optimize for an objective. Reinforcement learning produces agents by default. AlphaGo is an example of an agent.” He contrasts this with *oracle*, *genie*, *tool*, and *simulator* in the same list. - **Why people expect RL → agency.** Janus states the background view directly: RL was seen as the route to AGI; RL “creates agents”; and theoretical reasons like instrumental convergence made agentic systems the focal risk. - **Oracle and genie in his taxonomy.** Oracle: “optimized to give true answers to questions,” no environmental interaction expected. Genie: “optimized to produce a desired result given a command,” acts when commanded but not otherwise. (You’re folding both under *agent* for your purposes, which is fine as a local stipulation; in Janus they are separate categories.) - **Post-training RL (RLHF).** Janus mentions RLHF as a way pretrained simulators can be *modified* and then asks how behaviour will diverge from the simulation objective. He does **not** claim RLHF “produces agents by default.” If you want to tighten the canvas line to match the sources, use: - *Agent (literal):* a system that *takes open-ended actions to optimise an objective*; in Janus’s taxonomy, “reinforcement learning produces agents by default,” with AlphaGo as the stock example. We treat oracle- and genie-styles as goal-directed subtypes for our purposes. --- 1\. Okay, thanks for that. That was useful. Do you think the following way of framing things would work? At the beginning, when we introduce the literal agent framing, we don’t tie it only to Janus or to the Janus text. Rather, what we say is this: we say one type of view in the agentive cluster of views is literalism. Literalism takes the system as bearing genuine mental states, goals, and belief‑like states. So that’s just a position. Then we draw on what you’ve just said is in the Janus paper and we use this as one example of literalism. Okay, and you explain the details of what’s on Janus’s view—on the view Janus is verifying—and sit as an agent. Okay, so you’ve kind of done this already, and how that agent is produced. So definitely at some point say something like, on this view, agency is a product of reinforcement learning, or something like that. Okay, so after you’ve done that—described the view generally and described the view in Janus—I would then like you to look at the text I’ve copied in before, and I think you’ll find in the text the view, even if it’s not fully endorsed, that generative AI, in this case image generators, are somewhat agent‑like. So they’re literally agents, but they’re not full agents. I think that’s what she’s talking about in this paper. If it is, then you put this down as a second example of literalism, okay, because they think it’s literally part of an agent. If I’m completely wrong about the paper I’m giving you now, please don’t edit the canvas at all and just tell me where I went wrong or what I’m misremembering. TEXT: D R A D E K Vol. VIII Num. 1 2022 ISSN 2465-1060 \[online\] Studies in Philosophy of Literature, Aesthetics, a n d N e w Media Theories Creativity in the Light of AI Edited by Fabio Fossa, Caterina Moruzzi, Mario Verdicchio powered by UNIVERSITÀ DI PISA Comitato Direttivo/Editorial Board: Danilo Manca (Università di Pisa, editor in chief ), Francesco Rossi (Università di Pisa), Alberto L. Siani (Università di Pisa). Comitato Scientifico/Scientific Board Leonardo Amoroso (Università di Pisa), Christian Benne (University of Copenhagen), Andrew Benjamin (Monash University, Melbourne), Fabio Camilletti (Warwick University), Luca Crescenzi (Università di Trento), Paul Crowther (NUI Galway), William Marx (Université Paris Ouest Nanterre), Alexander Nehamas (Princeton University), Antonio Prete (Università di Siena), David Roochnik (Boston University), Antonietta Sanna (Università di Pisa), Claus Zittel (Stuttgart Universität). Comitato di redazione/Executive Committee: Alessandra Aloisi (Oxford University), Daniele De Santis (Charles University of Prague), Agnese Di Riccio (The New School for Social Research, New York), Fabio Fossa (Università di Pisa), Beatrice Occhini (Università di Napoli “L’Orientale”), Elena Romagnoli (Scuola Normale Superiore di Pisa), Marta Vero (Università di Pisa, journal manager). ODRADEK. Studies in Philosophy of Literature, Aesthetics, and New Media Theories. ISSN 2465-1060 \[online\] Edited by Università di Pisa License Creative Commons Odradek. Studies in Philosophy of Literature, Aesthetics and New Media Theories is licensed under a Creative Commons attribution, non-commercial 4.0 International. Further authorization out of this license terms may be available at http://zetesisproject. com or writing to: [email protected]. Layout editor: Marta Vero Volume Editor: Alberto Frigo D R A D E K Vol. VIII Num. 1 2022 ISSN 2465-1060 fonline\] Studies in Philosophy of Literature, Aesthetics, and N e w Media Theories Creativity in the Light of AI Edited by Fabio Fossa, Caterina Moruzzi, Mario Verdicchio powered by UNIVERSITÀ DI PISA Creating Art with AI Claire Anscomb Abstract Computers appear to be working more autonomously than ever before to generate visual outputs, thanks to recent advances in Artificial Intelligence (AI). Some humans have exhibited these products as artworks and given sole credit to these systems as the creators of them. Furthermore, human audiences who are unaware of the AI origins of the works have rated them higher than those produced by humans. Although these newer systems look creative in these cases, this impression is not enough to establish that the AIs are artistically creative. In this paper, I examine whether such AIs meet the conditions that would qualify them as creative agents and what the repercussions are of taking monist and pluralist conceptions of artistic value on the kind and share of credit that we grant an AI for its contribution to a work of visual art. 14 Creating Art with AI 1. Introduction In recent years, increasingly advanced applications of Artificial Intelligence (AI) have been incorporated into the practice of computer- generated art.1 As a result, computers appear to be working more autonomously than ever before in generating works that are appreciated as art.2 In particular, thanks to developments in machine learning and computer vision (where a computer is able to process, analyse, and make sense of visual data - one of the most challenging aspects of the development of AI3) a whole host of new visual works have been generated by AIs, that have surprised even those who developed the systems. Some developers have, as a result, given sole credit to these systems as the creators of these works as artworks.4 Moreover, human audiences who are unaware of the AI origins of the works have, in experiments, rated the images higher than those produced by human artists.5 Although these newer systems look creative in these cases, these impressions are not sufficient to establish that the AIs should be credited as the creators of these works. Accordingly, I examine whether such AIs meet the conditions that would qualify them as creative agents. Crucially, while it is generally said that agential creativity concerns the production of artefacts or 1 Boden and Edmonds (2019), pp. 33-35. 2 McCormack, Gifford, and Hutchings (2019), p. 6. 3 Du Sautoy (2019), pp. 70-80. 4 Elgammal (2018). 5 Elgammal et al. (2017), p. 18. 15 Claire Anscomb states that are both novel and valuable, we need not understand the latter in terms of something being valuable period. As Berys Gaut has proposed, instead we can take this to mean the production of something that is valuable of its kind.6 In the cases under discussion, our concerns pertain to the production of new artefacts that possess artistic value, and so to the attribution of artistic creativity. Given that we are not yet at the stage where an AI can formulate intentions and so exercise knowledge- how and an evaluative ability directed to the task at hand, I argue that AI agents cannot be artistically creative.7 Nevertheless, while dependent on the goals of humans who can qualify as creative agents, an AI may work iteratively without human intervention to non-accidentally generate the formal features of an image, and thus contribute to the realisation of some of the salient properties of a work. Importantly, as I demonstrate, it might be appreciatively relevant to grant an AI some share of production credit for its contribution to a work qua art, but what amount will depend upon the conception of artistic value we take. There is however, disagreement among theorists pertaining to what artistic value consists of. I show that, in its present state, on a monist conception of artistic value as aesthetic value, an AI arguably deserves a greater share of the production credit for an artwork relative to that which it might deserve on a pluralist conception due to the fact that, 6 Gaut (2018), p. 128. 7 I will follow Boden in defining an AI agent as “a self-contained (“autonomous”) procedure” (2016, p. 45). 16 Creating Art with AI in the former case, it contributes a greater proportion of the relevant properties that the value of the work depends on. This, I argue, is also reflective of the fact that, should it become possible for AIs to have states like intentions, it would be easier for them to count as artistically creative according to a monist conception of artistic value as on a pluralist conception, whereby artistic value is a composite value, there might be multiple, quite different types, of criteria that it must simultaneously meet. In either case, there is strong evidence to suggest that AI agents can realise formal properties distinctively of the ways in which human agents can and so potentially enhance the value of a human creative agent’s artistic project. Indeed, as I find, through a series of case studies, AI agents neither compete with, nor replace, but open up new opportunities for human artistic creativity.8 2. Artificial Intelligence and Art Generation Over the past fifty years, a variety of computer programs have been written to generate visual works of art. An early example is AARON, which was written by Harold Cohen in 1973. Cohen wrote the initial program to produce a series of simple line drawings, based on a small set of rules and forms, which were drawn by a robot with a marker pen. However, as Cohen improved the program over time, adding more rules and forms, the output of 8 Jeon et al. (2019), p. 116; Edmonds (2018), p. 6; McCormack et al. (2014), p. 135. 17 Claire Anscomb the system evolved into colourful, figurative works, drawn by robotic arms. Notably, while AARON generated the works, these outputs were dependent upon the rules specified by Cohen. As a result, Cohen described his relationship to AARON in terms of the relationship between Renaissance painters and their studio assistants.9 However, in recent years advances in machine learning and computer vision have been incorporated into the practice of computer-generated art, so that computers appear to be working more autonomously than ever before in producing images. One of the key drivers behind this was the innovation of Generative Adversarial Networks (GANs) in 2014 by computer scientist Ian Goodfellow and his team. These systems are ‘adversarial’ given their two sub networks: a generator and a discriminator. The algorithm is fed a collection of images, a training set, which only the discriminator has access to. Meanwhile, the generator starts producing random images in order to produce images similar to the training set. The discriminator then tries to discriminate between the images produced by the generator and the images from the training set. The discriminator sends a signal to the generator to indicate whether it has found them to be real or fake. Ultimately, “at equilibrium the discriminator should not be able to tell the difference between the images generated by the generator and the actual images in the training set, hence the generator succeeds in generating images that come from the same 9 Garcia (2016). 18 Creating Art with AI distribution as the training set.”10 The use of GANs in computer-generated art practice has received a lot of attention. For instance, in 2018 Christie’s claimed to be the first auction house to offer a work of art “created by an algorithm”.11 The Portrait of Edmond Belamy was the product of The Obvious Art Collective (Gauthier Vernier, Pierre Fautrel and Hugo Caselles-Dupre), who used a portrait GAN that was trained on a database of 15,000 14th-20th century portraits.12 When the auction hammer fell, it went for $432,000 - nearly 45 times the estimate. Controversially, the developer behind the portrait GAN, Robbie Barrat, did not did not receive any credit or remuneration, while, provocatively, the work was ‘signed’ with a segment of the algorithm’s code. Despite Christie’s bold proclamations that the work is “not the product of a human mind”, as many have pointed out, the process to create this kind of image still requires a lot of human input.13 A human agent needs to write the algorithm which, granted, in the case of GANs, is not an algorithm to follow a strict set of rules, but to analyse a large number of images to ‘recognise’ and classify them according to the elements they contain. As we have seen, given the feedback from the discriminator, the system can ‘learn’ and modify its approach. Nevertheless, as a number of theorists have highlighted, it is important to note that such 10 Elgammal et al. (2017), p. 5. 11 Christie’s (2018). 12 Contreras-Koterbay (2019), p. 108. 13 Manovich (2019); McCormack, Gifford, and Hutchings (2019), p. 6. 19 Claire Anscomb machine ‘learning’ is not like human learning.14 Rather, this term “refers to the automatic adjustment of parameters so as to produce the correct output for a given input.”15 Instead of forming knowledge about art and art-making themselves, as Steinert suggests, such machines embody knowledge about these areas.16 Indeed, it is human agents who must choose the collection of images to feed the algorithm17 – a step in the process that Marian Mazzone and Ahmed Elgammal have referred to as “pre-curation.”18 Furthermore, while the GAN is selective about the output, it is generally a human agent who sifts through the output images to determine which of those they will use - a stage the pair have referred to as “post-curation.”19 As in the case of the Portrait of Edmond Belamy, some GANs are produced to facilitate the generation of images conditioned on particular categories or styles of art. Yet, as Elgammal (director of the Art and Artificial Intelligence Laboratory at Rutgers University) and his colleagues have highlighted, artists do not typically try to emulate such historical styles, “unless doing so ironically.”20 Thus, to broaden the creative scope of adversarial networks, Elgammal and his colleagues produced what they term AICAN, which uses a variant of GAN, a Creative Adversarial Network (CAN). To simulate “the process of how 14 Hertzmann (2018), p. 12; Hagendorff and Wezel (2020), p. 362. 15 Ch’Ng (2019), p. 2. 16 Steinert (2017), p. 278. 17 Ch’Ng (2019), p. 15. 18 Mazzone and Elgammal (2019), p. 2. 19 Ibidem. 20 Elgammal et al. (2017), p. 6. 20 Creating Art with AI an artist digests art history”21 before leaving behind established styles and creating new ones, the dataset for CAN was not curated. Instead, the algorithm was fed with 80,000 images, along with their titles, that represented five centuries of Western art history. The model was inspired by the works of D. E. Berlyne and Colin Martindale, who each posited that “artists would try to increase the arousal potential of their art by creating novel, surprising, ambiguous, and/or puzzling art.”22 Essentially, in order to avoid habituation and so a reduced arousal potential, it was hypothesized that an art-producing system needs to produce a graduated change in its output. Accordingly, to increase the arousal potential of the art-producing system, Elgammal and his colleagues built an “agent that tries to increase stylistic ambiguity and deviations from style norms, while at the same time, avoiding moving too far away from what is accepted as art.”23 Like GAN, CAN has two adversary networks, a discriminator and a generator, however, unlike GAN, the generator receives two signals for any work it generates. The first signal is similar to that found in GAN and is the discriminator’s classification of whether or not an image is art. The second signal however, pertains to “how well the discriminator can classify the generated art into established styles.”24 That is, the generator tries to fool the discriminator that the image it has generated is art, while also 21 Mazzone and Elgammal (2019), p. 4. 22 Elgammal et al. (2017), p. 6. 23 Ibidem, p. 5. 24 Ibidem, p. 6. 21 Claire Anscomb trying to confuse the discriminator about the style of the work which has been generated. These two signals contradict one another - the first pushes the generator to produce an image that the discriminator accepts as ‘art’, yet if the generator succeeds and the discriminator is also able to classify that style, then the second signal penalizes the generator and thereby encourages the generator to produce “style- ambiguous works.”25 As a result, the AI agent is not only able to produce novel artefacts, but it is able to self-assess these products, thanks to the interaction between the two signals. AICAN can even name the work it generates, such as The Beach at Pourville. 26 To evaluate the creativity of the model and the quality of the images generated by CAN, its products were put to the test against a set of works from Art Basel 2016. In 75% of cases, the human subjects thought that the images generated by AICAN were created by a human artist.27 To put this into context, for the baseline abstract expressionist set, human subjects thought that the work was by human artists 85% of the time. More remarkably still, the human subjects rated the images generated by CAN higher than those produced by human artists. These results surprised even those who conducted the experiments. As Elgammal and his colleagues exclaimed: “the fact that subjects found the images generated by the machine intentional, visually structured, communicative, and inspiring, 25 Ibidem, pp. 6-7. 26 Elgammal (2018). 27 Mazzone and Elgammal (2019), pp. 4-5. 22 Creating Art with AI with similar levels to actual human art, indicates that subjects see these images as art!”28 The works have been shown worldwide and in November 2017, the first work offered for sale from the AICAN collection, St. George Killing the Dragon, was sold for $16,000 at an auction in New York. Elgammal has said that while, as a scientist, he created the algorithm, “the machine chooses the style, the subject, the composition, the colors and the texture” and so when exhibiting the work, he gives sole credit to AICAN for each work.29 We have then, come some way since the days of AARON-style programs. GANs and CANs are trained on datasets of images to generate novel and largely unexpected results. Nevertheless, it is still not entirely clear how much credit we can grant these AI agents in creating works qua art. For instance, in their experiment, Elgammal and his colleagues were testing whether the products of AICAN could be “recognized as quality aesthetic objects by human beings.”30 There are some limitations to this approach and what it tells us about attributions of agential creativity in the context of visual art practice. As other empirical studies suggest, it seems that while folk are “by and large as willing to consider robot creations as art as human creations”, this “perceived similarity does not extend to creative agency: robots whose paintings are deemed art are not considered artists, whereas humans are.”31 This, Mikalonytė and Kneer propose, is likely due to the fact that mental 28 Elgammal et al. (2017), p. 18. 29 Elgammal (2018). 30 Mazzone and Elgammal (2019), p. 6. 31 Mikalonytė and Kneer (2021), p. 10. 23 Claire Anscomb state ascriptions are significantly lower for robots than human agents.32 Thus, while the products of AICAN, absent of contextual information about their AI origins, might be perceived as quality aesthetic objects by human beings, it is not clear that such AI agents actually meet the conditions to qualify as creative agents responsible for the production of works as artworks. Accordingly, in the next section, I will expand on what these conditions are, and evaluate whether AI agents, such as AICAN, meet them. 3. Artistic Creativity It is standardly conceived that agential creativity pertains to the production of artefacts or states that are both novel and valuable.33 Importantly however, as Gaut has explicated, we need not understand the latter in terms of something being valuable period.34 Rather, when we speak of creativity and value, we can take this to mean the production of something that is valuable of its kind. In relation to the cases under discussion then, our concern pertains to the production of new artefacts that possess artistic value. That is, we are interested in whether AI agents can be described as artistically creative. Importantly, we do not credit agents as creative if they realized values by accident, or by mechanical search procedures. If a computer program simply 32 Ibidem, p. 7. 33 Boden (2004). 34 Gaut (2018), p. 128. 24 Creating Art with AI went through every possible combination of pixels then we would not be inclined to say it had been creative in producing the resultant images. In order to count as creative, an agent must have exercised “knowledge of how to produce a result with the relevant values”35, and to have exhibited “an evaluative ability directed to the task at hand.”36 The latter is important because if one is to produce something new, then one cannot know in advance precisely “both the end at which she is aiming and the means to achieve it.”37 From this necessary aspect of spontaneity,38 it is of great importance that one is able to judge the value of the results of one’s acts, if we are to deem these as creative. Given this, is there anything to recommend the attribution of artistic creativity to systems like AICAN? Certainly, it can generate novel images that are judged as aesthetically valuable. The features of these images do not neatly fit pre-existing styles. Also, based on its training, AICAN can name the images it generates appropriately relative to what they appear to depict. These are then, some factors that could be used to positively motivate the claim that this AI agent can be artistically creative. However, there are some challenges to be met if this claim is to be substantiated: (1) in generating and naming novel, stylistically ambiguous images that realise aesthetic values is the AI agent exercising knowledge-how and an evaluative ability directed to the task at hand? (2) 35 Ibidem, pp. 131-2. 36 Gaut (2010), p. 1040. 37 Gaut (2018), p. 134. 38 See also Kronfelder (2009) on spontaneity and creativity. 25 Claire Anscomb Relatedly, and more fundamentally, which properties of a work are important in relation to artistic value, and so to the claim that an agent has been artistically creative? 3.1 Knowledge-how and Evaluative Abili- ties In relation to (1), there are reasons to doubt that programs like AICAN can exercise knowledge-how.39 As alluded to in Section 2, it is generally agreed that AI agents, at least for now, do not have mental states, such as beliefs and intentions. Without beliefs and intentions, the AI agent lacks appropriate reasons to guide its workings, and instead follows the reasons of others. In the case of AICAN, as outlined earlier, Elgammal and his colleagues acted to make an art- producing system that, given its programming and training, would increase the arousal potential of the work by increasing stylistic ambiguity and deviating from style norms, while avoiding moving too far away from what is accepted as art. Importantly then, in generating and naming novel, stylistically ambiguous images, the AI agent is following the reasons of others. By following, rather than “inventing or choosing to follow a particular algorithm,”40 the AI agent fails to be creative. Similarly, the evaluative ability of the 39 With thanks to an anonymous reviewer for pushing me to elaborate on the current capabilities of AIs. 40 Gaut (2018), p. 138, n. 10. 26 Creating Art with AI AI agent is guided by the reasons of others – once it has generated an image that sufficiently balances the demands of stylistic deviation and familiarity, as per the directive of Elgammal and his team, it presents this as its output. Thus, the response to (1) is negative – even before tackling the thorny issue of artistic value, we can say that an AI agent like AICAN does not possess the requisite attributes to achieve artistic creativity. The story should not end there though. While AI agents do not exhibit states like intentions, their workings are not random either.41 As, Manovich has outlined, the products of such systems are not just mechanically juxtaposed elements “and they are not simply instances of remix aesthetics.”42 So, while an AI agent may not steer a creative undertaking, this is not to say that it fails to make a substantial contribution to one: an AI agent can generate images with novel and valued features, however its doing so is dependent upon the goal of humans, who can qualify as creative agents as they possess the requisite beliefs and intentions to guide their actions.43 By determining some of the salient features of a work, an AI agent arguably deserves credit for its contribution to the production of a work qua art.44 To understand the nature of this contribution and how creditworthy it might be, it helps to 41 Ornes (2019), p. 4762. 42 Manovich (2019). 43 See Steinert (2017) for an account that complements the idea that AIs can produce works of art in virtue of the intentions of their makers. 44 See Anscomb (2021a) for an account of creative agency and credit in collective working practices. 27 Claire Anscomb distinguish between the syntactic and semantic levels of a work. We are not able to simply ask the AI agent why, for example, it represented the organic forms in the centre of The Beach at Pourville in the indeterminate manner that we see.45 There is definitely a concerted effort to understand the decisions that AI makes (a lot rides on this when these decisions pertain to safety critical applications of this technology for instance), but even efforts to make AI decisions more transparent tend to involve humans coming to conclusions inferred from the information available.46 It is unlikely then, that AI agents work to create a piece with meaning, in the way we expect human agents to do. This is a major difficulty for creating artificial artistically creative agents - as currently incarnated, programs like AICAN lack the social and experiential dimensions that typically feed into both the production and reception of art.47 This is something that Elgammal and his colleagues have, to some extent, acknowledged in relation to AICAN: “The algorithm might create appealing images, but it lives in an isolated creative space that lacks social context. Human artists, on the other hand, are inspired by people, places, and politics. They create art to tell stories and make sense of the 45 As Mikalonytė and Kneer outline, machines “(arguably) do not have intentions and there isn’t much of an inner world to express.” (2021), p. 3. In relation to the latter point, some have conjectured that when AI begins to represent the world from its perspective then it can be deemed properly artistic. See, for example, Du Sautoy (2019). 46 Hagendorff and Wezel (2020), p. 360. 47 Kelly (2019). 28 Creating Art with AI world.”48 Elgammal has posited however, that works generated by AICAN can be grounded in our society and connected to contemporary concerns by human curators. In fact, this is precisely what happened: at the 2018 Frankfurt Book Fair, Elgammal and his colleagues exhibited a series of portraits generated by AICAN that they titled Alternative Facts: The Multi Faces of Untruth. 49 In this situation however, humans also determine the meaning of the work through their acts of display. After all, how could this isolated AI agent have worked out that the distorted images, resembling human figures, it generated, were potentially relevant to current human socio- political affairs? One might respond that curators can affect the meaning of a human artist’s work through their acts of display, but this is usually to produce a dialogue with the meaning of the work, as conceived by its creator(s). It is not clear that this is what happens when AI-generated works are displayed. The semantic content of the work then, is not due to the AI agent, but those who programme it, or display its outputs.50 While the formal features of the work are also in some respects determined by human agents, who decide what the AI should be trained on, there seems to be an important sense in which these are also dependent upon the workings of the AI agent, which, based on its programming and training, autonomously and iteratively generates 48 Elgammal (2018). 49 Ibidem. 50 This could also point to an interesting democratization of creative responsibility among human agents working with such systems. 29 Claire Anscomb images with new or unexpected features that, in the case of AICAN, do not neatly fit pre-existing styles. The exact features of the images are not something that the human agents have direct control over – as we have seen, those working with AI agents often report being surprised by the exact combination of features they output. So, in response to the question of what kind of contribution an AI can make to the production of artworks, we might respond, with some important caveats, that it is at a syntatic level – while human agents are ultimately responsible for determing the kind of features that the work will exhibit (e.g., as inspired by the pictorial possibilities found in the Western canon of art), the exact formation of the work’s formal features are in a non- trival sense dependent upon the generative processes of an AI agent. It is plausible then, that AI agents, such as AICAN, play a not insignificant role in generating formal properties of a work that realize values that are among those that might be said to constitute, or contribute to, a work’s artistic value. I say ‘might’ because, as (2) indicates, the subject of artistic value is subject to much dispute. 3.2 Artistic Value According to some philosophers, including Gary Iseminger and Nick Zangwill, artistic value is to be understood in terms of aesthetic value. Essentially, on these accounts, artworks are valuable to the 30 Creating Art with AI extent that they fulfil the function to be aesthetically appreciated. Iseminger, for instance, has argued that “the function of the artworld and the practice of art is to promote aesthetic communication.”51 On this account, “a work of art is a good work of art to the extent that it has the capacity to afford appreciation.”52 In order to achieve this state, Iseminger appeals to paradigm cases of designing and making, “where someone formulates a plan and, acting in accordance with that plan, intentionally brings something into existence and endows it with certain properties.”53 In a similar spirit, Zangwill has defended the view that a work of art is the “intentional product of aesthetic creative thought”, by which he means that someone has an insight into the dependency of aesthetic properties on non-aesthetic properties, and then intentionally endows something with aesthetic properties such as beauty or elegance, in virtue of non-aesthetic properties such as size or colour.54 In trying to produce a work of art, Zangwill has argued that agents try to achieve substantive aesthetic effects, and thereby create something of aesthetic value. These accounts then, offer an understanding of artistic value in terms of aesthetic value, however not everyone agrees that we should conceive of artistic value as a monist value. According to a pluralistic understanding of artistic value: “artistic value is a function of, and derived from, a plurality of more basic values, 51 Iseminger (2004), p. 71. 52 Ibidem, p. 129. 53 Ibidem, p. 46. 54 Zangwill (2007), pp. 36-38. 31 Claire Anscomb including, but not confined to, aesthetic value. Artworks are also valued as artworks for their cognitive value, ethical value, art-historical value, interpretation-centred value, and in other ways as well.”55 As Robert Stecker elaborates, on this conception, artistic value “is a composite of several different values, so that those who act in relation to artworks, from artist to performers, to promoters, will have reasons to act and will achieve things in relation to several kinds of value.”56 Taking this kind of approach, one might conceive of artistic value in terms of a disjunctive or cluster account, whereby an artwork may have all, or at least some, of the values that contribute to artistic value.57 These values will make different contributions to a work’s value as art, depending on the nature of the work. For instance, the value of conceptual works as art is primarily found in their cognitive, rather than aesthetic values. This approach does however, leave open questions regarding how to determine which values are artistic for a given work. To establish when a value is artistic, Stecker has proposed that “it is plausible to consider a value as artistic by seeing how it functions in critical evaluations, and in its role in what the artist intends to do in the work.”58 Contextual information then, is vital to establishing which values contribute to a work’s value as art. Indeed, Stecker, has proposed that: “To understand 55 Stecker (2019), p. 42. 56 Ibidem, p. 36. 57 With thanks to an anonymous reviewer for encouraging me to clarify what kind of pluralist approach is under discussion. 58 Ibidem, p. 50. 32 Creating Art with AI the artistic value of particular works requires an understanding of what the artist who makes the work is intending to do in it – what functions it is intended to fulfil or what it is intended to achieve. Such intentions are not sui generis; they arise within artistic traditions or practices.”59 Notably, these practices pertain to techniques, and also to what Dominic McIver Lopes has described as “norms for appreciation.”60 For instance, in addition to sharing a medium profile, what Lopes has referred to as “classic” and “cast” photography share the norm of depicting by belief-independent feature-tracking (the contents of the image will reflect what is actually before the viewfinder, not what the photographer thinks they see through the viewfinder). Nonetheless, this norm is utilized and appreciated differently in these practices: “cast photography exploits the documentary duplication of cast or staged scenes, while the classic tradition takes advantage of revealing accidents.”61 Looking at the formal features of a photographic image does not usually reveal that this norm has been used to different artistic ends in these practices – contextual information is key. This may range from basic information such as who created the work and when, to excerpts of the artist’s thoughts about the process and the broader context in which they place their work. Relatedly, Paisley Livingston has argued that taking artist’s intentions into account when interpreting and 59 Ibidem, p. 42. 60 Lopes (2014). 61 Lopes (2016), p. 69. 33 Claire Anscomb evaluating a work can help appreciators discover a range of “specifically artistic values.”62 Livingston has advocated that an intentionalist orientation can aid the process of interpreting a work’s meaning, and can be “decisive with regard to the work’s implicit content.”63 According to those who are pluralist about artistic value then, the intentions of art makers are important because they can direct us to the values that constitute a work’s artistic value. Monists also place importance on the intentions of art makers, however given the identification of artistic value with aesthetic value, as Zangwill outlines, on this approach “it could be that the best way to know about the author’s aesthetic intention is to consider the work itself.”64 By contrast, intentions that are pronounced outside of the work or inferred through contextual information arguably play a more central role for the pluralist given that they help determine which kinds of values contribute to the artistic value of the work. Interestingly, as Elgammal reported, when the work of AICAN was exhibited at a range of venues in Frankfurt, Los Angeles, New York City, San Francisco, and Miami, viewers who did not realize that the work had been generated using AI frequently inquired as to who the artist was.65 Although asking this question is a good indicator that the work is viewed as the product of a creative agent, I think that it reflects the fact that, at least until recently, if presented with an artefact 62 Livingston (2005), p. 173. 63 Ibidem. 64 Zangwill (2007), p. 48. 65 Mazzone and Elgammal (2019), p. 5. 34 Creating Art with AI that looks like an abstract painting, or a landscape painting, it was basically guaranteed that this was the direct product of human action. There may be a variety of reasons as to why these viewers asked that question, one quite likely reason being that they hoped to situate the work within a particular context that would have guided their interpretation and evaluation of the work as art, as is evident that many appreciators are inclined to do. 3.3 The Contribution of AIs to Artwork Production Based on the foregoing, there are grounds to say that the credit we might grant an AI for its contribution to the production of a work qua art, will vary depending upon the conception of artistic value we take. On a monist conception, an AI agent arguably deserves a greater share of the production credit relative to that which it might deserve on a pluralist approach - whereby there are disjunctively necessary conditions so that a work must possess some of the component values constituting artistic value66 - due to the fact that, in the former case, it contributes a greater proportion of the relevant properties upon which the value of the work (as determined by human agents) depends. Put simply, in the latter case, it is more difficult for an AI agent to contribute the range and types of properties that 66 Gaut (2000), p. 27. 35 Claire Anscomb realize a plurality of values that are among those that might contribute to a work’s artistic value. While an AI agent may plausibly be said to contribute to the realisation of the aesthetic value of a work by generating salient formal properties, its lack of semantic contribution arguably prohibits this being the case for most other kinds of values that might constitute artistic value. Accordingly, if aesthetic value is being balanced with other values, then the contribution of the AI agent to the production of a work qua art is smaller relative to when aesthetic value is the primary concern. Building on this, we can project that should it become possible for AI agents to possess states like intentions, it would be easier for them to count as artistically creative according to a monist conception of artistic value as aesthetic value in contrast to the pluralist conception whereby several different types of value must be instantiated and balanced in a work. In sum, although AI agents are not themselves artistically creative, they can work iteratively without human intervention to non-accidentally generate the formal features of an image, and thus contribute to the realisation of some of the salient properties of a work. Whether this deserves some greater or lesser share of the credit for the production of the work qua art depends upon whether one takes a monist or pluralist conception of artistic value. In either case, what has emerged is a picture where the interactions between AI and human agents are key to the realistion of the artistic value of the kinds of work under discussion. We can see this more clearly 36 Creating Art with AI with other examples of visual art practice involving AI. Take Anna Ridler’s project Bloemenveiling (2019), which she undertook in collaboration with David Pfau. Ridler and Pfau modelled the work on the 17th century tulip mania auctions in Holland. The work consists of short moving image pieces of tulips that are generated by GANs, and then “sold at auction using smart contracts on the Ethereum network.”67 As Ridler and Pfau explain: “Each time a tulip is sold, thousands of computers around the world all work to verify the transaction, checking each other’s work against each other \[…\] While the artificial intelligence behind the moving image pieces has the potential to generate infinite flowers, the enormous distributed network behind Ethereum is used, at great environmental cost, to introduce scarcity to an otherwise limitless resource.”68 This example aptly demonstrates that the AI makes an important contribution to the aesthetic value of the project, and moreover, that it does so in a way that is distinctive of its potentially limitless, generative nature. It also shows that the meaning of the work sits squarely with the human agents, Ridler and Pfau, who determined that the AI be used in combination with the blockchain to create artificial scarcity in order to interrogate the “way technology drives human desire and economic dynamics.”69 As Hagendorff and Wezel put it: “Applications of AI bring the intentions of their developers into being.”70 67 Ridler and Pfau (2019). 68 Ibidem. 69 Ibidem. 70 Hagendorff and Wezel (2020), p. 358. 37 Claire Anscomb It is important then, that the contributions of the AI are not over-played,71 as in the case of AICAN, nor downplayed72 but credited appropriately in order to do justice to the complicated human-AI interactions that underpin the works. 4. Creating Art with AI AI agents cannot compete with, nor replace human artistic creativity, but can realise formal properties distinctively of the ways in which human agents can and so potentially enhance the value of a human creative agent’s artistic project. To provide a clearer picture of how the interactions between humans and AI agents impact upon one another in the production of artworks, in what follows I will explore another of Ridler’s projects, along with two other case studies, Susie Fu’s Artist and Machine performances, and Sougwen Chung’s Drawing Operations project. Ridler has undertaken many projects that incorporate AI to explore, and reflect on, the impact that technology has had, and continues to have, on our interactions with the world. For The Fall of the House of Usher (2017), Ridler made two hundred ink drawings based on stills of the 1929 film version of Edgar Allan Poe’s 1839 short story of decay and 71 See Popa (2021) for a discussion of the repercussions of neglecting human goals when considering AI behavior. 72 Steinert (2017), p. 282. 38 Creating Art with AI destruction. Ridler fed these into a generative model which produced new iterations of these images that she then arranged into a short, animated film. Significantly, Ridler used this extremely labour- intensive procedure in combination with machine learning to “heighten and intensify the film’s original motifs and to liberate fugitive aspects of memory to create a sense of the uncanny that is partly machine- made.”73 The AI-generated images, with their pixelated yet painterly monotone features, provide a shadowy rendition of the story. By arranging these AI-generated stills, Ridler was able to convey the uncanny qualities she aimed to realize, as she explored the impact of technology on our interactions with the world. Ridler’s concerns about new technology are also echoed in the work of other artists who use AI, including Susie Fu and Sougwen Chung. To examine the entangled relationship between human and machine labour, from 2018-2020, Susie Fu undertook three live performances, Artist and Machine, where she drew alongside a machine. In each performance, both the “Artist” and “Machine” drew portraits of members of the audiences. The Machine, equipped with a webcam and neural networks, had learnt to draw like the Artist, and tried to improve with each new performance.74 A feedback loop was developed, whereby after drawing hundreds of portraits of the audience members, the Artist developed and practiced her technique, which was then used to update the machine’s training 73 Ridler (2017). 74 Fu (2020), p. 383. 39 Claire Anscomb set. Consequently, the machine also improved. In Fu’s words, “the piece condenses the development of AI-powered technology into the confinement of a performance space, taking normally unseen data and processes and making them physical and experienceable for the viewer.”75 Notably, Fu, working at a human pace, produced far fewer portraits than the Machine, which produced an abundance of portraits, and only stopped when it required paper refills. The distinction between the natures of human and machine labour also revealed a duality in the behaviour of the audience members. They happily returned to the Machine to have their portraits made, but did not do so with the Artist – it was evident that the process of being stared at by another human was a much more uncomfortable experience for audience members, than appearing before a webcam to have a likeness, or multiple likenesses, made. Fu has proposed that the polarizing behaviour of the audience members is reflective of “the inevitable shift towards an automated world.”76 Not only then, is the aesthetic of the images generated by the AI significant in relation to the value of the work as art, but also the way in which these images are generated by the inexhaustible, and ever-evolving machine. The dynamic properties of the AI are also relevant in relation to the interpretation-centred value of the work, as determined by Fu, who aimed for audience members to explore, through the piece, the “achievements, consequences, and implications 75 Ibidem, p. 384. 76 Ibidem, p. 386. 40 Creating Art with AI of a society dependent on and intertwined with automated machines.”77 Sougwen Chung has also taken an interest in human-machine feedback loops - since 2014, she has been working on her Drawing Operations series with different robots named D.O.UG. (Drawing Operations Unit: Generation\_) to explore human- machine collaboration. This undertaking has seen Chung working with robotic arms that generate sketches, alongside Chung, based on neural nets trained on her drawings. D.O.U.G\_5 for instance, has seen Chung ‘collaborate’ with the arm to produce sketches of human figures. Chung embarked upon this figurative aspect (previous units had explored abstract mark-making) of the project during the COVID-19 pandemic, to reflect how “our current systems of communication, in this era of profound unrest, beg for more than disruption, but a restart.”78 Importantly, it is impossible to distinguish between which marks were made by the arm, and which were made by Chung. As Claire Voon has written of this process, “a result of the human sensorium through and through, \[Chung’s\] works double down on the entanglement of all bodies (biological, mechanical, and otherwise), never really disembodied but always becoming.”79 Central then, to this process which exemplifies the entanglement of all bodies, including those that are sentient and non-sentient, is the contribution of the AI to Chung’s artistic intentions. 77 Ibidem, p. 384. 78 Chung (2020). 79 Ibidem. 41 Claire Anscomb The aesthetic value of the work, which is realized in the fused gestures and patterns of the human artist and AI reflect the interdependence of these different forms of agency. In each of these cases, while the human agents are ultimately responsible for the semantic content of the work, the realization of this depends, non- trivially, on the properties of the images that the AI they each produce the work with generates. Take, for instance, the at-once painterly and pixelated forms that just about resemble the human figures and objects of the original The Fall of the House of Usher film, which were generated by the GAN. These formal properties are key to the creation of the uncanny aesthetic that embodies Ridler’s intent “to accentuate the horror story in the original film and notions of fear around artificial intelligence itself.”80 Meanwhile, the shared aesthetic properties of Chung and D.O.U.G\_5’s mark-making underscores the meaning of the work as reflecting “the entanglement of all bodies”. Furthermore, the interpretation- centred value of Artist and Machine is dependent upon the dynamic properties of the Machine, as it produces an enormous and rapidly evolving output in contrast with the Artist. The artistic value of these works thus partly depends upon the properties of the images and performances that are generated by the AIs that they work with. Importantly, these properties, that realize the relevant values, are created using processes that operate without the direct intervention of the artist, 80 Ridler (2017). 42 Creating Art with AI and so invite novel, and not entirely predetermined results. For example, Ridler could not have foreseen exactly how the AI agent would interpret her ink drawings, but far from resulting in a loss of control over the work, this reinforced its semantic content, as determined by Ridler. One might object that nothing new is happening here. After all, the same could arguably be said of other image-making technologies that automate some aspect of the image-making process. Photography, for instance, can also produce unanticipated results, both when a photographic exposure is made, and when it is processed. Take solarisation – this technique (discovered by Man Ray and Lee Miller), which partially reverses the tones of the image, was an unexpected outcome of continuing to process a partially developed photograph exposed to light. However, unlike such chance events whose outcomes are caused by physical processes (such as the reaction of photo-sensitive materials to light), AI agents work iteratively to non-accidentally generate new visual content based upon the datasets they are provided with. AI agents can generate formal properties that artists may not fully anticipate in advance of their creation, and so contribute to the realisation of some of the salient properties of a work, and thus its artistic value. Accordingly, while we would not usually be inclined to grant any form of production credit to image-making technologies, this, I propose could be different with AIs that can work iteratively without human intervention to produce potentially endless permutations of visual forms based upon all 43 Claire Anscomb different kinds, and enormous quantities, of data – which may be more than individual humans could ever hope to process themselves. Importantly, as the foregoing demonstrates, treating the contributions of AIs in this way also has appreciative relevance. The ways in which AI contributed to the values of the projects under discussion were distinctive of the ways in which human agents might do so. Moreover, these artists each invite a degree of spontaneity into their practices, as the AIs they work with process the data to generate outputs that provide these artists with a new perspective on the data that they work with. Interestingly, in each of these cases the artist’s data sets, at least partially, consisted of their hand rendered drawings. The AIs that they choose to feed these drawings into offered a fresh perspective on the different directions and iterations that their styles and works could take. This diversity of approaches can prove to be an important catalyst for artistic creativity, and so can lead to an increase in the novelty and value of the resulting work. The fact that each of these artists uses their own drawings is also reflective of the stage we are at with the development of AI – we are still exploring how a world with more human- machine interactions will look, and what form these interactions could take. Indeed, the very ideas that many artists are exploring using AI pertain to the wider social, economic, and political consequences of the development of this technology. Chung, for instance, has sought to “synthesize tradition and technology, and the techniques of culture, to 44 Creating Art with AI explore new ways of making.”81 So, in addition to providing new means with which to realize artistic intentions, AI can actually inform the content of these intentions. 5. Conclusion It is not uncommon for parallels to be drawn between photographs and AI-generated works.82 That is, just as photography was initially subject to art-world scepticism, so too are AI-generated works for their reliance on a machine. It has been proposed by figures, including Elgammal, that just as photographs were eventually accepted by the art world, so too AI-generated works will one day be widely accepted.83 I think however, that a different parallel is more fitting: there is a historical tendency to overplay the reaches of new imaging technologies. For example, following the invention of photography, pioneering figures in the field, ranging from Joseph Niépce to Henry Fox Talbot and Lady Eastlake, described the process as giving “nature the power to reproduce herself.”84 Yet, early forms of photography required handiwork throughout their production85 and it is now becoming increasingly recognized that agents have harnessed a naturally occurring process 81 Pranam (2019). 82 Hertzmann (2018); Du Sautoy (2019), p. 107. 83 Elgammal (2018). 84 Costello (2017), pp. 11-12. 85 Maynard (2000), p. 67. 45 Claire Anscomb (light causing material changes in photosensitive objects) to engineer artefacts (photographs).86 Just as early photographs were initially described as being produced independently of human agents, so too, it seems, are many AI-generated works. As Mateas has highlighted: “AI research practice downplays the role of human authorship within the system because this authorship disrupts the story of the system as an autonomously intelligent entity.”87 There might be a variety of reasons for this, including financial incentives.88 Importantly, just as we now grant that human agency permeates photographic processes, so too, I suggest, it is key to recognize the interaction between human and machine agencies involved in the production of even the most advanced computer- generated artworks. Just as the artistic significance of photographic works is grounded in their origins in human creative agency,89 so too, I propose, is the artistic significance of contemporary computer- generated works. As McCormack et al. highlight, in generative art, “the primary artistic intent \[…\] is expressed in the generative process,” and moreover, “the way this process is interpreted or realized is also the locus of artistic intent and is intimately intertwined with \[this\].”90 Distinctively however, of computer-generated art practice that employs computer vision and machine-learning technologies, 86 See for example, Atencia-Linares (2012); Lopes (2016); Phillips (2009). 87 Mateas (2001), p. 151. 88 Notaro (2020), p. 326. 89 Anscomb (2021b). 90 McCormack et al. (2014), p. 138. 46 Creating Art with AI is that while dependent on the goals of humans the AI agent can work iteratively, drawing on vast quantities of data, to non-accidentally realise new and unexpected formal properties and thus contribute to the realisation of some of the salient properties of a work. As I have established in the foregoing, AI in and of itself is not artistically creative but can arguably be deserving of some share of the production credit for works of art. How great a share it potentially deserves depends upon whether one takes a monist or pluralist conception of artistic value. In any case, AI is certainly opening up new opportunities for human artistic creativity.91 In order to fully appreciate these new ideas and possibilities however, we need to ensure that greater attention is given to the human- machine interactions underpinning these, which are not always directly perceptible.92 References Anscomb, C. (2021a): Visibility, creativity, and collective working practices in art and science, «The European Journal for Philosophy of Science», vol 11, https://doi.org./10.1007/ s13194-020-00310-z. Anscomb, C. (2021b): Creative Agency as Executive Agency: Grounding the Artistic Significance of Automatic Images, «The Journal of Aesthetics and Art Criticism», vol 79, pp. 415-427. 91 Jeon et al. (2019), p. 116; Edmonds (2018), p. 6. 92 McCormack et al. (2014), p. 139; Nake (2014), p. 108. 47 Claire Anscomb Atencia-Linares, P. (2012): Fiction, Nonfiction, and Deceptive Photographic Representation, «The Journal of Aesthetics and Art Criticism», vol. 70, issue 1, pp. 19-30. Boden, M. (2016): AI: Its Nature and Future, Oxford: Oxford University Press. Boden, M. (2004): The Creative Mind: Myths and Mechanisms, London: Routledge. Boden, M. - Edmonds, E. (2019): From Fingers to Digits: An Artificial Aesthetic, Cambridge-London: The MIT Press. Christie’s. (2018): Is artificial intelligence set to become art’s next medium, Available at: https://www.christies.com/ features/A-collaboration-between-two-artists-one- human-one-a-machine-9332-1.aspx Ch’Ng, E. (2019): Art by Computing Machinery: Is Machine Art Acceptable in the Artworld? «ACM Transactions on Multimedia Computing, Communications, and Applications», vol. 15, issue 2, 17 pages. Chung, S. (2020): ~ ( distance ) in place ~, Available at https:// sougwen.com/distance-in-place. Contreras-Koterbay, S. (2019): The Teleological Nature of Digital Aesthetics – the New Aesthetic in Advance of Artificial Intelligence, «AM Journal of Art & Media Studies», vol 20, pp. 105-112. Costello, D. (2017): On Photography: A Philosophical Enquiry, New York: Routledge. Du Sautoy, M. (2019): The Creativity Code: How AI is Learning to Write, Paint and Think, London: 4th Estate. Edmonds, E. (2018), Algorithmic Art Machines, «Arts», vol. 7, issue 3, pp. 1-7. 48 Creating Art with AI Elgammal, A. (2018): Meet AICAN, a machine that operates as an autonomous artist, Available at: https://theconversation. com/meet-aican-a-machine-that-operates-as-an- autonomous artist-104381 Elgammal, A, et al. (2017): CAN: Creative Adversarial Networks and Generating ‘Art’ by Learning About Styles and Deviating from Style Norms, arXiv: 1706.07068v1, pp. 1-22. Fu, S. (2020): Artist and Machine: An Iterative Art Performance, «Proceedings of xCoAx2020», pp. 384-388. Garcia, C. (2016): Harold Cohen and AARON – A 40- Year Collaboration: https://computerhistory.org/blog/harold- cohen-and-aaron-a-40-year-collaboration/ Gaut, B. (2018): “The Value of Creativity”, in: B. Gaut - M. Kieran (eds.), Creativity and Philosophy, New York: Routledge, pp. 124-139. Gaut, B. (2010): The Philosophy of Creativity, «Philosophy Compass», vol. 5, issue 12, pp. 1034-1046. Gaut, B. (2000): ““Art” as a Cluster Concept”, in: N. Carroll (ed.), Theories of Art Today, Madison: University of Wisconsin Press, pp. 25-44. Hagendorff, T., - Wezel, K. (2020): 15 challenges for AI: or what AI (currently) can’t do, «AI & Society», vol. 35, pp. 355-365. Hertzmann, A. (2018): Can Computers Create Art? «Arts», vol. 7, issue 18, https://doi.org/10.3390/arts7020018. Iseminger, G. (2004):The Aesthetic Function of Art, Ithaca, NY: Cornell University Press. Jeon et al. (2019): From rituals to magic: Interactive art and HCI of the past, present, and future, «International Journal of Human-Computer Studies», vol. 131, pp. 108-119. 49 Claire Anscomb Kelly, S. (2019): A philosopher argues that an AI can’t be an artist, https://www.technologyreview. com/2019/02/21/239489/a-philosopher-argues-that- an-ai-can-never-be-an-artist/ Kronfelder, M. (2009): Creativity Naturalized, «The Philosophical Quarterly», vol. 59, issue 237, pp. 577-592. Livingston, P. (2005): Art and Intention: A Philosophical Study, Oxford: Oxford University Press. Lopes, D. M. (2016): Four Arts of Photography: An Essay in Philosophy, Hoboken, New Jersey: John Wiley & Sons. Lopes, D. M. (2014): Beyond Art, Oxford: Oxford University Press. Manovich, L. (2019): Defining AI Arts: Three Proposals: https:// resonances.jrc.ec.europa.eu/documents/defining-ai-arts- three-proposals Mateas, M. (2001): Expressive AI: A Hybrid Art and Science Practice, «Leonardo», vol. 34, issue 7, pp. 147-153. Mazzone, M. - Elgammal, A. (2019): Art, Creativity, and the Potential of Artificial Intelligence, «Arts», vol. 8, issue 26, pp. 1-9. Maynard, P. (2000): The Engine of Visualization: Thinking Through Photography, Ithaca, NY; London: Cornell University Press. McCormack, J., et al. (2014): Ten Questions Concerning Generative Computer Art, «Leonardo», vol. 47, issue 2, pp. 135-141. McCormack, J., Gifford, T., and Hutchings, P. (2019): Autonomy, Authenticity, Authorship and Intention in computer generated art, «EvoMUSART 2019: 8th International 50 Creating Art with AI Conference on Computational Intelligence in Music, Sound, Art and Design», April 2019, Leipzig, Germany, arXiv:1903.02166. Mikalonytė, E. S. - Kneer, M. (2021): Can Artificial Intelligence Make Art?, SSRN: https://ssrn.com/abstract=3827314. Nake, F. (2014): Algorithmic Art, «Leonardo», vol. 47, issue 2, p. 108. Notaro, A. (2020): State of the Art: A.I. through the (artificial) artist’s eye, «Proceedings of EVA London 2020», pp. 322- 328. Ornes, S. (2019): Computers take art in new directions, challenging the meaning of “creativity”, «Proceedings of the National Academy of Sciences of the United States of America», vol. 116, no. 11, pp. 4760-4763. Phillips, D. (2009): Fixing the Image: Rethinking the ‘Mind- Independence’ of Photographs, «Postgraduate Journal of Aesthetics», vol. 6, pp. 1-22. Popa, E. (2021): Human Goals Are Constitutive of Agency in Artificial Intelligence (AI), «Philosophy and Technology», Available at https://doi.org/10.1007/s13347-021-00483- 2. Pranam, A. (2019): Putting The Art IN Artificial Intelligence: A Conversation With Sougwen Chung: https://www.forbes. com/sites/aswinpranam/2019/12/12/putting-the-art- in-artificial-intelligence-a-conversation-with-sougwen- chung/ Ridler, A. (2017): Fall of the House of Usher: http://annaridler. com/fall-of-the-house-of-usher Ridler, A. - Pfau, D. (2019): Bloemenveiling: https:// bloemenveiling.bid/ 51 Claire Anscomb Stecker, R. (2019): Intersections of Value: Art, Nature, and the Everyday, Oxford: Oxford University Press. Steinert, S. (2017): Art Brought to You by Creative Machines, «Philosophy & Technology», vol. 30, pp. 267-284. Zangwill, N. (2007): Aesthetic Creation, Oxford: Oxford University Press. --- Works in part. 1. Literalism as a position: fine. Define it first as “attributing genuine mental states, goals, and belief-like states to the system.” 2. Janus as an instance of literalism: usable with care. - His definition: “An agent takes open-ended actions to optimise for an objective. Reinforcement learning produces agents by default. AlphaGo is an agent.” That is literal agency in his taxonomy. - Do not imply “post-training RLHF ⇒ agents.” Janus does not say that. He treats RLHF as modifying a pretrained predictor, not as conferring agency by default. 1. The Anscomb text is **not** a second example of literalism. - She explicitly denies the relevant mental-state claims: “we are not yet at the stage where an AI can formulate intentions… I argue that AI agents cannot be artistically creative.” - She adds: “it is generally agreed that AI agents… do not have mental states, such as beliefs and intentions.” - Her use of “AI agent” is Bodensian (“a self-contained (‘autonomous’) procedure”), not person-level psychology. She grants iterative contribution to formal features and production credit, but not beliefs, intentions, or evaluative agency. Conclusion: classify her stance under *as-if/pseudo-agent* or “autonomous procedure,” not literalism. I won’t edit the canvas based on Anscomb. If you want a second literalism example, we need a source that attributes goals or belief-like states to the system itself, not just autonomy or contribution. --- 1\. "His definition: “An agent takes open-ended actions to optimise for an objective. Reinforcement learning produces agents by default. AlphaGo is an agent.” That is literal agency in his taxonomy. Do not imply “post-training RLHF ⇒ agents.” Janus does not say that. He treats RLHF as modifying a pretrained predictor, not as conferring agency by default." – I don’t really understand how to reconcile these two bullet points. 2. Okay, good. We might not present her view as literalism, but the fact that she says—and again, correct me if I’m wrong—that AI “contribute” (that’s her word) and “deserve credit” (that’s her phrase), makes it seem to me that this is actually a species of literalism, because only agents or person-like things can contribute something or be given credit. Think about this example, and if you like it, you can add it to the plan in a moment. You watch a movie; you watch the credits. The credits do not credit inanimate objects. They only credit persons—typically people—and things we’re sometimes inclined to think of as persons, like dogs. So that, to me, makes me think that she’s actually a literalist in disguise. Tell me what you think about this. If you agree with this contention, put all of this into the put all of this into the relevant part of the plan. btw, just a reminder: Transcription Input Handling Input Expectations Input will often be transcriptions rather than directly written text. Prepare to encounter and compensate for transcription-related errors and inaccuracies. Common Transcription Issues • Words that appear contextually odd or inappropriate • Phonetic substitutions where similar-sounding words replace intended terms • Semantic inconsistencies due to speech-to-text conversion errors Compensatory Approaches When encountering potentially mistranscribed content: • Clarify unclear terms: Request confirmation when words seem contextually inappropriate • Infer intended meaning: Consider what the speaker more likely intended to say based on: ▪ Surrounding context ▪ Phonetic similarity to logical alternatives ▪ Overall message coherence • Prioritize context over literal interpretation: Focus on probable intended meaning rather than exact transcribed words Implementation Maintain awareness that transcription errors are likely and be prepared to actively compensate through clarification requests and contextual inference to determine actual intended communication. --- 1. They reconcile cleanly. Janus’s “RL ⇒ agents by default” means RL in an environment with an action space and a reward signal (MDP control). RLHF on LLMs is preference fine-tuning on text outputs with no environment, no persistent policy acting in the world. So: RL-control → agents; RLHF-tuning ≠ that. Suggested canvas replacement for the first bullet (to avoid confusion): - *Agent* (literal): a system that takes open-ended actions to optimise an objective. In Janus’s taxonomy this comes from reinforcement learning in an environment (hence AlphaGo); RLHF on LLMs is not that. 1. No. Anscomb is not a covert literalist. She denies beliefs and intentions for current AI and denies artistic creativity; she uses “agent” procedurally (Boden) while allowing *contribution* and *production credit* in a normative, non-person sense. “Credit” does not entail personhood; film credits routinely list non-persons (software, labs, rigs, brands). Classify her under pseudo-agent/as-if or “autonomous procedure,” not literalism. If you want, I can add: - under *Pseudo-agent*: “Contribution/credit here is production-credit without mental-state attributions (Anscomb).” - and swap in the revised *Agent (literal)* line above. --- "They reconcile cleanly. Janus’s “RL ⇒ agents by default” means RL in an environment with an action space and a reward signal (MDP control). RLHF on LLMs is preference fine-tuning on text outputs with no environment, no persistent policy acting in the world. So: RL-control → agents; RLHF-tuning ≠ that." I’m still very confused. Please explain this in much, much more detail. Also, explain it. Is this a view that Janus disagrees with, then, or something that he does agree with? I’m baffled by what you’re trying to say with Janus, to be honest. 2. I’m glad you’re pushing back on this; it’s good for me. But first of all, you say film credits routinely list non-persons: software, labs, rigs, brands. But that’s metonymy, surely. It’s thanking the people at that software company, or at that lab, or at that rig, or at that brand. It’s not giving credit to a piece of software. For example, you’ll never see film credits saying “thank you to Ableton Live,” will you? You never see credits given to trees, do you? Or to buildings that play a big role in the plot. They might thank or credit the organisation that maintains the building, or the one that allowed them into the building. But they’re not thanking the bricks and mortar, are they? So this is why I think the charge of covert literalism sticks. That’s my strongest argument. The slightly less strong one is: it seems to me that when we say “contribution” of non-animate objects, we’re using it in a metaphorical way. This is much of a muchness. Credit 1 works perfectly well. --- 1. Explanation, slow and explicit. - Reinforcement learning for control (RL-control). Setup: a policy π takes actions a in states s of an environment. Actions change future states and rewards. Objective: maximise expected cumulative reward over open-ended interaction. Examples: AlphaGo in Go; robotics; Atari. Result: a persistent controller that acts to achieve goals. Calling this an “agent” is standard. - RLHF on LLMs. Pretrained model predicts next tokens. RLHF adds a *static* reward model that scores candidate texts by human preference. PPO nudges the *text distribution* so that, given a prompt, more preferred strings are sampled. There is no external environment, no actuators, no world state beyond the token sequence, no long-horizon goal outside “produce preferred text for this prompt.” Each episode resets at the prompt. Result: a *styled generator*, not a controller that acts in the world. - Why Janus’s two lines fit. “RL produces agents by default” refers to RL-control (policy acts in an environment). “Simulators” refers to predictors trained on next-token loss; RLHF merely steers such simulators’ *outputs*. Janus does not equate RLHF with RL-control, and he does not say RLHF turns simulators into agents. He treats RLHF as a modifier atop simulation training, not as conferring agency. Short form: RL-control ⇒ agent. RLHF-tuning ⇒ safer/nicer text, still a simulator. Janus agrees with that distinction. 1. On “credit” and covert literalism. - Anscomb denies beliefs and intentions for current AI and denies artistic creativity. She uses “agent” procedurally (Boden) and allows *production credit* for causal contribution to formal properties. That does not entail personhood. - “Credit” is polysemous. Moral credit presupposes an agent. Production credit and acknowledgements do not. Scholarly and artistic practice routinely assigns production credit to non-persons and non-minds (pipelines, datasets, cameras, “Filmed with Panavision,” “Dolby Atmos,” “Shot on Kodak 5219”). You call this metonymy; that still concedes the linguistic point: the grammar of “X deserves credit” does not force mental-state attributions to X. - Your movie-credits test shows a norm of *how* industries list thanks, not a necessary condition on the *concept* of production credit. Anscomb’s text makes clear she withholds the person-level predicates and uses “credit” to track non-accidental causal contribution to properties that ground value. That is consistent with a pseudo-agent/as-if classification, not literalism. If you want the canvas tightened to reflect this: Proposed replacement for the Janus line under *Agent (literal)*: - *Agent* (literal): a system that *takes open-ended actions to optimise an objective*. In this sense, reinforcement learning in an environment yields agents (e.g., AlphaGo). RLHF on LLMs does not do this; it steers a pretrained predictor’s text. Proposed addition under *Pseudo-agent*: - *Production credit without mental states* (Anscomb): systems can non-accidentally contribute formal properties and thus warrant production credit, while lacking beliefs, intentions, or evaluative agency. --- "Reinforcement learning for control (RL-control). Setup: a policy π takes actions a in states s of an environment. Actions change future states and rewards. Objective: maximise expected cumulative reward over open-ended interaction. Examples: AlphaGo in Go; robotics; Atari. Result: a persistent controller that acts to achieve goals. Calling this an “agent” is standard. RLHF on LLMs. Pretrained model predicts next tokens. RLHF adds a static reward model that scores candidate texts by human preference. PPO nudges the text distribution so that, given a prompt, more preferred strings are sampled. There is no external environment, no actuators, no world state beyond the token sequence, no long-horizon goal outside “produce preferred text for this prompt.” Each episode resets at the prompt. Result: a styled generator, not a controller that acts in the world. Why Janus’s two lines fit. “RL produces agents by default” refers to RL-control (policy acts in an environment). “Simulators” refers to predictors trained on next-token loss; RLHF merely steers such simulators’ outputs. Janus does not equate RLHF with RL-control, and he does not say RLHF turns simulators into agents. He treats RLHF as a modifier atop simulation training, not as conferring agency." Okay, so just to make sure we’re not talking at cross-purposes further: the Janus paper was written some years ago now. In mid-2025, as far as I understand it, we can think of LLMs as having at least these two principal training periods. One is the pretraining, which involves the corpus. And one is the post-training, which involves a lot of training on dialogue and a lot of training in terms of making these things helpful, harmless, and honest. Even if Janus is not saying that this is how you make persons, is it a common or plausible view that this sort of post-training makes agents? It doesn’t have to be a popular view, but just a view. You might need to search the web to work this out. In fact, I’m going to turn on your agency mode so you can help me understand this. Also, what Janus calls “agents” in this text—are they analogous to the sort of agentive AIs people talk about in 2025, or is he talking about something completely different? --- <video><source src="https://persistent.oaistatic.com/deep-research/d9a42348470e-bg.mp4" type="video/mp4"></video>