The free access to this article was made possible by support from readers like you. Please consider donating any amount to help defray the cost of our operation.
Human and AI Perspectives on Academic Aesthetics
Gioia Laura Iannilli and Ossi Naukkarinen
Abstract
The purpose of this article is to examine how current AI tools can be used in aesthetics research and teaching. We focus in particular on one theme: how traditional ways of describing the field of aesthetics as a whole could relate to descriptions made with the help of artificial intelligence. Describing the field and its main questions is unavoidable when we teach and research aesthetics; our question is, how can AI be used for that? Starting from a new meta-aesthetic tripartite model of aesthetics that we introduce, we raise a set of questions to pave the way to a more detailed examination of how new AI tools can impact academic aesthetics. We are particularly interested in what can—and what should not—be expected from AI models at the moment and whether experts in aesthetics should update their skill-set and approaches to the field.
Key Words
AIsthetics; ChatGPT; history of aesthetics; meta-aesthetics; prompting
1. How it started
One of the basic tasks of all academic disciplines is to outline their own main features. It cannot be avoided, for example, when writing introductory books. The aim is to explain what the field consists of, what is researched and taught, what kinds of questions are asked, what types of answers are given, and what approaches and methods are used. In aesthetics, this can be called meta-aesthetics, and forming such an overview requires plenty of experience and good understanding of different aspects of the field.
Useful divisions can be made in many ways, but typically some models become more popular than others, and they may be repeated for decades in textbooks and teachings. It has been customary to see aesthetics as one of the main subfields of philosophy and to divide this subfield, for example, into philosophy of art, philosophy of taste, philosophy of beauty, and philosophy of criticism (or meta-criticism), with numerous further subareas within these topics. Environmental aesthetics is often mentioned as a category of its own, although it overlaps with the others and some parts of it can be closer to natural sciences than philosophy.[1] The field of aesthetics can also be analyzed from chronological and geographical perspectives, and they are often used when writing about the history of aesthetics. The possible combinations are many, and they are further multiplied when philosophical schools to which single authors or groups of authors belong are taken into account.
As researchers and teachers of aesthetics, we wondered from time to time whether it would be possible to devise alternative ways that could shed light on the field in a slightly different manner and perhaps open fruitful alternatives. A few years ago, we came up with the idea that aesthetics could be divided into three main groups: 1) action- or verb-driven aesthetics; 2) object- or noun-driven aesthetics; and 3) quality- or adjective-driven aesthetics. We describe this tripartition in more detail in section 2.
Rather soon, we realized that the potential usefulness of our tripartite division should be tested with students, and we designed courses and workshops in which this could happen. After explaining our tripartition to the students, we asked them to apply the model and carry out qualitative-stylistic analyses of selected philosophical texts by John Dewey, Frank Sibley, and Arthur C. Danto.[2] We wanted to see whether the students could detect our three perspectives in these seminal texts. We also asked them to do this by taking into consideration specific aspects of their philosophical style, hence qualitatively and not quantitatively, that is, not by simply highlighting the number of occurrences of a specific word or types of words—verbs, nouns, and adjectives, for instance. We wanted to verify whether the tripartition made sense to the students and helped them to understand differences among real examples of aesthetics scholarship.
At the same time, artificial intelligence quickly was developing and becoming more accessible. We started to think about how it could be used to test our own model, refine it further, or to develop completely different ones. With the availability to the general public of Large Language Models (LLMs), and in particular the release of ChatGPT in November 2022 (various versions of it have been launched since), the scene changed completely. Before that, we had been thinking about and discussing with computer scientists what type of artificial intelligence or some other computer-based computational approach we could use for our purpose. We wondered, for example, whether and what kind of supervised or unsupervised machine learning would be suitable for this, and how we could organize a research project to test different approaches.[3]
However, it quickly became apparent that ChatGPT is one of the main AI tools that researchers, students, and many others will use.[4] Other tools based on LLMs such as Bard, Gemini, Llama, and Claude have since appeared, but we decided that, at least for this article, it makes sense to focus on only one, and especially one that is more widely known and used by the public at large.
How could we test and develop our own way of distributing aesthetics with ChatGPT, and what could we learn from such a process? It is clear that those in the field must learn to use such instruments intelligently, judiciously, wisely, and responsibly, because they are here to stay. As professors and lecturers, we have a duty to understand them and cannot exclude them from our own professional toolbox, while neither can we blindly trust them. The question is how to contribute positively to this unprecedented situation as humans and humanists, aesthetic beings and aestheticians? How do we combine our human approach, as professionals and students, with the latest technology?
So, pondering on how and if ChatGPT could contribute to developing and corroborating our model soon became a wider question on the status of academic aesthetics in the age of AI. We soon realized that in order to understand how AIs in general and ChatGPT in particular work, we needed to work closely with computer scientists; in turn, we realized that they needed us as domain experts to check and verify whatever results they might get.[5] This also made us wonder whether aestheticians ought to be trained differently nowadays in order to most effectively educate students.
It soon became evident that at this stage we have more questions than definite answers. And in fact, we only reached the setup point in our work; in many respects, implementation is yet to come. We, as many others,[6] are at a problematizing stage rather than at a stage where things are clear-cut and settled. This is why this article, too, includes more openings than conclusions. But, after all, it is also something that connotes philosophizing at its core: discussing and showing possible new ways of seeing and dealing with change.
2. Our human-made tripartite model
Through our human discussion process we ended up with a human-made model formulated through years of study, research, and teaching. We came to believe that the field of aesthetics could be divided into three areas: 1) action- or verb-driven aesthetics; 2) object- or noun-driven aesthetics; and 3) quality- or adjective-driven aesthetics. This is a model that to the best of our knowledge has never been suggested.
‘Aesthetics’ is a very loaded term, in the same way as, for example, ‘politics,’ ‘ethics,’, and ‘economics.’ In different contexts, they mean different things. That is why they must be defined again and again for various purposes, and there is no universally accepted understanding of them. Because the concept of ‘aesthetic’ is so widely used and influential, one should be aware of its different meanings.[7]
In our model, ‘aesthetics’ functions as an umbrella concept that covers discussions about aesthetic phenomena that can be organized into three partly overlapping categories in which a certain aspect or approach is emphasized. Moreover, we believe that those categories often indicate an author’s own way of writing about, conceptualizing, or philosophically expressing certain aesthetic phenomena.
2.1 Action- or verb-driven aesthetics
Action- or verb-driven aesthetics designates theories that focus on aesthetic activities that are typically referred to by verbs. They emphasize what and how an artist, for example, or an audience, or a person walking in nature does or makes; aestheticians try to illuminate what characterizes aesthetic activities in relation to other types of activities. Such approaches pay attention to what people do, make, and how they act when, for example, they create works of art, play music, do sports, cook, dress up, and look at or interact with something or someone. The activity itself is aesthetically defined and interesting or relevant. The point is not to focus on end results of such activities—objects such as artworks, for example—but on activities themselves.
This kind of approach can also be defined as process- rather than end product-oriented. John Dewey and some of his later followers such as Joseph H. Kupfer, and Friedrich Schiller long before Dewey, took this approach in their theories of aesthetics. Of course, they were not only interested in activities, but they emphasized their importance. A pragmatist, anti-essentialist approach, such as Dewey’s, is particularly emblematic in this framework. On the one hand, he adopted a strategy emphasizing the verbal form of a noun, as to stress its dynamic import in order to avoid its hypostatization. On the other hand, he claimed that any conceptualization must find its root and test-bed in concrete experience, hence, in practice. Concepts must in this sense be operative, rendering aspects that would otherwise remain implicit in experience more perspicuous, and hopefully positively impact them. Even expressions like “work of art” point in this direction. For him, ‘work’ is an activity taking place in the environment, a doing-undergoing, an interaction, that in some cases grows into a consummatory and active aesthetic experience. Another relevant example concerns the verbal acceptation he provides of ‘mind,’[8] thus meant not as a substance but as a verb, insofar as it has to do with an activity that concretely modifies the environment in which the organism lives. A further Deweyan indicator of this (almost obsessive!) action- or verb-oriented aesthetics has to do with his usage of the present continuous “ing” form, which emphasizes the processual, continuist import of his conception of aesthetic experience.
Speaking of an action or verb-oriented approach can be analyzed by aestheticians in many ways, not only in Dewey’s manner. For instance, other activity- or verb- driven approaches to aesthetics may be found in David Hume and his modern-day followers who analyze development of taste through activities, in somaesthetics initiated by Richard Shusterman, in Jacques Rancière and others who focus on political activities intertwined with aesthetics, and in the work of scholars who address potential aesthetic activities of animals by observing what they do, such as David Rothenberg.[9] A further, more general and recent area of application could potentially be found in enactivist approaches to aesthetics,[10] which are partly piggybacking from the Deweyan active-passive conception of the relationship between mind and environment.
2.2 Object- or noun-driven aesthetics
Some other aestheticians pay more attention to objects (broadly speaking), and thus for them systematizing nouns is important. Many definitions of art come close to this approach, as do analyses that stress genres. Objects can be single pieces of visual art, events, performances, entities in nature, or even categories. Knowing what one is approaching is seen to be essential for guiding a well-reasoned understanding process. This basic approach has been presented, for example, by Kendall L. Walton in his “Categories of Art,” where he points out that it is crucial to be able to position a work of art in a correct group of objects so that we can understand and evaluate it in a fruitful way.[11] Historically, many essentialist definitions listing supposedly objective features or properties of artworks by authors such as Clive Bell can also be seen from this perspective, and it is present even in Aristotle’s Poetics. It is clear that in some cases objects can be close to processes or activities, but in the objects case we more often are talking about defined and completed processes with more or less clear boundaries. In contemporary aesthetics, this approach can be said to have its strongest roots in the analytic tradition. The analytical philosophy of art initially characterized itself as a “meta-criticism” interested in the language used by art critics, and progressively focused on defining what art is. Monroe Beardsley, Ernst Gombrich, Richard Wollheim, Nelson Goodman, and Arthur C. Danto have provided seminal texts in this framework. However, this kind of demarcation is also action, a process, and so there is no absolutely strict division between the two approaches, but rather a difference in emphasis. Arthur C. Danto’s aesthetics is exemplary in this sense,[12] in that an interpretive activity produces works of art, that is, objects that are different from non-art objects, or mere real things.
2.3 Quality- or adjective-oriented aesthetics
The third approach is interested in qualities or qualitative properties. The starting point is that certain types of properties would be particularly aesthetic: beauty, ugliness, cuteness, and so on. This approach uses lots of adjectives and manifests a quality-oriented emphasis. An example of this is Frank Sibley’s aesthetics, most famously his classic article, “Aesthetic Concepts,” which analyzes how aesthetic qualifiers, which are adjectives, are used. Also Göran Hermeren’s The Nature of Aesthetic Qualities, Jerrold Levinson’s Contemplating Art: Essays in Aesthetics, and Thomas Leddy’s The Extraordinary in the Ordinary (partly) analyze how such properties are described and handled in the aesthetic discourse, spanning art criticism and everyday language. Umberto Eco has published illustrated books on both beauty and ugliness, and older classics, such as by Immanuel Kant and Karl Rosenkranz, have addressed similar themes, as have contemporary authors such as Sianne Ngai in her book on aesthetic categories and Max Ryynänen and Anna-Sofia Sysser when analyzing qualities of tropical kitsch. One of the interesting and tough questions for this type of aesthetics is, do such qualities actually exist? Over the course of history, opinions have differed between extreme objectivism and subjectivism and everything in between. Of course, this way of seeing aesthetics is close to the philosophy-of-beauty approach, but covers a wider conceptual area than just beauty.[13]
2.4 A dynamic meta-aesthetic model
These three approaches can be intertwined. They can be used side by side not only academically but also when we discuss aesthetic phenomena in everyday life: “It is great to sip this tasty coffee freshly brewed with my new coffee maker!” They can only be analytically separated. But they reveal different aspects of the whole, and that is why none of them can completely replace others. While some aestheticians or theories are more strongly representative of one of these, typically the two other perspectives are also involved to some extent. Some sort of activities are needed when we create or make objects, and objects can be categorized as art, or deeds as aesthetic activities, in respect to their qualities—and qualities can be qualities of some sort of objects or situations. Nonetheless, the tripartition can be useful in analyzing and discussing different strands and aspects of aesthetics.
It cannot be stressed enough that our tripartition is based on our personal experiences and the perception of both academic aesthetics where aesthetic phenomena are analyzed; and non-academic aesthetic phenomena, arts and otherwise. And while it seems to make sense to us to address aesthetic phenomena by using these three main approaches while also interlacing them, or using them side by side, the tripartition also describes our aesthetic language, that is, how we speak and write about aesthetic phenomena.[14] Meta-aesthetic sketches can be used to examine the field of aesthetics today as well as the history of aesthetics, and a good model must be fit to do just that. It would be an interesting project of its own to systematically explore the history of aesthetics in more detail from the perspective of how the positions epitomized by the three approaches have varied in different contexts: when and why is there perhaps more interest in action, when in object-thinking, and when in quality? Was eighteenth-century aesthetics, for instance, more “verbal,” “substantival,” or “adjectival”?
Nevertheless, in our discussions and workshops with students the model seemed to work, as with its help they could rather easily understand how to classify and analyze different texts.[15] Moreover, it also appears to describe the field of aesthetics in a slightly different way than other models, that is, it does seem to renew the field to a certain extent. And it by no means replaces previous ones, but complements them. Different models seem to be fruitful to use crosswise as well.
How did we end up with this model? As said, it is not easy to retrace the whole process. We did not have a clear method for producing it; it just evolved in our common discussions. We have both been active in the field for a long time, studying it, reading a large amount of literature, discussing with colleagues, teaching, attending conferences, and participating in associations in the field. In other words, we have formed our model through a longtime exposure to aesthetic discourse. All of this has given us an idea of what aesthetics as an academic discipline is, and for some reason a few years ago it just started to feel like we could see it a little differently. And even though this might sound dubious, we believe this is probably how academic thinking has always worked and still is fairly common in humanities.
3. Toward an AI-assisted model
We had a human-made model that seemed to be understandable and useful in discussions with students, but it had yet to be circulated to the academic community for assessment; this article is the first published version.
As a matter of fact, we suggest our model. When we were developing it, we had lengthy discussions about how to test whether it worked. An evident problem that came up is that when one has the basic idea of the tripartition, it very easily turns into a self-fulfilling prophecy. Once you understand the idea, you start seeing examples and cases that fit into it—which is typical in philosophy in general. That is not testing a hypothesis in any strict, methodological, or empirical way. It simply is having an idea and seeing whether there is something that could support it. When one is selective, there usually is. So, besides our more traditional ways to tackle the issues, we felt the need to resort to other, data-driven kinds of tools in a more structured way.
This consideration brings us to a wider issue, namely the fact that we have entered the era where different ways of answering the question, “What is aesthetics?” are almost inevitably affected by various digital tools, sometimes by artificial intelligence. We, human beings, create our interpretations also with the help of computational machines. Before this era, our conceptions were formed more by listening to lectures, attending seminars and conferences, reading books and articles, observing the world around us, and discussing with our colleagues. Not long ago, even libraries functioned with paper-based classification tools, and there was no internet, no personal computers, no smart phones.
Now, computer-based tools to gather and organize information are everywhere. Moreover, we are dealing with fast-changing and updating technologies that affect academia more directly than before, hence also more directly the way we work. Whether we are first-year students or retired academics, it is crucial to understand at least the most important ways how this can happen.
The typical process probably is that we continue the old-school way of reading books, discussing with our colleagues, attending conferences, and so on. However, at the same time, we Google things, make use of various sorts of digital materials, and use Word, Pages, and other software to formulate our texts, while sometimes forgetting that there is, for example, lots of materials in libraries and archives that have not been digitized and are not on the internet. Moreover, the digital academic universe is dominated by materials in English, and especially less widely used languages don’t have much visibility.
Over the past decades and years, we have also seen the rise of something different. We can make use of digital tools in a much more systematic, targeted, and customized way. When you think, “What is aesthetics?” you can use computers to do things you could never do yourself as a human being. You can use a massive body of data—texts, speeches, pictures, and videos—that other people and human-machine pairs have created on aesthetics and let the computer analyze it. Of course, it won’t analyze it all by itself, for it needs your input, the parameters you set.
We initially played with the idea that we could use computational methods for testing our model. Could an algorithm find similar categories in a body of literature? Could we teach an AI to detect them? Yes, probably we could—but then the algorithm or AI would do exactly what we’d ask it to do. If we teach it to find certain kinds of things, it will find them. So, it would just repeat our ideas in an effective way. And this is what happened when we tested various AIs ourselves: after a few attempts they gave us a somewhat acceptable answer for our standards. With them, we ended up in self-fulfilling prophecies probably more easily than through human reasoning, and that is why AI cannot be used straightforwardly to verify our ideas.
At the beginning of this project, we might have let an algorithm categorize a body of literature in its own way through topic modeling, which still would have required us to select the parameters of the algorithm—in that respect, it is quite dependent on us, too. But it would still categorize the data differently than our preset AI model. Depending on parameters, it would have a different number of categories, and they would contain different things. Would this approach produce something like our tripartition? Hardly. It could produce something else, but equally useful—but how could it be used to test and examine our model, if at all possible? At this stage we were already in discussion with a computer scientist, who was providing us feedback on our questions and needs, and our attempts were not escaping that feared “self-fulfilling prophecy” mentioned earlier. Nor was the approach clarifying how certain processes would work, which is a crucial aspect if one aims at understanding them outside of their “black box.” On the one hand, our humanmade model can overrule the computational process and thus add nothing to the understanding of aesthetics; on the other, the AI-created model could be something completely different and detached from it, again adding nothing to understanding and examining our model. How to proceed then?
Our initial aim was to be able to create a matrix for the three main approaches starting from three classics of aesthetics that exemplify each of them—John Dewey, as a verb-oriented aesthetician, Arthur C. Danto, as a noun-oriented aesthetician, and Frank Sibley, as an adjective-oriented aesthetician, and then test what “reaction” or results its application would produce when used with some of the most recent body of literature available in academic aesthetics.[16] We would have thus used the same texts we discussed with our students as base-texts, offering reference points to a computational process analyzing digitized materials in aesthetics available online. Could we computationally categorize contemporary publications in this way? If yes, what could we learn from that, and would there be a way to really see what the used computational tool would do in detail? There constantly seemed to be the risk of a self-fulfilling prophecy and also that of the black box, that is, that there would be no way of really knowing how the selected computational tool, AI or otherwise, would come up with the results it provided.[17] Yet, trying to find a way out of these risks made us realize that we can actually aim at a larger scale analysis on the status of academic aesthetics today and in the future.
4. The AIsthetics Project
Much of what is described above was underway before the release of ChatGPT, but it was already clear that the problems and the shortcomings we encountered in the first phase of our project were productive in at least directing our path. As soon as we realized that the kind of analysis the method we envisaged provided would only scratch the surface of our problematic field, we realized that we needed an even more articulated plan. We hence involved in the project—that we named AIsthetics—two colleagues from computer science,[18] who also have philosophical and linguistics training and a specific interest topics like knowledge extraction and pattern recognition from data, ontology design, and ontology engineering (and that’s why they can also be called “knowledge engineers”).
At the same time, we enlisted a group of interns active in the Almæsthetics Research Center[19] in a more systematic qualitative-stylistic analysis of the philosophical texts we had previously only intermittently experimented with. Our aim was to provide our two computer science colleagues material to work on, with the idea of comparing a humanmade and an AI-made qualitative-stylistic analysis of philosophical texts, while trying to avoid the pitfalls of our previous attempt. We tried to find more refined techniques of knowledge representation, aiming at preserving the highest possible amount of semantics from human analysis, and represent it formally to reinject this knowledge into AI systems. At this stage, it was unclear what kind of AI-made model would be created, whether ChatGPT would have a role in it, and how it would relate to our human-made model. It quickly became very clear that finding a fruitful way of making use of the latest technologies in aesthetics is far from simple.
Our interns still play a crucial role in this rather lengthy and slow process.[20] We had asked a few students to carry out a qualitative-stylistic analysis of the selected philosophers and their texts in their original language, as to preserve as much as possible their expressive features that sometimes get lost in translation. They were asked to provide analyses that would basically show where, how, and why such and such author could or should be categorized as an action- or verb-oriented, object- or noun-oriented, and quality- or adjective-oriented aesthetician. The same task was now presented to a group of interns. But before letting them proceed with the analysis, which would end up in the form of a traditional paper, we had to teach them our model and to recognize instances of our tripartition. In doing so, they became domain-aware, and domain annotators, supervised by us, the domain experts. In other words, we had to create what in the framework of machine learning is called a ‘ground truth,’ which consists of information that is known (or agreed) to be true or valid. In data analysis, ground truth designates the portion of data that is annotated by domain-aware actors, and therefore can be considered reliable and valid. In our case, it was the model we were assuming as working in and informing our research, which would also be employed to train or validate a machine learning model.
This more traditional way of analyzing a philosophical text had to be transformed into an Excel sheet as AI experts suggested, because the more discursive analysis would not be digestible enough when later fed to ChatGPT. This more traditional way could work in a pseudo-Socratic dialogical context, namely prompting LLMs such as ChatGPT, and receiving answers. Although LLMs have incredible generative language capability, they pose problems of two kinds. The first set of problems relate to their coherence: generative AI systems are usually nondeterministic; therefore, even posing the same question with the same phrasing in two different chats—or even in the same chat, but in two different moments—could result in different answers, with a different structure, and possibly conflicting content. The second relates to reuse: in order to fix a conceptual structure that would allow a meta-aesthetic analysis independent of the author, computer scientists must create a more structured methodology. That’s why tabular data finally was adopted as good midway, allowing domain experts and annotators to fill in with glosses and comments, leading the way to a shared conceptual model reusable for several cases. It is important to note that the final articulation of the sheet is the outcome of a long mediation between the “needs” of the aesthetics domain experts and annotators, who were used to lengthier and more discursive analyses, and the AI experts in need of something more synthetic but that would also make the nexuses among the various parts of the argumentation more perspicuous for the machine learning process. In addition, tabular data are a very common data shape; therefore, LLMs are perfectly able to read and interpret correctly the semantics of cell values.
The headers of the tables are the elements that we (aestheticians and AI experts) think are going to be pivotal for further experiments with ChatGPT, and they are a translation of the indications provided in order to carry out the traditional qualitative-stylistic analysis in the first place.
Let’s consider the table available at the following link, where readers will find an example of the analysis carried out by interns (namely of Sibley’s classic essay, “Aesthetic and Non-Aesthetic”). From its articulation,[21] it is plain to see that understanding the logic of the Excel sheet and its relation to the reading exercise the interns went through is not easy. Nor is it easy to take the next step: apply the Excel elements—that still need to be reworked through another process, as we explain below—to an AI-run process, where ChatGPT or some other AI makes use of them to process text masses in aesthetics to find patterns and thus detect styles of authors. The point of showing the table here is exactly this: using the latest technologies in a deeper way is far from automatic and easy.
In addition to the Excel sheet, our AI experts suggested a further process, namely the formulation of a diagram or conceptual map providing a visualization for better representing elements and nexuses included in the table. An example concerning the same analysis of Sibley’s essay can be found here. Tables are in fact a machine-readable format, but humans do not typically organize their cognitive information in columns and rows. In fact, humans seem to be more prone to adopt something similar to a diagrammatic form, resembling the mathematical structure of a graph (so-called ‘schemas’ in cognitive psychology suggest this).
When we conducted this first part of the project, we had limited human resources, and our interns each analyzed one chapter from the three exemplar works. Of course, to have more reliable data to work on, we would need multiple analyses of each single chapter, but for the time being we were able to have interns doublecheck each other’s analysis, providing further corroboration.
The interns also provided feedback about the pros and cons of working in these three ways: a traditional philosophical paper; an Excel sheet; and a conceptual map. This is an interesting way of having a glimpse on what philosophy students might expect or not expect in their philosophical education. Generally speaking, the most familiar and comfortable way of carrying out their analyses was, quite unsurprisingly, through the more discursive or traditional procedure. The least appreciated one was the translation of their previous analyses into the Excel sheet, as this activity was considered too restrictive and not providing “enough room” to detect the stylistic and qualitative aspect that they were required to identify. Some interns agreed, however, that this step helped them to develop an ability to be more straightforward, synthetic, and clearcut in their work, and in certain cases allowed them to understand better certain aspects of the authors. Finally, some found the conceptual map an interesting exercise, perhaps easier to develop than the more rigid Excel sheet, as it was closer to the way they sometimes organize the material they are required to study for their exams.
Working closely and more systematically with interns, our interest in understanding at least some potentialities of AI for academic aesthetics expanded further. We even started thinking of producing a guide that would allow for a more critical usage of AI tools when dealing with aesthetics. Yet, the work around this has actually just started and won’t be finished any time soon. Much more experimentation still must be done. Some further possibilities are explained more in detail by our colleagues in their article, “The Beauty and the Bit: AI-driven Philosophical Aesthetics,”[22] included in this volume, and two examples of possible future directions can be given here:
First, we can aim at a formal knowledge-based comparison. This means making explicit the “conceptual components” of a certain theory—namely all the elements that are mentioned by an author as being fundamental for their theory—and having them organized in a formal structure that individuates the complex network of relations among these components. This allows one to sketch a map of the conceptual elements of a certain theory and author. As a mind-map represents an abstraction of a certain idea, conceptual schema, or theoretical system, so does this kind of ontological formalization of semantic relations among entities in a certain theory. Furthermore, when several theories are represented with the help of a common formalizing template, it allows for a smooth formal comparison of theories and interpretations of matters that might otherwise remain vague.
Second, we can develop an AestheticsLLM or a refined version of an LLM with a certain aesthetic “sensibility.” This could deepen the analysis of LLM’s capabilities, and it could be done mainly in two ways:
1. By providing more background information to the LLM directly via chat, with the aim of creating a more precise context, and reducing the risk of vagueness or so called “hallucinations.”
2. Another, more radical, path to deepen the analysis could be done via fine-tuning a dedicated LLM. This would mean directly modifying the internal representation of the information, creating a more expert LLM “for specific domain use,” in our case for scholarly aesthetics analysis. ‘Fine tuning’ is a technical expression that consists in a (partial) retraining of the model. The information contained in an LLM, in fact, is represented internally in a vectorial space. The LLM, once trained, has “learned” certain patterns. A (partial) retraining consists in providing further material from a specific domain, in this case, aesthetics, to instruct the LLM to behave in a more domain specific way. In this way, the LLM will show more familiarity and “sensitivity” towards the desired domain, as if it was an expert.
In both cases, the tool we are envisaging should be able to make sense of contemporary academic aesthetics, its history, as well as nonacademic aesthetic phenomena such as art or taste choices in everyday life, through our tripartite meta-aesthetic model.
Broadly speaking, clever prompting seems to be key.[23] Simply asking questions as the ones we would ask each other in human-academic conversations does not work well, as the AI would generally tend to be either pleasing, namely eventually providing an answer that we want to hear,[24] or biased, depending on how the AI was devised at its core.
This emphasis on prompting partly answers our initial questions on what aestheticians should do and what skills they need to have nowadays to sensibly deal with AI. Examining our human-made meta-aesthetic model served as an example of what this means and requires in practice. What counts in such interactions, it seems to us, is not necessarily what is asked but how something is asked, and prompting provides just that: a way of asking certain things that would somehow articulate or even critically push the conversation in a specific direction, trying to avoid the pitfalls of a self-fulfilling prophecy on the one hand, or a biased response on the other, or even, most importantly, to avoid complete trust, belief, and automatism when dealing with AIs, but to develop, learn, and train oneself to a progressive understanding of criticality and awareness towards them. The importance of developing these skills in academia today, and more specifically when researching, teaching, and learning aesthetics, is also proven by the fact that AIs clearly need to be constantly instructed and checked at various levels. We are trying with our project to understand what AIs can and cannot be used for and what human skills are still needed, both in terms of prompting, of instructing the AI, and checking the results obtained. This is not an easy task, and lengthy analysis and interdisciplinary process has proved necessary. It is clear that the traditional aesthetician’s skills need to be integrated with new skills, which can be obtained only in a cooperative and technologically competent context. It is true that there will most probably always be black boxes in computational processes that we don’t fully understand—but it is equally true that there is no way to understand and follow all details of purely human interaction either. In many respects, we are black boxes to ourselves and each other. As regards aesthetics, we can try to make all the boxes a bit less opaque, with all the tools we have at our use.[25]
Gioia Laura Iannilli
gioialaura.iannilli2@unibo.it
Gioia Laura Iannilli is Associate Professor of Aesthetics at the University of Bologna. She served in the Executive Committee of the Italian Society of Aesthetics and as the Secretary of the Experience Research Society. She has authored or edited the books John Dewey. Il senso delle qualità: saggi sulla percezione (2024), Co-operative Aesthetics. A Quasi-Manifesto for the 21st Century (2022), The Aesthetics of Experience Design. A Philosophical Essay (2020) and L’estetico e il quotidiano: design, Everyday Aesthetics, esperienza (2019).
Ossi Naukkarinen
ossi.naukkarinen@aalto.fi
Ossi Naukkarinen is Professor of Aesthetics and former Vice-President of Aalto University, Finland. He is President of the Finnish Society for Aesthetics and the author of the books, Aesthetics as Space (2020), Art of the Environment (2007), and Aesthetics of the Unavoidable (1998), and he has also published books and articles in, for example, Finnish, Italian, and Chinese. He currently is finalizing a book addressing the nature of the humanities.
Published on July 14, 2025.
Cite this article: Gioia Laura Iannilli & Ossi Naukkarinen, “Human and AI Perspectives on Academic Aesthetics,” Contemporary Aesthetics, Special Volume 13 (2025), accessed date.
Endnotes
[1] Different versions of describing the field can be found in the following books that have been widely used for some time and have thus had considerable impact: Colin Lyas, Aesthetics (London and Bristol: UCL Press, 1997), Marcia Muelder Eaton, Basic Issues in Aesthetics (Belmont: Wadsworth Publishing Company, 1988), Robert Stecker, Aesthetics and the Philosophy of Art: An Introduction (Lanham: Rowman & Littlefield Publishers, Inc., 2005), Richard Eldridge, An Introduction to the Philosophy of Art, Second Edition (Cambridge: Cambridge University Press, 2014).
[2] John Dewey, Art as Experience, LW, Vol.10 (Carbondale and Edwardsville: Southern Illinois University Press, 1989); Arthur C. Danto, The Transfiguration of the Commonplace (Cambridge, MA: Harvard University Press, 1981); and Frank Sibley, Approach to Aesthetics. Collected Papers on Philosophical Aesthetics, ed. by John Benson et al. (Oxford: Oxford University Press, 2001).
[3] An earlier example of this interest can be found in Ossi Naukkarinen, Aesthetics as Space, Chapter 4 (Espoo: Aalto ARTS Books, 2020), where the author experimented with topic-modeling. Before ChatGPT was released, we made some initial tests with our colleague Wenjie Fan. They were very useful steps forward, but did not result in a functional solution.
[4] The relevance of this tool is so recognized (and also prone to be exploited) that, for instance, a specific, first version of ChatGPT specifically for education was released in June 2024.
[5] In knowledge engineering, a “domain expert” is a person who is knowledgeable in a specific field, someone who is assigned to the production or the supervision of the production of data that then is used to perform any kind of further analysis or elaboration.
[6] A recent, widely read book addressing potential AI-driven futures largely through questions and alternative scenarios is Genesis: Artificial Intelligence, Hope, and the Human Spirit by Henry A. Kissinger, Craig Mundie and Eric Schmidt (London: John Murray, 2024).
[7] Our own point of view is theoretical, but aesthetics has a concrete and fundamental role in most people’s everyday life dynamics. However, as such, it should also be accessible thematically, and not only “practiced implicitly” by people. So, one of the goals of this research is also, eventually, to make aesthetics more widely understandable, as to generate some sort of “aesthetic literacy” extra-academically by providing new ways of perceiving it.
[8] Dewey, Art as Experience, 268.
[9] Jacques Rancière, Le partage du sensible: esthétique et politique (Paris: La Fabrique, 2000), David Rothenberg, Survival of the Beautiful (London: Bloomsbury, 2012).
[10] For an overview, see John M. Carvalho, “Enactivism and Aesthetics,” in Routledge Encyclopedia of Philosophy (Milton Park: Taylor and Francis, 2021).
[11] Kendall L. Walton, “Categories of Art,” Philosophical Review 79, no. 1 (1970): 3, 34-367.
[12] And, interestingly enough, while being clearly interested in defining the essence of the objects of his analyses, Danto also wrote an Analytical Philosophy of Action (Cambridge: Cambridge University Press), 1973, proving that it is not necessarily (or at least exclusively) the subject of a research that defines the style of an author. In this case, we have a “substantive” philosophical style applied to the topic of action.
[13] See Göran Hermeren, The Nature of Aesthetic Qualities (Bromley: Chartwell-Bratt, 1988); Jerrold Levinson, Contemplating Art: Essays in Aesthetics, Section V (Oxford: Oxford University Press, 2006); Thomas Leddy, The Extraordinary in the Ordinary (Peterborough: Broadview Press, 2012); Sianne Ngai, Our Aesthetic Categories: Zany, Cute, Interesting (Cambridge, MA: Harvard University Press, 2012); and Max Ryynänen & Anna-Sofia Sysser, “Making Sense of ‘Tropical’ Kitsch,” Contemporary Aesthetics, Volume 19 (2021), https://contempaesthetics.org/2021/01/08/making-sense-of-tropical-kitsch/.
[14] Of course, other labels could be used. For instance, if we talk about a verb-, substantive-, and adjective-oriented aesthetics, why not about an “adverbial” one? In fact, in philosophy something referred to as ‘adverbialism’ does indeed exist (see at least Pierre Steiner, “Pragmatism in Cognitive Science: From the Pragmatic Turn to Deweyan Adverbialism,” Pragmatism Today, 8/1 (2017), 9-27). In the framework of theories of perception, it indicates the “modal” component of the latter, namely, emphasizing the “hows” and not the “whats” of perceptual experience. Why is this relevant to our discourse? Because it clearly has to do with the question of aesthetic qualities (it is one way to deal with it) on the one hand, but also with the question of the processes through which experience is carried out actively, hence showing an overlapping of possible approaches characterizing one single author. For instance, a A) verbal, B) adjectival or C) adverbial aesthetics would be equally fitting for Dewey’s aesthetics since, as for point A) see the discussion above, and as for B) and C) in his writings he maintained that our experience with the environment is always qualitatively connoted, never “neutral,” and it orients our action in certain ways; but also that in order to express such experienced qualitativeness, we often resort to adjectives and adverbs.
[15] It is particularly interesting to note that based on the provided instructions, they were able to classify one author who seemed to be more straightforwardly emblematical of a type of aesthetics in a different but totally reasonable and thoroughly explained way.
[16] For this purpose, we thought of three exemplary forums of aesthetic discussion: The British Journal of Aesthetics, The Journal of Aesthetics and Art Criticism, and Contemporary Aesthetics because they mostly are available online and published in English. We thought that focusing on just one language would allow us a more homogeneous basis to ground our research on, and choosing e-journals (although some of them are also available in a print version), offered more freedom in applying digital tools and aids. Yet in the future, testing the model with other languages and in different cultural contexts is necessary and might open up completely new aspects of aesthetics.
[17] In principle, the “reasoning processes” of AI systems and other computational tools can be opened up by writing an algorithm that would make the system explicate all the steps it takes. However, in practice, it would be an extremely laborious exercise to go through all of them, and it is not possible to add such a function in, say, ChatGPT controlled by OpenAI.
[18] Aldo Gangemi and Stefano De Giorgis.
[19] https://centri.unibo.it/almaesthetics/it/tirocinio-alm-sthetics/aisthetics-project.
[20] Starting from the submission of our research project for evaluation to the University of Bologna Bioethics Committee, which granted us permission to proceed, but just this step took us several weeks to complete.
[21] In the example we provided, readers will find an explanation of each header in the notes section we added (see upper right corner in each header). A more technical overview will be provided by our colleagues in their paper.
[22] “The Beauty and the Bit: AI-driven Philosophical Aesthetics” provides a focus on Danto’s first Chapter of The Transfiguration of the Commonplace, along with access to both an AI-assisted and a humanmade analysis (via graph) of it is provided. We refer readers to the Appendix they provide in their article to view this material.
[23] Here’s a few links to articles, cheat sheets, guides, introductions, and so on to prompting: https://www.kdnuggets.com/publications/sheets/ChatGPT_Cheatsheet_Costa.pdf; https://dl.acm.org/doi/10.1145/3560815; https://www.promptingguide.ai/introduction; https://www.reddit.com/media?url=https%3A%2F%2Fpreview.redd.it%2Fthe-complete-chatgpt-cheatsheet-v0-ydhr03od285b1.png%3Fwidth%3D3000%26format%3Dpng%26auto%3Dwebp%26s%3D0a1b3f2564933897bd6d9a2654c7a21603579106; from Open AI/ChatGPT: https://platform.openai.com/docs/guides/prompt-engineering; from Google/Gemini: https://ai.google.dev/gemini-api/docs/prompting-strategies.
[24] See, for instance, this interesting article on how tagging and geographical areas where tagging takes place can really affect the human-AI interaction and the vehicular language of this interaction, but also how behind AI surface there is a whole human exploitation reality taking place especially in certain areas of the world: https://www.theguardian.com/technology/2024/apr/16/techscape-ai-gadgest-humane-ai-pin-chatgpt.
[25] Although the whole paper and each section have been planned, discussed, and co-written by the authors in detail, in the final version, Gioia Laura Iannilli authored sections 2, 2.1, 2.2 and 4, while Ossi Naukkarinen authored sections 1, 2.3, 2.4, and 3.