Bernabé Mallo
Doctor en Filosofía por la Universidad del País Vasco (UPV/EHU)
Investigador
en neurofilosofía, evolución humana y origen del arte. / PhD in
Philosophy – University of the Basque Country (UPV/EHU)
Researcher in neurophilosophy, human evolution, and the origins of art.
A review of the study by Yali Liu and Can Zhu (2025) in Scientific Reports: The use of deep learning and artificial intelligence-based creative teaching in art education
Introduction: when the machine looks at a painting
What does a machine see when it contemplates a painting? It experiences no aesthetic pleasure, is not moved by colours, evokes no memories. But can it learn to recognise what makes a work what it is? Can it decompose a brushstroke into its essential components —texture, colour, composition— and, from there, generate something new?
The study by Yali Liu and Can Zhu, published in 2025 in Scientific Reports (Nature Portfolio), addresses precisely this question. The authors propose and optimise an innovative artistic creation system called the Creative Intelligence Cloud (CIC), which combines generative adversarial networks (GANs) and convolutional neural networks (CNNs) to explore how deep learning can be applied to art education. The goal is not only to generate images, but to understand the process by which a machine acquires what we might call an "aesthetic knowledge".
This approach places us before a question that runs through our research on the origin of art: what does it mean to know a work of art? Is it enough to identify patterns, styles, and compositional structures, or does it require something more —a lived experience, a body that feels, a symbol that is inhabited— that the machine, for now, does not possess? The answer, as we shall see, forces us to distinguish between two forms of knowledge: the one based on statistical correlation and the one emerging from the single function of the nervous system.
The CIC system: anatomy of a visual learning process
The system proposed by Liu and Zhu is structured around two fundamental technical components: generative adversarial networks (GANs) and convolutional neural networks (CNNs). GANs, introduced by Ian Goodfellow in 2014, consist of two networks competing with each other: a generator, which produces images, and a discriminator, which tries to distinguish generated images from real ones. This adversarial competition forces the generator to progressively improve the quality of its creations.
CNNs, for their part, are the tool through which the system "reads" images. Unlike a traditional neural network, CNNs are designed to process data with a grid structure —such as the pixels of an image— and are capable of detecting local features (edges, textures, patterns) that are then combined into higher-level representations (forms, objects, compositions). It is, in a sense, an algorithmic imitation of how the human visual cortex processes information: from simple features to complex configurations.
The CIC system is trained on massive art repositories, such as the WikiArt dataset, which contains tens of thousands of paintings labelled by style, artist, and period. From this training, the model learns to decompose works into their constituent elements: brushstrokes, colour palettes, compositional structures. It does not "see" the painting as a meaningful whole, but as a matrix of features that it can analyse, compare, and recombine.
The "data lens": how the machine learns the visual language
The learning process of the CIC system illustrates what, in another review on this blog, we have called the "data lens" of artificial intelligence. Unlike a photographic camera, which captures the light of a real scene, AI does not see the world directly. It fragments billions of pre-existing images into minimal units of visual information —the so-called tokens— and associates them with linguistic labels (words such as "impressionist", "portrait", "blue"). In this way, the machine learns a "visual language of the world" that allows it, once trained, to move fluidly and autonomously across artistic movements, styles, and eras.
This process has profound implications. On the one hand, it allows unprecedented versatility: a model trained on WikiArt can generate an image in the style of Van Gogh, a hyperrealistic photograph, or a geometric abstraction with equal ease. On the other hand, it raises the question of what "originality" means when the system is trained on millions of existing works. The machine does not invent from scratch; it recombines what it has learned. Its "creativity" is, at bottom, a sophisticated form of statistical interpolation.
What the machine cannot learn: the symbolon gap
Here is where technical analysis meets philosophy. Liu and Zhu demonstrate that the CIC system is capable of analysing works of art, identifying styles, and generating new images. But does it understand what it does? The answer, from the perspective of our research, is negative. The machine can manipulate visual and linguistic tokens, but it does not inhabit the symbol. There is no one "inside" who recognises or is recognised in the act of creation.
This distinction is crucial. Symbolon, in its original etymological sense, is an act of recognition through shared codes. When a human artist creates, they do not merely combine forms: they project their interiority, communicate an experience, seek an encounter with the other. The work of art is a bridge between subjectivities. The machine, by contrast, produces objects that can be interpreted as art, but it does not generate the living process —the survival substrate, the symbolic dimension, the drive toward wholeness— that gives meaning to human creation.
The CIC system, by decomposing works into statistical features, loses precisely what makes a work a work: its character as an encounter. The machine can learn about art, but it cannot learn from art. It has no body that feels, no history that hurts, no need that drives it. Its knowledge is, at best, a shadow of human aesthetic knowledge.
A reading from the S/Y/C model
The study by Liu and Zhu offers a concrete case in which the functioning of the Law of Biological Coherence S/Y/C that forms the core of our research can be observed (Mallo, 2023, 2025, 2026a, 2026b). The three dimensions of our thesis unfold here with uncommon clarity, precisely because the contrast with the machine allows us to see what we take for granted in the human.
The S (Survival) dimension manifests in the usefulness of the CIC system. Its stated goal is to improve art education, provide automated feedback, and broaden access to creative tools. This is a legitimate adaptive function: technology, like Neolithic pottery, is adopted when it solves a practical problem. But the survival of the system is not that of the organism using it. The machine does not need to learn art in order to survive; the human who uses it does. AI is a tool for cultural survival, not a subject that survives through it.
The Y (Symbolon) dimension is the point of greatest contrast. The CIC system manipulates symbols —labels, styles, visual representations— with impressive efficiency. But it does not inhabit the symbol. There is no mutual recognition, no community of lived shared codes. The machine can classify a painting as "impressionist", but it does not know what it means to be impressionist, has not felt the light of southern France, has not argued with an academic about brushwork. The human symbolon is lived; the machine's is simulated.
The C (Wholeness) dimension manifests in the machine's search to close a form, to generate a coherent image. GANs, in particular, operate through a process of convergence: the generator and the discriminator compete until they reach an equilibrium, a point at which the generated images are indistinguishable from real ones. This convergence is a form of algorithmic wholeness. But it is a wholeness without yearning, without the biological drive that in humans seeks to integrate parts into a coherent whole. The machine "closes" a form because its loss function tells it to, not because it needs to make sense of its experience.
Surgical Philosophy invites us to make a precise analytical cut in the study by Liu and Zhu. It is not about rejecting technology nor idealising it, but about distinguishing levels. At the level of technical production, the CIC system is a powerful tool that can expand the capacities of the artist and the educator. At the level of aesthetic experience, AI cannot replace the embodied subjectivity that constitutes the heart of art. And at the level of cultural responsibility, an ethical framework is needed —such as the technological humanism we have discussed in other reviews— to ensure that these tools are used to enhance the human, not to replace it.
Final considerations: art as encounter, not as classification
The study by Yali Liu and Can Zhu has the merit of offering a rigorous and well-documented vision of how deep learning can be applied to art education. Their CIC system demonstrates that machines are capable of analysing, classifying, and generating images with a degree of sophistication that a decade ago would have seemed impossible. But, at the same time, their work reveals —perhaps unintentionally— the limits of that approach.
The machine can learn to recognise a style, but not to understand why that style matters. It can decompose a brushstroke, but not to feel what the brushstroke expresses. It can generate a new image, but not to inhabit the process of creation as an act of encounter with the other.
Art, in its origin and in its function, is not a classification of visual features. It is an encounter between embodied subjectivities. And that encounter, for now, remains the exclusive patrimony of the living. AI can be an ally in art education, a tool that broadens access and enhances creativity. But it cannot occupy the place of the artist, nor of the spectator, nor of the educator. Because art, like survival, symbol, and wholeness, is something that is lived before it can be measured.
Now, what does this research contribute to the whole of our inquiry into the origin of art? We believe that its most notable contribution is not only what it reveals about AI's capacities, but the comparison it makes possible. By observing how an artificial system, organised by us, decomposes and recomposes humanity's visual legacy, we are offered an unexpected mirror for thinking about how the biological system itself, tens of thousands of years ago, might have fragmented, assimilated, and reformulated the world to create something that today encompasses us all and everything.
Comparing these two ways of constructing creation —the artificial, guided by algorithms and data; the biological, guided by needs, biases, and yearnings— can be useful for consolidating a possible theory about how biological systems in general came to produce art. It is not about equating both processes, but about using the contrast as a heuristic tool: the machine shows us, through absence, what life contributes. And in that absence —the lack of a body that feels, of a history that hurts, of a need that drives— we find precisely the keys to what art, in its origin, was and continues to be.
References
Liu, Y., & Zhu, C. (2025). The use of deep learning and artificial intelligence-based creative teaching in art education. Scientific Reports, *15*, Article 15763. https://doi.org/10.1038/s41598-025-00892-9
Mallo, B. (2023). La construcción neuro-simbólica. Una aproximación al funcionamiento del cerebro desde una perspectiva multidisciplinar [Doctoral thesis, University of the Basque Country - Euskal Herriko Unibertsitatea]. ADDI Repository. http://hdl.handle.net/10810/62701
Mallo, B. (2025). Arte y biología: Una aproximación neurofilosófica al origen de la experiencia estética. https://www.amazon.com/dp/B0E8Y5WZMK
Mallo, B. (2025). Art and biology: A neurophilosophical approach to the origin of aesthetic experience. https://www.amazon.com/dp/B0E8Y6C2XN
Mallo, B. (2026a). De la filosofía quirúrgica a la ley de coherencia biológica S/Y/C: Hacia una investigación sobre el origen del arte en la especie Homo. https://isbn.bibna.gub.uy/catalogo.php?mode=detalle&nt=57196
Mallo, B. (2026a). De la filosofía quirúrgica a la ley de coherencia biológica S/Y/C: Hacia una investigación sobre el origen del arte en la especie Homo [Kindle edition]. https://www.amazon.com/dp/B0GYGTJD5C
Mallo, B. (2026b). From surgical philosophy to the law of biological coherence S/Y/C: Toward a study of the origin of art in the Homo lineage. https://isbn.bibna.gub.uy/catalogo.php?mode=detalle&nt=57197
Yu, C., & Liu, Z. (2026). Multi-task convolutional network for artwork analysis and personalized guidance. Scientific Reports, *16*, Article 70478. https://doi.org/10.1038/s41598-026-70478-6
Autor / Author
Bernabé Mallo
Doctor en Filosofía – Universidad del País Vasco / Euskal Herriko Unibertsitatea (UPV/EHU)
Investigador independiente en neurofilosofía, evolución humana y origen del arte.
Bernabé Mallo
PhD in Philosophy – University of the Basque Country / Euskal Herriko Unibertsitatea (UPV/EHU)
Independent researcher in neurophilosophy, human evolution, and the origin of art.
Enlaces / Links
Página de autor Amazon / Amazon Author Page: https://www.amazon.com/author/bernabemallo
ORCID: https://orcid.org/0000-0001-9002-9728
Plataforma EHUenRed / Link EHUenRed: https://www.ehu.eus/es/web/masterrak-eta-graduondokoak/red-latinoamericana-de-posgrados
Canal YouTube / Channel YouTube: https://www.youtube.com/@neuroideas815
Canal YouTube / Channel YouTube: https://www.youtube.com/channel/UCBsf6OZ482NjST6QA-hvYtQ
Publicaciones y proyectos en desarrollo / Publications and projects:
https://www.amazon.com/author/bernabemallo
https://ehuenred.theglocal.network/ideas/el-origen-del-arte-en-el-cerebro-de-makapansgat-al-moma-del-primate-al-sapiens