Abstract
This paper examines the intersections between Kensuke Koike’s collage practice and convolutional neural networks (CNNs), proposing a dialogue between human creativity and machine vision. Koike’s method, defined by the rule “nothing added, nothing removed,” decomposes and recomposes single photographs to generate aesthetic doubt in viewers, unsettling perception and inviting reflection. CNNs, by contrast, decompose and reassemble images through layered filters and probabilistic computations, producing classifications or segmentations without aesthetic intent. By comparing these processes, the paper highlights formal similarities, such as segmentation, recomposition, and constraint, while emphasizing their divergent engagements with uncertainty. Koike cultivates doubt as an artistic and conceptual force; CNNs reduce uncertainty mechanistically. This contrast informs of the distinct roles of doubt and automatism in human–machine creativity, contributing to a broader philosophical inquiry into technological humanism and the evolving aesthetics of the post-digital age.
Keywords
Kensuke Koike, Convolutional Neural Networks, Doubt, Philosophy of AI, Classificators.
1. Introduction
In an era in which AI is everywhere in our lives, the relationship between humans and technology has become a focal point for understanding creativity and culture. The creative process today is increasingly entangled with advanced algorithms and automation, raising new questions about authorship, agency, and the role of uncertainty in making art [1]. The impact of technology on human existence is undeniable and growing. While technology evolves, we must also critically examine how it transforms the way we create and perceive art. This paper explores these themes by putting into conversation two seemingly disparate domains: the image manipulation art of Kensuke Koike and the image-processing methods of convolutional neural networks (CNNs). Koike is a contemporary artist known for his meticulous collages made from single photographs, a practice guided by the rule “nothing added, nothing removed”. Meanwhile, CNNs are state-of-the-art computational models that learn to analyze images through layered mathematical operations. At first glance, one might doubt that a human artist’s analogue collage technique has anything in common with an artificial neural network. Yet, a closer look reveals intriguing parallels in how both decompose and recompose images, albeit for very different purposes and in very different ways. By examining these similarities and differences, we can gain insight into how humans and machines process visual information, and what that means for the evolving human-technology relationship in art. Crucially, this comparison is not about equating human creativity with machine computation on a cognitive level, but about drawing formal and methodological connections. We will see that Koike’s labour-intensive collage method and CNNs’ algorithmic image processing both involve forms of segmentation and reassembly of visuals. However, they embed and evoke “doubt” in fundamentally different ways[2]. Koike deliberately cultivates aesthetic doubt in the viewer, as a playful uncertainty about what one is seeing and how it was made. In contrast, CNNs operate through quantifiable uncertainties (randomized weights, probabilistic outputs, error minimization) that are mechanistic, lacking any conscious or aesthetic quality. Exploring these distinctions will help clarify what doubt means in an artistic context versus a computational one. This paper proceeds as follows. First, I introduce Koike’s art practice, explaining his unique technique of decomposing and recomposing photographs under strict constraints. I highlight the formal discipline and conceptual rigor of his method, which exemplifies an embrace of creative constraints and uncertainty as part of the artistic process. Next, I present an overview of convolutional neural networks, describing in general terms how CNNs process images. I then undertake a comparative analysis of Koike’s methods and CNN methods, focusing on how each performs image decomposition/segmentation and recomposition/integration, and noting both similarities and divergences. In the penultimate section, I analyse how each engages with the concept of “aesthetic doubt”. I argue that Koike’s work instils doubt as an aesthetic and conceptual strategy in the human observer, whereas CNNs involve forms of uncertainty that, while important for their function, do not amount to an aesthetic of doubt. I consider whether any notion of “doubt” can be meaningfully applied to CNN operations, or if it is purely a human projection. Finally, the conclusion affirms the philosophical significance of comparing these approaches on formal grounds. This comparison, I suggest, sheds light on the human–machine creative dialectic championed by technological humanism. It does so without conflating human cognition with algorithmic processing, and indeed underscores why the human element, exemplified by Koike’s creative doubt, remains vital. By mapping this unexplored connection, the aim of this paper is to contribute a new perspective on how artistic practices and AI techniques might inform each other in understanding creativity in the post-digital age.
2. Kensuke Koike’s “Nothing Added, Nothing Removed” Art Practice
Kensuke Koike (b. 1980) is a Japanese visual artist based in Italy, recognized for a singular approach to photographic collage[3]. Koike’s practice revolves around found vintage photographs and postcards, anonymous images from bygone eras, often sourced from flea markets or antique shops. What he does with these forgotten photos is both simple in concept and astonishing in effect: using only a sharp blade and his imagination, Koike cuts apart and reconfigures a single image to create a new visual artwork. Crucially, he adheres to a strict rule: “nothing is removed, nothing is added”. Every fragment of the original photograph remains in the final composition, and no outside material is introduced. In other words, the image is deconstructed and then reconstructed entirely from itself. Koike’s commitment to this rule gives his work a distinctive formal purity. It is collage, but not in the usual sense of assembling pieces from disparate sources – rather, it is a solitary collage, a puzzle crafted from one source image. As one commentator describes, «He has one rule for making his work. Nothing can be removed and nothing can be added to the piece he works on»[4]. The process is akin to solving a visual anagram: the “letters” (pieces of the image) can be rearranged to yield a new “word” (composition), but none may be discarded or supplemented. This constraint demands extreme rigor. Each cut must be planned with foresight, since a mistake cannot be corrected by fetching a new piece: «to cut wrong would mean that the next piece will not fit the portion removed»[5]. Koike approaches the task with what has been called a “cool and calculating” mindset, almost mathematical in its precision. The cutting is done slowly, «with great mental architecture and meditation»[6], excising pieces carefully like a surgeon or a geometer. We might say that chance and improvisation are minimized; instead, preparation and exactitude are paramount. The formal constraint becomes a driving force for creative problem-solving, as Koike must find novel ways to utilize every fragment of the image. Despite (or rather, because of) these constraints, the results are remarkably imaginative. Koike’s finished works often take the viewer by surprise, transforming the original image into something surreal, witty, and visually puzzling. For example, in one piece a portrait of a man’s face is sliced into vertical strips that are reassembled out of alignment, so that the eyes and mouth appear multiple times in a staggered, almost cubist arrangement. The familiar human face becomes an uncanny architectural form, yet it is undeniably made of the very face itself. In other works, faces “float” away from the heads they belong to, or a person’s silhouette might be cut and recomposed into the shape of a spider, as reported in an art review. Such images carry a surrealist edge, reinventing found imagery into unexpected, witty scenes. Koike’s «keen ability to re-envision the vernacular and enliven the mundane»[7] breathes new life into these discarded photographs. The viewer is often left both amused and unsettled, confronted with a picture that is at once the original and not the original. The artworks have been aptly termed “renewed photographs,” as they challenge the viewer’s expectations and invite new associations – revealing humour, curiosity, absurdity, and beauty in images once considered ordinary. Formally, Koike’s method requires not only precision but also creative vision. Because he cannot add anything new, the transformation he seeks must lie latent in the original image, to be discovered and activated by cutting. Koike has described himself as a kind of “alchemist” of images, metaphorically turning photographic “lead” into visual “gold.” He often starts by studying a found photo and imagining what hidden possibilities it might contain. As Koike explains, «I first choose an image randomly and then contemplate what to do with [it]. I create prototypes with random images and begin to play around, to have a clearer idea of what kind of interactions are possible»[8]. This experimental phase is crucial: Koike makes many trial runs and prototypes, frequently using photocopies or digital tools to test an idea before ever cutting the vintage print. In fact, Koike notes that people mistakenly think he creates the final collage in one go, but in reality, it takes 20 times to get it right[9]. His typical workflow «starts in Photoshop and ends with endless prototypes trying to get it right»[10], after which he executes the final cuts on the original photo only when he’s confident in the solution. This iterative process underscores the conceptual rigor behind the apparent whimsy of his pieces. Koike is acutely aware of the risk involved — each found photograph is unique and irreplaceable, so any error is irreversible. «More risk means that I have to think twice before cutting the originals, and that is important», Koike remarks[11]. The risk, in other words, enforces a kind of discipline and doubt that ultimately deepens the meaning of his work. By «delv[ing] deeper into the meaning of an image»[12] through such careful manipulation, Koike treats the photograph not just as raw material, but as something to be interrogated and re-imagined. Conceptually, Koike’s rule of “nothing added, nothing removed” can be seen as a statement about the nature of images and reality. On one hand, it shows how a single snapshot of reality can hold endless possibilities for reinterpretation—by simply rearranging its parts, entirely new narratives or forms can emerge. On the other hand, it imposes a strict economy of means: creation comes not from adding content, but from adding insight or changing perspective. This approach resonates with broader artistic themes of constraint as a catalyst for creativity (one is reminded of the Oulipo movement[13] in literature, where writers impose rules on themselves to spur ingenuity). In Koike’s case, the constraint produces works that are playful yet intellectually provocative: the viewer is prompted to consider the photograph’s original truth versus its transformed state, and perhaps to reflect on the fragile line between reality and illusion in images. As one description put it, Koike is a “seeker of the unseen”, using «the tight precision of a knife and ample humour»[14], to reveal hidden facets of familiar photos. The formal discipline (precise cutting, exhaustive planning) combined with conceptual playfulness (surreal reconfigurations, visual puns) gives Koike’s oeuvre a unique place in contemporary art. It invites us to doubt what we see, to question how much of an image’s meaning is inherent and how much is in the arrangement of its parts. This sets the stage for our later discussion of aesthetic doubt. Before that, however, we turn to a very different but equally fascinating system of image transformation: the convolutional neural network.
3. Convolutional Neural Networks: Machines that Decompose and Recompose Images
Convolutional neural networks (CNNs) are a class of artificial neural networks that have revolutionized computer vision in the past decade. In simple terms, a CNN is a computational model designed to analyse visual data by mimicking aspects of how animal visual cortices process images. While Koike manually cuts and rearranges pixels on paper, a CNN “cuts” and analyses images through mathematical filters. In essence, CNNs perform a layered decomposition of images into features, and then recombine those features to make predictions (such as recognizing an object in the image). CNNs rose to prominence after 2012 when they achieved breakthrough results in image recognition competitions, and they have since become dominant in various computer vision tasks, from classification and detection to segmentation and image generation[15]. A convolutional neural network is composed of multiple layers of artificial neurons arranged in a sequence. Each layer transforms its input data to produce an output that becomes the input to the next layer[16]. The key types of layers in a CNN include convolutional layers, pooling layers, and fully-connected (dense) layers. The convolutional and pooling layers serve as feature extractors, while the fully-connected layers (usually at the end) serve as a classifier or decision-maker. Through training on many examples, the CNN learns appropriate weights for all these layers, enabling it to automatically recognize patterns in images. Convolutional layers are the core building blocks. In a convolutional layer, a set of small filters (also called kernels) is applied to the input image (or to the feature map from the previous layer). Each filter is essentially a tiny matrix of weights that slides (“convolves”) across the image, computing a dot product at each position. This operation produces a new image-like map called a feature map or activation map. Thus, the original image is decomposed into a set of feature maps, each highlighting a different aspect of the image’s content. This is analogous to looking at the image through many different “lenses,” each isolating a particular feature (like edges, corners, or blobs). By applying multiple filters, a convolutional layer can detect multiple features in parallel. Pooling layers typically follow one or a few convolutional layers. The purpose of pooling is to downsample the feature maps, reducing data size while preserving the important information. Pooling thus contributes to the CNN’s robustness and helps control overfitting by reducing the number of parameters. After several convolutional and pooling layers, a CNN often ends with one or more fully connected layers (the same type as in a standard neural network). These take the high-level feature maps produced by the final convolutional layer, flatten them into a vector, and process them to produce an output. The fully connected layers perform the integration and decision-making based on the features extracted by earlier layers. In an image classification task, the final output might be a probability distribution over possible labels, effectively recomposing all the detected features into a semantic judgment (“This image is a cat with 95% confidence, not a dog”).
One of the powerful aspects of CNNs is that they learn the right features automatically from data. Through a training process (usually using backpropagation and gradient descent), the CNN adjusts the weights of its filters and connections so as to improve its performance on a given task. The process is data-driven and adaptive, which is a major reason for CNNs’ success; given enough training images, a CNN can discover very intricate and optimized ways to parse images far beyond any manually designed feature extraction. To illustrate, consider a photograph of a human face. A well-trained CNN’s early convolutional filters might produce feature maps highlighting all the edges in the image (outlining the chin, the hairline, the eyes, etc.). Deeper in the network, some neurons might specifically fire when they see an eye-like shape or a round object (perhaps detecting the iris or the whole eye region), and others might respond to mouth shapes or ears. Deeper still, a neuron might combine those inputs to detect the presence of a face as a whole. By the final layer, the network might output a high score for the class “face” or identify the person if it’s a recognition task. All of this happens through learned convolutions and weighted connections that incrementally segment and recompose the image’s information.
In summary, CNNs provide a mechanized counterpart to what Koike does by hand. They take an image, analyse it by breaking it down (through convolutions into feature maps), often reduce it to constituent informative parts (through pooling and multiple stages of feature extraction), and then recompose an understanding or output (through deeper layers and possibly inverse operations in generative/segmentation models). A CNN doesn’t create a whimsical collage, of course; it produces classifications, detections, or segmentations according to how it was trained. But at a formal level, both Koike’s collages and CNNs involve a transformative reassembly of visual information. Next, we explore these two processes side by side, examining how image decomposition and recomposition play out in each case, and how the notion of “doubt” is or isn’t present in their operation.
4. Comparing Koike’s Method and CNN Image Processing
Placing Kensuke Koike’s artistic technique alongside convolutional neural networks yields an intriguing study in contrasts and comparisons. Both involve cutting an image into parts and putting it back together, yet one is an intuitive human-guided process aimed at aesthetic expression, and the other is an automated mathematical process aimed at visual analysis or recognition. In this section, we explore the parallels in how each approach handles images (formally breaking and combining components), and we highlight key differences in motivation, execution, and outcome. By examining these similarities and differences, we can better understand the unique nature of human creativity versus machine computation, while appreciating the structural analogies between the two. At a structural level, Koike’s collages and CNNs both perform a kind of image segmentation followed by reconstruction (or integration). Consider the following step-by-step analogy.
Koike physically segments a photograph by cutting it into pieces. Each piece is a contiguous region of the original image (for example, a person’s eyes, or strips of the image, or a shape cut out along certain lines). Similarly, a CNN analytically segments an image by applying convolutional filters that respond to local patterns. While the CNN doesn’t literally produce separate image scraps, it produces multiple feature maps each of which isolates certain content (edges, textures, object parts) from the whole. This can be seen as a form of conceptual cutting of the image, splitting the information into parallel channels, each representing one aspect of the image’s structure. In a deep CNN, later layers effectively segment the image into more abstracted components (for instance, one internal neuron might strongly activate for “eye-like” regions across the image, effectively isolating eye regions in its activation map). Furthermore, in explicit segmentation networks (like U-Net or other semantic segmentation CNNs), the network indeed divides the image into segments by labelling each pixel, which is a direct computational analogue of cutting an image into regions corresponding to objects or surfaces.
After cutting, Koike rearranges the pieces of the photo in a new configuration. The spatial relationships between elements are altered: e.g., he might swap two pieces, rotate one, or offset them from their original positions. This creates a novel composition while still using the original visual content. In CNNs, after the initial decomposition into feature maps, the features are transformed and combined through subsequent layers. For example, an early edge-detecting feature might be combined with another feature in a next layer to detect a corner (which is like two edges meeting). There is a recombination happening as we go deeper, features get aggregated (via weighted sums in neurons) to form more complex features. By the final layers, the network has effectively reassembled a high-level understanding from the low-level pieces. If one considers a CNN that generates an output image, then there is a literal recomposition of pixel values at the end (often using decoding layers that are roughly the inverse of convolution). Thus, both systems end up putting pieces together to make something new out of the original image’s data.
Interestingly, both processes use only information from the original image. Koike’s rule ensures no outside imagery is introduced; CNNs similarly do not pull in external data when processing a single image (the learned weights are from prior training, but the actual processing of one image is self-contained on that image’s content). In other words, all the features and outputs the CNN produces are derived from the input image (with the assistance of learned filter weights). This is analogous to Koike deriving all visual elements from the photo itself. In both cases, the “DNA” of the final result traces entirely back to the input image.
Koike’s operation can sometimes be a multi-stage cut-and-paste. For instance, he might cut a photo into large sections, then further cut one section into sub-pieces and rearrange those, achieving a hierarchy of recomposition. Likewise, CNNs operate in a multi-layer hierarchy, where each layer works on the output of the previous, effectively performing a series of transformations. The end result is not the product of one single cut or one single filtering, but a compounded effect of many small operations. This multilevel processing in both cases contributes to the richness of the outcome.
To make this more concrete, imagine a specific visual element like a human face in a photograph. Koike might cut out the eyes and relocate them elsewhere in the picture (perhaps swapping left and right, or moving them to the forehead). A CNN, on the other hand, will “cut out” the eyes in a figurative sense: some of its filters will focus on detecting eyes independent of where they are, and later layers might treat “eye detected here” as a unit of information to combine with “nose detected there” when recognizing a face. In effect, the CNN has isolated the eyes (like a cut-out) and then virtually re-positioned that information in relation to other features to decide that it’s looking at a face. Both processes must preserve certain relationships (Koike must ensure pieces fit without gaps or overlaps that betray the original material; the CNN must maintain correct spatial correspondence when combining features, otherwise it could mistake the arrangement). The difference is that Koike intentionally breaks logical relationships (placing eyes where they shouldn’t be to create an artistic effect), whereas the CNN is generally trying to infer the true relationships (like, eyes and nose in proper arrangement equal a face). To use a metaphor, Koike’s collage and a CNN’s analysis are like two forms of visual grammar. Koike takes a “sentence” (image) and rewords it by shuffling its “words” (components) into a new poetic order, making us see new meanings. A CNN diagrammatically parses the “sentence” into grammar (parts of speech like edges, shapes) and then uses the grammar to understand the sentence’s meaning (what is in the image). Both require understanding the constituents of an image and of how their recombination can yield either a new image or a new interpretation.
5. Aesthetic Doubt vs. Mechanistic Uncertainty
One of the central themes introduced was the concept of doubt and how it plays into creativity and technological processes. In artistic contexts, doubt can be an engine of originality and a space of reflection. It is often said that artists question and doubt appearances to reveal deeper truths or new perspectives. In computational contexts like CNNs, what might correspond to “doubt” are things like uncertainty in predictions or the trial-and-error of an optimization algorithm. However, these are not conscious doubts but algorithmic mechanisms. This section analyses how Kensuke Koike’s work actively engages aesthetic doubt, and contrasts that with the entirely different nature of “uncertainty” in CNNs. We also consider whether CNNs contain any embedded form of doubt or if doubt remains an exclusively human (and humanistic) domain.
5.1 Koike’s Cultivation of Doubt in the Viewer (and Himself)
Koike’s collages are deliberately crafted to make viewers do a double-take, to doubt their first impression of the image. At first glance, one of his pieces might look like a normal vintage photograph; a moment later, the eye catches that something is impossibly off. For instance, seeing a man’s face with duplicate eyes in strange places, the viewer might momentarily doubt the evidence of their senses. This moment of uncertainty is exactly what Koike often aims for. By presenting an image that defies expectations (yet is made solely of expected components), Koike creates an effect of aesthetic doubt, the comfortable trust we place in photographic reality is shaken. The viewer begins to question what they see, and in doing so, becomes acutely aware of their own act of seeing and assuming. This strategy relates to a broader concept in aesthetic theory: art can introduce ambiguity and uncertainty to invite active interpretation by the audience. Umberto Eco famously described the “open work” as a work of art that is intrinsically ambiguous, requiring the audience to complete it in their own minds. Koike’s collages exhibit a kind of controlled ambiguity. There is a definite transformation, yet the meaning of that transformation isn’t handed to the viewer on a plate. The viewer must reconcile two truths: everything in the image is original, yet the overall image is altered. This generates a productive tension and doubt: we doubt the completeness or integrity of the image even as we see it’s all there. In a sense, Koike is playing on what we might call the ontological doubt of images, the realization that an image can lie or tell new truths without altering its fundamental pieces. From a philosophical perspective, doubt has often been seen as a driver of deep engagement. This he considered essential for artistic creativity, the artist does not rush to resolve doubt but lives in it to find inspiration. By imposing the constraint of no additions or removals, he forces himself into a corner of uncertainty: How can I make something new out of this one photo? He must dwell in doubt, exploring various cuts and arrangements (with many discarded prototypes) until a satisfying solution emerges. That is the creative doubt internally. Externally, he passes a version of that doubt to us: we, too, must puzzle and think twice about the image to grasp it. In this way, doubt becomes an aesthetic experience, a kind of play between artist and viewer, where the viewer’s slight confusion or hesitation is actually a source of engagement, wonder, and perhaps insight. Koike’s use of doubt is fundamentally aesthetic: it enhances the expressive impact of the work. The doubt is not about factual truth (we know the image is manipulated), but about perceptual and interpretive truth. It raises questions rather than answers them: Why does the image feel so uncanny? What does it suggest about identity when eyes are relocated? The viewer might even doubt the boundary between photography (usually seen as truth-bearing) and collage (which is evidently constructed), as Koike’s work blurs that line. In sum, doubt in Koike’s context is an invitation to engage; it slows down perception and forces a reflective, curious attitude.
5.2 Doubt (or Lack Thereof) in CNN Operations
If we turn to convolutional neural networks, can we find anything analogous to “doubt”? At first glance, a CNN is an algorithm and has no consciousness or emotion; it cannot feel doubt as a human does. However, there are a few places where one might metaphorically or mechanistically talk about uncertainty in a CNN.
5.3 Training Phase Uncertainty
During training, the CNN starts with random weights and gradually adjusts them to reduce a loss function (which measures error). One could say the network initially is “very unsure” of how to represent the data, essentially it has high loss (lots of doubt) and through learning it reduces that uncertainty, converging to a solution that explains the training images well (low loss, more confidence). This is, of course, a mechanical process. The CNN isn’t aware of being unsure; it’s simply following a gradient. But mathematically, one can view the progress as resolving uncertainty about the correct internal model. In Bayesian interpretations of neural networks, this can even be made explicit: one can treat the weights as having a probability distribution and update beliefs about them. In that sense, during training the network’s weights encode a distribution of possible explanations. However, standard CNN training doesn’t explicitly maintain a distribution (unless using Bayesian methods); it just finds one set of weights that (hopefully) works. There is no active doubt or metacognitive check, the algorithm either converges or not. If it doesn’t, we see it as failure; if it does, it “believes” the model is set (even if it’s a wrong solution, the algorithm doesn’t doubt it unless we validate and force adjustments).
5.4 Model Ambiguity and Errors
CNNs can be very confident and yet wrong, especially when confronted with tricky or adversarial inputs. For example, an adversarial image might look like static to us but the CNN “sees” a clear object with high confidence. In such cases, one cannot even say the CNN is doubtful, it is confidently fooled. The concept of doubt doesn’t straightforwardly apply because the network doesn’t have a mechanism to realize its own error or second-guess itself beyond what is encoded in the output probabilities.
In essence, CNNs lack an analogue of aesthetic doubt. They do have uncertainties in a mathematical sense, but these are incidental to their function, not intentionally cultivated. A CNN does not value uncertainty; rather, the goal in training is often to reduce uncertainty (improve confidence and accuracy). In creative work like Koike’s, doubt is seen as valuable, it keeps questions open and fosters exploration. In CNNs, uncertainty is something to either quantify or eliminate for better performance. The “forms of automatism” referenced in the call for papers are exactly these kinds of automated processes that seem to proceed without the deliberation or hesitation that characterize human creativity. CNNs epitomize automatism in image creation/analysis: they execute learned rules without pause to reflect or question. They do not have an internal dialectic of “Is this the right interpretation or should I try something else?” beyond what gradient descent enforces during training. Interestingly, when we integrate CNNs into human workflows (like using AI to generate art or assist in creative tasks), the doubt re-enters through the human. A human might doubt the appropriateness of a CNN-generated outcome and then adjust something. But the CNN itself is not a participant in doubt; it’s more an object that might cause doubt in us. For example, if an AI art generator produces a weird collage, the human artist might doubt whether the composition works and then select or guide differently. The CNN doesn’t share that doubt, it just outputs according to its weights. One could poetically say that CNNs have a form of blind faith in their learned parameters. They apply them rigidly, and if the input distribution shifts (leading to mistakes), they do not notice unless re-trained. There is no built-in scepticism. Koike’s work shows an artist consciously leveraging doubt for artistic ends; CNNs show an automated system where any doubt must be externally interpreted or managed. If we stretch the idea, one might ask: Could we imagine a CNN-like system that had a loop of doubting its output and revising it? That starts bordering on more advanced AI (perhaps systems that do iterative self-refinement or generative models that produce variations and choose one). Some generative adversarial networks (GANs) have a generator and a discriminator; the discriminator could be seen as “doubting” the authenticity of generator outputs, forcing the generator to improve. Even there, it’s an adversarial game, not introspective doubt. True aesthetic doubt would require the system to have a notion of what it wants to achieve aesthetically and to judge its own intermediate outputs against that, this is far beyond current CNNs. In conclusion, any ‘doubt’ in CNNs is at best a metaphor for uncertainty quantification, whereas doubt in Koike’s context is an experiential and creative principle. The dialectical relationship between human doubt and machine automatism, hinted at in the call for papers, becomes clearer: human creators like Koike embrace doubt to push creativity, whereas machines like CNNs operate with deterministic or probabilistic rules that do not include doubt unless engineered in. This dichotomy perhaps suggests why, despite the power of automation, there remains something irreducibly human in the creative process.
6. Conclusion
Bringing together the worlds of Kensuke Koike’s image collages and convolutional neural networks has allowed us to examine, through a new lens, the evolving relationship between human creativity and technological processes. Formally and methodologically, there are resonances: both Koike’s art and CNNs deal with breaking images down and building something up from those pieces. The exploration kept the analysis to formal and methodological grounds, deliberately avoiding any suggestion that CNNs are artistically creative in the human sense. The dialectic between doubt and automatism emerged as a key theme: Koike’s work affirms that the artist’s doubt, the willingness to question an image, to reconsider form, to embrace uncertainty, is a generative force. The CNN affirms that automation can achieve remarkable outcomes by brute pattern-learning, yet it operates with a kind of unquestioning determinism that lacks the exploratory nature of human creativity. In answering the question posed at the outset: Yes, the idea of doubt is still necessary as a constitutive element of creation, at least, for human creators striving for meaning and originality. Koike’s art exemplifies this necessity: without his double-thinking, his risk-taking and careful doubt, the magical recompositions he achieves would not be possible. At the same time, forms of automatism (like CNNs) have their place as tools and methods; they can produce, but they do so in a different, often less richly questioning, manner. The relationship between the two is not zero-sum but complementary. We can imagine, for instance, Koike using a neural network as part of his process (perhaps to generate prototype ideas), or conversely, artists interpreting CNN visualizations to inspire artworks. In such collaborations, the human brings doubt and critical evaluation, the machine brings generative or analytical power, and together they could yield new creative frontiers. This speaks to the broader Technological Humanism goal of reaffirming our humanity in the face of powerful new technologies. The lesson is that technology can extend our capabilities, but the human element, our ability to doubt, to question, to imbue artifacts with meaning, remains irreplaceable and indeed crucial. In the spirit of continuing this inquiry, further research or creative experiments could explicitly combine these realms: perhaps training a CNN on images of Koike’s collages to see if the machine can learn the visual “trick” and produce its own “nothing added, nothing removed” transformations, or conversely, using Koike’s approach as a metaphor to design new interpretable network architectures (e.g., networks that constrain themselves to use all parts of an input in generating an output). Such bridges might deepen our appreciation of both human and machine creativity. In conclusion, comparing Kensuke Koike’s collage methods with CNN methods has provided a thought-provoking case study in how formal analogies can illuminate conceptual differences. It reaffirms that doubt, that quintessentially human spark of curiosity and caution, remains at the core of creative endeavour.
Giovanni Galli
(Giovanni Galli is a Postdoctoral Researcher at University of Teramo. He earned his PhD at University of Urbino and a Master of Arts at Bologna University. His research focuses on issues related to the philosophy of science and the philosophy of artificial intelligence. Specifically his academic interests rotate around the study of explanatory theories and of scientific understanding. Furthermore he is interested in scientific communication, as a local administrator he focuses on the role that science plays in political decision-making processes.)
References
S. Colton and G. A. Wiggins, Computational creativity: the final frontier?, in Proceedings of the 20th European Conference on Artificial Intelligence (ECAI’12), IOS Press, NLD, 2012, pp. 21-26.
D. Cope, Experiments in Musical Intelligence, in Computer Music and Digital Audio, Series 12, WI: A-R Editions, Madison 1996.
D. Cope, Computer Models of Musical Creativity, MIT Press, Cambridge (Massachusetts) 2006.
M. Faini, Standing at the Crossroads. Stories of Doubt in Renaissance Italy, Legenda, Oxford 2023.
B. Feuerhelm, Kensuke Koike & Thomas Sauvin: No Cut Left to Chance, «ASX», 14 December 2017. https://americansuburbx.com/2017/12/thomas-sauvin-kensuke-koike-no-cut-left-chance.html#:~:text=He%20has%20one%20rule%20for,portion%20removed%20and%20vice%20versa (cited on 1 Sept 2025).
R. Flanagan, Kensuke Koike: Nothing Added, Nothing Removed, «Ignant», 7 March 2018. https://www.ignant.com/2018/07/03/kensuike-koike-nothing-added-nothing-removed/ (cited on 1 Sept 2025).
K. Koike, Photographers on Photographers: Su Ji Lee in Conversation with Kensuke Koike, interview with Su Ji Lee, «Lenscratch», 17 August 2022. https://lenscratch.com/2022/08/photographers-on-photographers-su-ji-lee-in-conversation-with-kensuke-koike/#:~:text=Does%20staring%20at%20strangers%20and,what%20to%20do%20with%20them (cited on 1 Sept 2025).
K. Koike, https://www.kensukekoike.com/, 2025. (Cited on 1 Sept 2025).
M. Mitchell, Artificial Intelligence. A Guide for Thinking Humans, Farrar, Straus and Giroux, New York, 2019.
E. S. Paul and D. Stokes, Computer Creativity is a Matter of Agency, in Institute of Arts and Ideas News, 11 November 2021.
Rosegallery, Kensuke Koike, special online exhibition, 2 October – 4 November 2020. https://rosegallery.net/viewing-room/3-kensuke-koike/ (cited on 1 Sept 2025).
X. Zhao, L. Wang, Y: Zhang et al., A review of convolutional neural networks in computer vision, in «Artificial Intelligence Review», LVII, 99, 2024, pp. 1-43. https://doi.org/10.1007/s10462-024-10721-6 (cited on 1 Sept 2025).