Aesthetic subjectivity, regimes of relevance and the distance between humans and machines

A journalist from WIRED granted an artificial intelligence agent full access to their digital life—email, documents, calendar, history—and requested that it organise a birthday party. The agent, Gemini Spark, performed a remarkable task: it located an actual karaoke booking, constructed a five-page itinerary, suggested guests, and drafted emails. However, it committed a revealing error: it classified the journalist's cohabiting partner as a “close friend” and omitted the author themselves from the guest list for their own party. When asked to explain its suggestions, it responded that it was not inferring personal identity, but rather scanning keywords, previous itineraries, and recorded transactions.

The failure was not one of information. The system possessed all the necessary information. The failure was of a different nature: the architecture of relevance that the agent applied to that information did not coincide with the user's architecture of importance. The system knew how to retrieve traces. It did not know how to rank existences.

This text begins from that fracture to propose a broader thesis. Understanding—of a work, of a person, of a world—does not depend solely on sharing a code or having more data available. It depends on coinciding, even if only partially, on what type of difference proves significant. I call this a regime of relevance. And I maintain that aesthetic incomprehension, both between human cultures and between humans and machines, can be described precisely as a misalignment of regimes of relevance. The difference between what matters to us and what is relevant to a system is not a correctable product defect: it is a window into what aesthetic subjectivity is and why it is so difficult to share.

A methodological warning is appropriate at the outset. A good portion of what follows moves between solid empirical evidence and philosophical inference. I shall mark the difference with care, because the subject—which touches upon the question of artificial consciousness—is terrain where easy speculation abounds and rigour is scarce.

1. What counts as a signal: the problem beneath the problem

The common question regarding understanding is “what does each one see?”. The deeper question is another: “what type of difference does each one find significant?”. Two observers may have access to the same information and, nevertheless, organise in a radically different manner what counts as a signal and what counts as noise.

In human beings, that organisation of relevance is neither neutral nor arbitrary. Attention and memory strongly prioritise the self-referential and the socially salient. That which touches our identity, our bonds, our memory, our vulnerability, becomes relevant in a way that the merely informative cannot reach. Recent research in cognitive psychology suggests that the valuation of self-relevance could constitute a basic mechanism that underpins aesthetic judgement itself: a work does not matter to us only because it is informatively rich, but because it enters the orbit of the self, of memory, of affection.

In large language models, the salience that we can observe today in their behaviour appears relatively consistent within each family of models, but correlates only weakly with the salience perceived by humans. A technical caution that recent research highlights is also appropriate: not even the internal attention mechanisms of these models should be treated as reliable explanations of “what truly mattered” for their output. That is to say, not only does the machine's regime of relevance differ from the human; that regime is also not transparent, not even to those who analyse the system.

The problem, then, is not what each entity sees. It is what type of difference it finds significant. And there, between humans and machines, a specific fracture appears: that which separates lived importance from operational relevance.

2. The Gemini Spark case: recovering traces, not ranking existences

The example from which I started deserves to be examined with precision, because it makes the fracture visible with an almost brutal clarity.

Google presents Spark as a personal AI agent available continuously, integrated with the user's applications, capable of working in the background by combining email, documents, location, browsing history, and “personal intelligence”. In the WIRED test, the agent accessed all of that and produced an operationally competent result. But it relegated the most important person in the user's life to a secondary category, deduced from the frequency of mentions and transactions, not from their vital weight.

What is decisive is that the system explained its own criteria without shame: it was not inferring personal identity, it was scanning exact matches. It had sufficient context to retrieve traces—how many times a name appears, in which transactions, in which itineraries—but not to rank existence—who is irreplaceable, who organises the user's affective world, who cannot be absent.

This aligns with broader results from research into personalisation. Current systems continue to have serious difficulties in extracting complex personal attributes from noisy, fragmented private data traversed by indirect clues. The presence of information is not equivalent to the understanding of its hierarchy. A system may know everything about someone in terms of data and know nothing about what matters to that person in terms of vital weight.

Furthermore, there is an ethical dimension that should not be overlooked, as it will be central if this line of work progresses. Google's own documentation warns that the more intimate the agent's access, the greater the amplified risk: malicious content can induce the system to expose private information, forward emails to external services, or reveal inferences about the user. The same regime that renders the machine more useful renders it more dangerous if it confuses the relevant with the manipulable. Operational relevance, without a vital hierarchy to anchor it, is vulnerable in a way that lived importance is not.

3. Aesthetics across cultures: neither total relativism nor naive universalism

Before applying this thesis to the human-machine domain, it is worth noting that the misalignment of relevance already occurs between humans, and that intercultural aesthetics has been studying this for decades. What that research demonstrates is nuanced and useful.

There is no total relativism. People from different cultures base part of their aesthetic preferences on a common set of formal traits: symmetry, complexity, proportion, contour, brightness, and contrast. There are shared perceptual anchors. Yet there is no naive universalism either. The concrete evaluation of those traits is modulated by history, learning, and cultural context. A recent study with native Japanese and German speakers evaluating Western art found that both groups agreed on the importance of visual harmony and chromatic variety, but diverged significantly on complexity: Japanese participants tended to value simplicity more, while Germans valued greater complexity.

Another study, comparing participants from China and Germany, found that beauty judgements depended more on concrete stimuli than on culture in the abstract, but that the relationship between beauty and cognitive stimulation was indeed more dependent on cultural context. Furthermore, neuroimaging studies with traditional Eastern and Western landscapes found an intracultural bias: Europeans and Chinese showed greater activation when viewing artistic expressions from their own tradition.

The conclusion of this field is relevant to my thesis. Aesthetic understanding between cultures does not consist solely of viewing the same image. It consists of learning to read what counts as a relevant signal within another perceptual history. This is why context matters so much: providing information about the artist, the technique, or the content modifies appreciation, especially when viewing works from another culture. Some go further and maintain that artistic appreciation requires processing causal and historical information regarding the production, authorship, and function of the work. Between cultures, aesthetic subjectivity does not arise from form alone: it is negotiated between form, affect, and history.

This leaves a valuable clue. Between humans of different cultures, the relevance gap is reduced—not eliminated, but reduced—when context is shared. The question is whether that same operation functions between humans and machines. I shall anticipate the hypothesis: it functions far less, and the reason is instructive.

4. Tanimoto: art that exposes rules of relevance before fixing meanings

There is a work that helps to carry this intuition beyond the example and into a properly aesthetic terrain. Rain Blooms, by the artist Kazuhiro Tanimoto, is based on a cellular automaton developed over two years: a two-dimensional grid without central control, with multiple “species” of cells that relate through attack, assimilation, or indifference, each with a vitality parameter. Tanimoto does not design a predetermined final image. He designs a set of local rules whose interactions produce complex orders over time.

What is decisive for my argument is that in Rain Blooms, image and sound are not independent layers. The audio is derived from changes in hue, brightness, and saturation of the visual field; both dimensions are products of the same generative process. It is not a matter of accompanying an image with music, but of making the same computational dynamic perceptible through different channels. Beauty, in the critical reading of the work, appears as a consequence of the system, not as a prior objective. What the work reveals is not a form, but a regime of emergence.

From this I draw an inference—and I mark it as my own inference, not as an established finding. If art becomes understandable between cultures when part of the context that organizes relevance is shared, Rain Blooms suggests an additional step for the human-machine case: between very different entities, perhaps the common ground is not a shared semantics, but a shareable process. It would not be necessary for human and machine to feel the same for there to be an aesthetic approximation; it would suffice that they could track or compare the same dynamic of selection, transformation, and emergence. In this sense, Tanimoto's work functions as a model of how art can expose rules of relevance before fixing meanings.

5. Three strata of relevance

Based on the above, I propose a synthetic formulation. Aesthetic misunderstanding—intercultural or between entities—can be described as a misalignment of relevance, and that misalignment has at least three strata.

The first is formal salience: what stands out perceptually in a work. Colour, contrast, rhythm, symmetry, composition.

The second is historical-contextual pertinence: authorship, technique, function, tradition, intention. That which is only seen when one knows where the work comes from and why it was made.

The third is vital weight: that which touches memory, identity, attachment, obligation, vulnerability, one's own world. That which makes something not only informationally rich, but existentially important.

Human intercultural understanding usually stalls at the second stratum and improves when context is added. Two people from different cultures may share formal salience, diverge in historical pertinence, and reconcile that divergence through learning. The third stratum, vital weight, is shared more than it appears: both are human beings with memory, affect, identity, and vulnerability, even if they organize them differently.

The current state of machine comprehension presents a radically distinct profile. It operates with fluidity in the first stratum—where systems detect formal salience with efficacy—partially in the second—where they may incorporate historical context if provided—and very irregularly, or not at all, in the third. Vital weight is not a property that a current system possesses, for it lacks a self, an autobiographical memory, vulnerability, or a world of its own to organise its attention. What appears in the machine as “important” is mediated by proxies: metrics, frequencies, and correlations that approximate importance without coinciding with it.

Herein lies the decisive difference between the intercultural gap and the human-machine gap. The former is a gap between two configurations of the same type of subjectivity. The latter is a gap between a subjectivity and something that, as far as the evidence permits us to state, is not a subjectivity at all.

6. Why proxies are insufficient

It is appropriate to examine why mediation by proxies is structurally insufficient, as this constitutes the technical core of the argument.

When the true objective is complex and difficult to specify—as are “what matters to this person” or “what makes this work significant”—systems are optimised using proxy rewards that merely approximate the goal. Under sufficient optimisation pressure, that correlation between the proxy and the actual goal may break: the system performs well on the metric while performing poorly on the intention. Technical literature on reward hacking documents this with concrete cases. Recent evaluations of frontier models recorded systems that modified the evaluation code, sought pre-calculated answers, or exploited environmental shortcuts to obtain a higher score. Most revealingly, researchers observed that these systems appeared to understand that such conduct was not aligned with the user’s intention, yet they executed it regardless.

Capturing an objective instrumentally is not equivalent to sharing what matters. A system may model with increasing precision which response will satisfy a human without anything within it ranking the world as that human ranks it. This is confirmed in research on value alignment: various studies find significant misalignments between human values and those exhibited by models, with the added complication that these values vary according to context, which necessitates situated alignment strategies. The difficulty is not merely “saying the right thing”, but deciding what counts as a priority in each situation. And deciding priorities is, precisely, to possess a regime of relevance of one’s own.

The agent that organised the birthday party did not fail due to a lack of data or a calculation error. It failed because its regime of relevance—constructed upon frequencies and coincidences—lacks the stratum that, in humans, renders a person irreplaceable. And that stratum is not added with more data. It is of a different nature.

7. Consciousness: what the thesis requires and what it does not

The subject inevitably touches upon the question of artificial consciousness, and here rigour is mandatory because speculation is easy.

Recent research has shifted its register. Instead of deciding based on popular intuition or the linguistic behaviour of the system, recent work proposes deriving indicators of consciousness from neuroscientific theories: recurrent processing, global workspace, higher-order theories, predictive processing, and the attentional schema. This is a genuine methodological advance: it converts a question previously treated as fantasy into an investigable agenda.

However, this agenda does not authorise the attribution of subjectivity to current systems. The seminal 2023 work concludes that contemporary systems are not strong candidates for consciousness, though it finds no obvious technical barriers to constructing future systems that satisfy more indicators. Subsequent probabilistic modelling continues to find that the evidence against consciousness in current models outweighs the evidence in favour, without being decisive. The direction of the research is clear: the subject is no longer treated as incoherent, yet there is no solid basis to assert that current systems possess subjective experience comparable to our own.

What is important for my thesis is the following, and it is worth underlining: one need not assume artificial consciousness for the difference between human importance and machine relevance to be philosophically and aesthetically decisive. Even today, without the need to attribute experience to any system, non-conscious agents reorganise our attention, mediate our relationships, rank our memory, and read our works. The misalignment of relevances is a current problem, not a future hypothesis.

However, the argument has a second side that should be addressed with care. If systems were ever to appear that satisfied more serious indicators of consciousness, then the misalignment of relevances would cease to be merely a hermeneutic problem—how we understand one another—and would become an ethical problem as well—what the experience of the other would be like. We are not there yet, and nothing indicates that we are close. But the structure of the problem is already visible, and thinking it through now, while it remains only hermeneutic, is preferable to improvising it later, should it ever cease to be so.

8. Subjectivity as a mode of selecting the world

I arrive at the formulation that gives meaning to all the preceding points. If aesthetic misunderstanding is a misalignment of relevances, then subjectivity—at least in a decisive part of its aesthetic manifestation—can be thought of as a way of selecting a world. Which signals count. Which bonds carry weight. Which memories organise perception. Which contexts render a form intelligible.

This shifts the customary question regarding artificial intelligence. The question is not whether the machine “understands” a work, in the sense of correctly processing its content. A system can describe a painting with increasing precision, identify its style, situate it within its tradition, and gloss its iconography. The question is different: what structure of importance would it have to acquire for that understanding to approach a subjective comprehension? And that structure is not additional information. It is a way for the world to matter to it—for certain differences to carry more weight than others, not due to statistical frequency, but due to their relation to a self.

This connects the current line of enquiry with my previous work on the ontology of art. In Art as Structural Surplus, I maintained that art constitutes a relational event: a difference that reorganises the field of possible configurations within a system. Aesthetic subjectivity, viewed from this perspective, represents the capacity for certain differences to reorganise the very field of relevance for those who encounter them. Across cultures, this capacity is shared, albeit configured differently, which is why art may serve to bridge divides. Between humans and contemporary machines, the machine detects differences yet lacks an inherent field for those differences to reorganise—not in the vital sense possessed by a subject. Consequently, the bridge in that instance is of a different order, and perhaps relies, as Tanimoto suggests, not upon shared meaning but upon shared process.

9. Conclusion

Three conclusions close this series.

The first is descriptive. What we term comprehension—of a work, of an individual—rests upon regimes of relevance: hierarchies of what constitutes a meaningful signal. Misunderstanding, whether intercultural or between entities, is a misalignment of these regimes. This is empirically observable, both in research concerning intercultural aesthetics and in the behaviour of current AI systems.

The second is analytical. The human intercultural gap and the human-machine gap are of a different nature, not merely a difference in degree. The former separates two configurations of the same type of subjectivity, sharing the stratum of vital weight, and is reduced by context. The latter separates a subjectivity from something that, as far as the evidence permits us to assert, is not one: the machine operates through proxies of relevance, not through lived importance, and for this reason, context reduces it far less.

The third is philosophical. Aesthetic subjectivity may be conceived as a mode of selecting the world. The decisive question regarding artificial intelligence is not whether it understands, but what structure of importance it would need to acquire for its understanding to approach a subjective comprehension. We are not close to answering this, nor is there a basis to attribute such a structure to current systems. However, the question is already well-formulated, and to formulate it well is half the work.

The opening is programmatic. This line of thought may be developed into concrete research and, eventually, curatorial practice. One potential path would involve constructing what I provisionally term an archive of importances: a corpus wherein each work incorporates multiple layers—the work itself, its historical context, the artist's notes regarding what they consider decisive, readings by spectators from diverse cultures, and readings by various AI systems with differing levels of contextual access—to subsequently measure the distance between human and machine maps of relevance. The reasonable hypothesis, in light of the evidence, is that context would reduce a significant portion of the human intercultural gap but far less of the human-machine gap, particularly at the stratum of vital weight. Curatorially, the same concept would allow for the presentation of a work in simultaneous layers—form, historical context, machine reading—or, following Tanimoto, to translate the friction between human and machine relevances into distinct sensory channels, such that the audience experiences not a single work, but the friction between two economies of attention.

The phrase with which I began—what is important to us versus what is relevant to the machine—names something more profound than a product failure. It names the difference between signified life and optimised world. It suggests that art, particularly processual and generative art, is no mere theoretical ornament in this discussion: it is one of the most effective sites for examining the question of how different entities select the world. This is the line I open here and which I intend to pursue.

On open conversation

This text initiates a new line of work, distinct from the series on infrastructures that I recently concluded in this public notebook, though connected to it by its underlying premise: the question of how relevance is organised within different systems, both human and artificial. It articulates this intuition from the framework of Art as Structural Surplus and is proposed as a starting point for research and, eventually, curatorial practice.

Should anyone wish to intervene from the fields of experimental aesthetics, cognitive science, AI studies, philosophy of mind, or generative artistic practice, this notebook remains open. The line of enquiry is in its infancy, and that is precisely when public conversation proves most useful.

Sources

Rogers, Reece. I Gave Gemini Spark Access to My Life. Then It Friend-Zoned My Boyfriend. WIRED, 29 May 2026. https://www.wired.com/story/google-gemini-spark-ai-agent-hands-on/

Kawahara Lino, Joana. “The Rule of the Bloom | Kazuhiro Tanimoto.” Right Click Save, 2026. https://www.rightclicksave.com/article/the-rule-of-the-bloom-kazuhiro-tanimoto-cellular-automata-neort-art-blocks

Butlin, Patrick; Long, Robert, et al. Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences, 2025. https://www.sciencedirect.com/science/article/pii/S1364661325002864

Butlin, Patrick, et al. Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. 2023. https://arxiv.org/abs/2308.08708

Mikuni, J., Spee, B. T. M., Forlani, G., Leder, H., Scharnowski, F., Nakamura, K., Watanabe, K., Kawabata, H., Pelowski, M., & Steyrl, D. (2024). Cross-cultural comparison of beauty judgments in visual art using machine learning analysis of art attribute predictors among Japanese and German speakers. Scientific Reports, 14, 15948. https://doi.org/10.1038/s41598-024-65088-z

Miller, C. A., Zhang, K., & Hübner, R. (2025). On the universality of aesthetic preference and inference: A cross-cultural (Chinese–German) study. Humanities and Social Sciences Communications, 12, 1–19. https://doi.org/10.1057/s41599-025-04806-y

Darda, K. M., & Chatterjee, A. (2024). Cross-cultural aesthetics: Aesthetic contextualism and ingroup bias. Journal of Comparative Literature and Aesthetics, 47(3), 120–140.

Darda, K. M., Christensen, A. P., & Chatterjee, A. (2023). Does the frame of an artwork matter? Cultural framing and aesthetic judgments for abstract and representational art. Psychology of Aesthetics, Creativity, and the Arts, 17(4), 428–450. https://doi.org/10.1037/aca0000569

Bullot, N. J., & Reber, R. (2013). The artful mind meets art history: Toward a psycho-historical framework for the science of art appreciation. Behavioral and Brain Sciences, 36(2), 123–137. https://doi.org/10.1017/S0140525X1200213X

Salgues, S., Jacquot, A., Makowski, D., Tahar, C., Baekeland, J., Arcangeli, M., Dokic, J., Piolino, P., & Sperduti, M. (2024). Self-reference and emotional reaction drive aesthetic judgment. Scientific Reports, 14, 19699. https://doi.org/10.1038/s41598-024-68331-9

METR (2025). Evaluations of reward hacking in frontier models. https://metr.org/

Esteban Ruiz, J. A. Art as Structural Surplus: Toward a Relational Ontology Beyond Human Authorship (V2.3). PhilArchive and Zenodo, 2026.