Abstract
How Computers Imagine Humans? began in 2016 from a small but persistent irritation I encountered while working with machine vision: a face detector was certain it had “seen” a face, although no photographed face was present. The installation used two computers facing each other: one generated drawn-like noise, while the other filtered the visual stream through a conventional face detector. When the detector fired, the detected patch was harvested and slowly layered into a portrait. The result looked human, not because a human was there, but because repeated trials had finally satisfied a criterion. This paper does not claim that the work technically anticipated contemporary generative models. Rather, it argues that the work made visible, at an artisanal and legible scale, a logic that continues to shape synthetic culture: brute force as a material and cultural regime of repeated sampling, filtering, optimization and computational throughput. Written from a practice-based media arts perspective, the article treats the installation as both an artwork and an epistemic device: a system through which recognition thresholds, misrecognition, machine authority and the labour of repetition become perceptible. The paper connects this loop to contemporary generative AI, through which plausible images, texts and identities can be produced cheaply and at scale, while verification, authorship and accountability become harder to stabilize. It also argues that synthetic appearance remains distinct from embodied experience: a face may be manufactured according to criteria, but it does not thereby acquire memory, history, or lived bodily presence. By revisiting How Computers Imagine Humans? nearly ten years later, in the generative AI era, the article proposes that computational art can expose what smooth AI interfaces often hide: the threshold, the archive, the failure, the selection process and the human responsibility required to turn a system into art.
Introduction
Generative AI is often explained as an engineering timeline: GANs appear after earlier procedural approaches, diffusion models follow, large language models become interfaces, and the next leap is explained through scale, larger datasets, greater compute and improved optimization. This story is useful, but it leaves something important aside. It explains how synthesis became feasible at scale, but it explains less well why scale itself has become such a decisive cultural logic. When images, voices, and texts can be produced on demand, the status of likeness, evidence, authorship and artistic labour changes. When plausibility becomes abundant, it becomes less informative. In this paper, brute force does not mean mere blind randomness. It names a regime of scale: repeated sampling, massive datasets, optimization, filtering and computational throughput through which plausibility is manufactured. I use the term as a cultural-material lens, not as a claim that contemporary AI architectures can be reduced to blind trial and error. AI should not, therefore, be understood only as a generation tool, but as a cultural technology that reconfigures practices of production, validation, and authority (González Díaz, 2021; Brandão et al., 2024). Computational art offers a useful position from which to examine this shift because it often treats algorithms less as solutions and more as cultural propositions. It asks not only what an algorithm can do, but what it makes visible, what it hides, and which forms of authority it invites. This becomes especially important when synthetic media ceases to be merely an aesthetic question and begins to affect trust, accountability and power relations (Carvalhais & Cardoso, 2015; Lee, 2019). I write from a media arts practice perspective, in which software is material and the exhibition context is part of the method. I do not use the artwork as a convenient illustration of a later theory; I treat it as a practice-based research case that produces knowledge by making recognition thresholds, repetition, selection and failure perceptible. Methodologically, I combine close analysis of the installation, technical contextualization of face detection and generative AI, and reflection on what the artwork made perceptible through exhibition and iteration. I therefore treat the installation as an epistemic device: not only as an object to be interpreted, but as a system that produces knowledge about recognition, thresholds and authority. Rather than separating the detector from its cultural work, I read the installation as a way to render thresholds, failures and misrecognitions tangible, and to discuss them as aesthetic and technical decisions. Initiated in 2016 and later presented in several contexts, How Computers Imagine Humans? marked a turning point in my trajectory because it raised questions that became mainstream only later, with the explosion of generative AI image and other media systems. It did not anticipate those systems technically. Instead, it made visible, in a small and inspectable loop, the logics of generation, testing, rejection, and accumulation that has since become central to synthetic media.
In the piece, the algorithm functioned as an operative artistic agent, not as an autonomous author, and the artistic force came from making the labour of brute force legible. It did not process an existing portrait. It generated bodies from elementary geometry, points, and lines, keeping the results close to abstraction and minimalism. I programmed the work in C++ using OpenFrameworks, an open-source creative coding framework, so that the real-time loop remained visible, responsive and materially close to the machine. The monochrome palette and visible noise were not decoration. They deliberately removed social context, so that attention remained on structure and on the machine’s criteria rather than on identity cues. This was important to the work. The noisy, imperfect outputs demystified the technology and made clear that computer vision was a mathematical interpretation, not a replica of reality. By “instructing” the system through code rather than through natural-language prompts, the work also foreshadowed prompt culture while keeping agency visible and contestable. By asking what remained of the human when the detector authorized a non-human image, it brought questions of identity and evidence to the center of the experience. Compared with contemporary GenAI portrait production, the contrast was clear. Current systems often rely on large-scale datasets of real images and tend to pursue realism or borrowed styles. The installation, by contrast, operated with hand-built rules and ran in real time, producing an infinite number of variations through repeated search. It pursued machine-vision aesthetics rather than human resemblance and made selection and rejection visible as craft. The work also rhymed with diffusion models through a shared relation to noise. Diffusion pipelines typically begin with random noise and iteratively remove it until a plausible face appears. This installation moved in the opposite direction, or stopped midway, keeping noise present so the viewer could see how form was being negotiated. Failure, glitch, and misrecognition were treated as productive outcomes. From this practice-based position, AI art becomes most interesting when the artist carefully, sometimes poetically, adjusts the system until its own rules begin to show.
How Computers Imagine Humans? served as the practical anchor for this inquiry. The work began with a mundane but consequential fact: false positives exist. A face detector does not search for a person. It searches for configurations that match its learned and engineered criteria, whether the input comes from a photograph, a drawing, a rendering or noise. The installation exploited this gap between referent and criterion: it generated images from noise and simple geometries, then collected regions that the detector mistakenly labelled as faces, compositing them into portraits that looked human yet depicted no human subject (Moura, 2017; Moura & Ferreira Lopes, 2017). These false positives matter here not as the paper’s final concept, but as visible evidence of a deeper mechanism. The work showed how brute force can make a criterion productive: generate enough, test enough, reject enough, and the accidental begins to resemble intention. On my website, I described the setup in simple terms: AI was used against AI, with a noise generator on one computer and a face detector on the other. The description also noted that the detector could be bypassed by holding a sign with one eye. In this slightly theatrical setup, a technical quirk became an experience: the face appeared as an apparition authorized by the detector, and viewers confronted the question of how recognition criteria could become generative aesthetics. The work’s exhibition reflected his public dimension. The installation was first presented in China in 2017, at ARTECH 2027, then in Germany in 2018, and later shown for an extended run in the curated programme of the 25th ISEA International Symposium on Electronic Art in Gwangju, Republic of Korea, in 2019. These contexts were significant because the piece functioned as a public experiment in algorithmic authority. The portrait was not a record of a person; it was a record of what repeated computation could eventually make acceptable, with no soul, no history and no memory (Moura, 2017). It also offered an example of how AI and facial parametrization reshape portrait regimes beyond mere resemblance (Camilo, Pires & Grigoraş, 2025).
Beyond this technical loop, the installation also functioned as a sculptural situation. Two ordinary computers faced each other like two bodies in a minimal encounter; no cable or wireless protocol carried the decisive exchange. Communication occurred through vision, across the air separating them, via the brightness of noise pixels generated by one computer’s screen and read by the other’s camera. This gave the work an analogue, almost poetic quality: computational randomness became visible light, and the space between the machines became part of the piece. When the installation was shown in Germany in 2018, in a deconsecrated church and placed at the main altar, this condition became even more explicit. The synthetic face appeared not only as a technical result but as a spectral image produced by two machines looking at one another (Moura, 2017). This altar installation is shown in Figure 1.
The article develops two related claims. First, How Computers Imagine Humans? can be read as a manual adversarial system: procedural generation proposed candidates, a fixed detector selected and accumulation produced an emergent portrait. Second, I argue that the work did not technically anticipate current generative models but made visible a principle that continues to shape generative AI: plausible likeness is often produced less by understanding than by large- scale search, repeated sampling and criteria of acceptance. In this work, “false-positive portraiture” was the visible surface of that operation. The deeper issue was brute-force plausibility: the manufacture, tuning, and accumulation of human-like form without referential grounding (Fallis, 2021; Vaccari & Chadwick, 2020).
From detection to manual brute-force aesthetics
Face detection has a foundational role in many computer vision pipelines because it reduces an unconstrained image to candidate regions for further interpretation. Classic surveys emphasize that detection is a matter of pattern recognition rather than an understanding of personhood (Yang et al., 2002). The boosted cascade approach, popularized by Viola and Jones (2001), achieved real-time performance by combining simple Haar-like features with boosting and a cascade that selectively allocates computation. Extensions, such as richer feature sets, were aimed at improving robustness to variation (Lienhart & Maydt, 2002). Whatever the variant, the practical consequence is a detector that encodes expectations about what a face should look like, shaped by efficiency constraints and training data. How Computers Imagine Humans? inverted that intention.
Instead of asking whether a face existed in an image, the work searched the space of possible images for those that triggered the detector. The generator was deliberately simple: a stream of visual noise built from repeated grey lines and occasional triangles. This choice mattered because the texture read as mark-making rather than television static, encouraging viewers to interpret the output as an image rather than a mere glitch. The detector evaluated each candidate frame. When a false candidate was detected, the region was cropped and retained. Additional checks, including small rotations and minimum thresholds for repeated detections, were applied to stabilize selection before compositing (Moura & Ferreira Lopes, 2017). The point was not to “improve” detection. But to reveal detection as a boundary in feature space that brute force could probe, push and eventually inhabit.
The work’s adversarial character was realized through brute-force search and selection, rather than through end-to-end gradient training. In canonical GANs, a generator and a discriminator co- evolve through optimization of data distributions (Goodfellow et al., 2014). Here, the discriminator was a fixed detector, and the generator was a procedural hypothesis engine.
Repetition and compositing served as substitutes for optimization. Analytically, this made the criterion legible. The portrait’s human likeness was not evidence of a human; it was evidence of a classifier’s priors under repeated pressure. Seen through today’s generative art discourse, the loop also aligns with attempts to formalize creativity as deviation rather than mere imitation. In Creative Adversarial Networks, Elgammal et al. modified the GAN objective so that generated images remain legible as “art” while intentionally confusing style classification, rewarding departures from established style norms. The loop in my work was manual and non-gradient, but the artistic question was related: when does a boundary stop being only a gatekeeper and become an engine of surprise, and what kinds of authority are built into that boundary? (Boden, 2010; Elgammal et al., 2017). Still, deviation in an objective function is not, by itself, an artistic act.
Without an authorial address, situated stakes, and responsibility for meaning, it remains optimization that merely looks like novelty. Compositing was essential to the piece’s aesthetics and argument. Detections were layered at low opacity, so a face seemed to emerge slowly, almost reluctantly. This span matters. It lets the viewer perceive brute force as time, accumulation, and labour, rather than as an invisible computational fact. It also invites a procedural reading of machine vision and gives the viewer time to notice how their own pareidolia aligns with the detector’s output. This aligns with accounts of ergodic artefacts, where interpretation involves non-trivial effort over time rather than immediate consumption (Carvalhais & Cardoso, 2015).
The emergent portrait became an interface between machine criteria, repeated computation and human pattern seeking. The resulting apparition of a face is visible in Figure 2.
Generative AI as a brute-force replication regime
The expansion of generative AI since the mid-2010s can be read as the industrialization of brute- force replication. The architectures differ, and the argument is not that diffusion models, language models, or adaptation techniques are technically equivalent to blind search. Rather, a shared production pattern appears across domains: generate many candidates, select, refine, regenerate. Diffusion models formalize image synthesis as iterative denoising guided by learned representations (Ho et al., 2020; Rombach et al., 2022). Large language models generalize predictive sequence modelling into a flexible interface for drafting, rewriting and summarizing text (Brown et al., 2020; Hoffmann et al., 2022). Scaling law research makes explicit that much of the improvement comes from increasing compute and data under broadly similar objectives (Kaplan et al., 2020). In practice, this becomes a production regime. It accelerates ideation, mock- ups, documentation, and synthetic variation while relocating labour towards validation, governance and liability. This framing clarifies why How Computers Imagine Humans? remains relevant beyond the specific issue of false positives. The core change is not only greater realism, but also the collapse of the marginal cost of plausible variation. When plausible outputs are cheap and abundant, plausibility becomes less informative. The cultural task shifts to discrimination: deciding which artefacts can be trusted, credited or situated. In the installation, discrimination was performed by a face detector. In society, discrimination is distributed across institutions, platforms and social norms.
Selection, validation, and the new craft of checking
A blunt description of the current moment is simple: generation is cheap, and checking is expensive. This is one of the cultural effects of brute force. Many workflows begin with the rapid production of plausible variants and end with the slower work of determining what is correct, safe or ethically sound. In earlier media regimes, some validation was indirectly embedded in production costs. In generative workflows, the cost moves downstream and often disappears from view. As González Díaz noted, a “pretense of transparency” can “lead to a lack of sense of information” (González Díaz, 2021). This relocation has social consequences. Those with time, organizational support, and domain expertise are better placed to verify outputs. Those who read fluent output as authority are more exposed to error, precisely because confidence is so easy to simulate. Expertise itself also changes its public form. Speed is easy to perform publicly; justification is slower and harder to display. The imbalance is therefore not only computational, but cultural: brute force accelerates production while leaving interpretation, verification, and responsibility to slower human and organizational processes. Edwards offered a plain illustration of this gap between function and grasp: “The code worked, but the understanding had never been his.” (Edwards, 2026, Chapter 3).
Empirical research suggests that evaluations of AI-mediated cultural objects depend heavily on attributed agency, not only on perceptual qualities. In visual art, perceived effort, intentionality, and authenticity shape value judgments and distinguish human from machine provenance (Samo & Highhouse, 2023). A meta-analysis reports a systematic penalty for works labelled as AI-generated, although effect sizes vary by context and participant expertise (de Rooij, 2025). Information about production processes can also recalibrate moral and aesthetic appraisal: factual descriptions of AI back-end mechanisms reduced moral acceptability and aesthetic appeal in some situations, while disclosure of AI collaboration can reduce admiration for both the artwork and the artist by triggering a perceived loss of creative authenticity (Bara et al., 2025; Messer, 2024). How Computers Imagine Humans? compressed this dynamic into a single loop. The system generated many candidates and retained only those accepted by a detector. Yet acceptance does not mean the system “understood” anything. It only means that the output survived a criterion after repeated trials. In contemporary synthetic media, selection often works in related ways. Outputs are filtered by user preferences, platform policies and distribution incentives.
Successful ones are amplified. Amplification is then easily misread as evidence of truth or quality. Verification, therefore, cannot be reduced to a purely technical act. It includes disclosure norms, documentation of datasets and prompts, and audit practices. Raji et al. (2020) argue for internal algorithmic auditing as an end-to-end framework spanning design, development and deployment. In cultural contexts, an analogous audit asks who is credited, whose data is used, how consent is handled, and what is disclosed to audiences.
Archives, priors, and the past bound model
Once checking becomes central, another property becomes difficult to ignore: most model competence is structurally past-bound. Generative systems learn from archives, and these archives are uneven. They overrepresent what is documented, widely circulated, and platform- friendly, and underrepresent situated knowledge that is less documented or less digitized. This is why debates about data governance are not external to aesthetics. They shape what can be imagined computationally. Brandão et al. (2024) frame this problem within media arts by noting that AI tools may produce “conservative and ‘average’ and predictable results”, a tendency linked to their statistical nature and dependence on pre-existing data. Bender et al. (2021) argue that large language models can produce fluent output that appears to reflect understanding while remaining grounded in probabilistic pattern completion. Datasheets for datasets were proposed to document how datasets were constructed and what they contain, so that social and ethical implications are not hidden (Gebru et al., 2021). These interventions matter because they resist treating training corpora as neutral raw material. Brute force is therefore never abstract. It is brute force applied to specific archives, particular priors and particular omissions. How Computers Imagine Humans? provided a functional analogue. The face is not discovered; it is assembled from what the detector is prepared to accept after enough attempts. Likewise, synthetic portraiture in contemporary systems is assembled from learned priors about faces, bodies and identities. Better realism can increase persuasive force without increasing grounding. This is central to evidentiary manipulation, where plausible media can enable both deception and denial (Fallis, 2021).
Counterfeit culture and evidentiary fragility
Synthetic media intensifies longstanding problems of forgery and misattribution by lowering the cost of plausible repetition. Deepfake research documents both the increasing capability of manipulation and the adversarial dynamics that complicate detection (Mirsky & Lee, 2021; Tolosana et al., 2020). Fallis (2021) emphasizes that the damage is not only direct deception, but also the “liar’s dividend,” where genuine evidence can be dismissed as fake. In cultural economies, the same logic supports counterfeit identities, automated portfolios and volume- driven content flooding. Brute force intensifies the problem because plausible production can be repeated until attention, attribution and verification are exhausted. Revisiting Benjamin’s analysis of mechanical reproduction, Rabinovich and Foley argue that AI intensifies reproducibility in ways that risk a an entertainment culture of consensus, while also inviting artists to problematize these conditions as a public practice (Rabinovich & Foley, 2025).
Questions of authenticity and originality cannot be reduced to whether an artefact appears novel. The notion of semi-aura is useful here because it names a hybrid authenticity produced by the interplay between human intention and algorithmic execution (Salas Espasa & Camacho, 2025). Photography scholarship describes a parallel shift in which images may no longer operate as indices of external events, but as artefacts whose referents may be absent; under these conditions, framing becomes a criterion of authorship and consequence (Schofield, 2024). A critical framework for AI art must therefore examine not only visual novelty, but also the technical arrangements, cultural stratification, and ideological filtering that accompany AI as a medium (Grba, 2022).
These debates make clear that counterfeit culture is not only a technical detection problem. What is at stake is how authorship, responsibility, and value are recognized when convincing images no longer guarantee a stable referent. Some philosophers resist granting AI outputs the status of art in a strong sense because they lack intention, responsibility, and communicative address, even when they produce compelling surface effects (Chiodo, 2024). Technical responses such as watermarking support traceability but remain constrained by adoption, adversarial adaptation and unmarked generation pipelines (Dathathri et al., 2024). Authenticity in the arts is also a property of situated practice, curatorial framing and social interpretation. This is why computational art remains relevant: it can make the gap between brute-force plausibility and accountability perceptible. How Computers Imagine Humans? isolated that mechanism by showing that a face-like image can be manufactured through repeated attempts against a criterion, even when no face exists.
Epistemic prototyping and artistic method
When people say that art is “ahead” of science, it makes more sense as a claim about method than about hierarchy. Artistic practice can foreground ambiguity, failure modes, and lived consequences before those consequences are stabilized as benchmarks. Cohen’s AARON treated drawing as a formalizable practice and made the limits of formalization visible as part of the work (Cohen, 2002). Krueger’s Videoplace explored sensing and embodiment long before contemporary immersive platforms became commonplace (Krueger et al., 1985). This anticipatory function matters for AI because many risks are not primarily technical. They are structural risks related to incentives, disclosure and accountability. Here, the method is grounded in studio iteration and in the gallery as a site of critique, where the system’s behaviour is refined through embodied observation. A common strategy in computational art is critical inversion. Instead of using a system for its intended purpose, the artist uses it to expose the assumptions that make it work. How Computers Imagine Humans? inverted detection into generation. The work did not need a new model to critique model culture. It required a legible loop to render the decision boundary and the labour of brute force visible. In this sense, it functioned as a compact pedagogy for synthetic media: it showed what it means for a criterion to produce an image when enough attempts are allowed to pass through it. Boden describes three routes to surprise in art making: combinational creativity, which produces unfamiliar juxtapositions; exploratory creativity, which searches a structured conceptual space such as an artistic style; and transformational creativity, which alters the constraints that define that space. This vocabulary helps to name what happened when detection was inverted into generation. The work explored the conceptual space defined by a fixed detector, but it also proved transformative by redefining the role of that constraint. The detector’s norm was no longer an invisible gate. It became a visible aesthetic operator that viewers could watch at work (Boden, 2010). The latter exhibition context in Gwangju, Republic of Korea, is shown in Figure 3.
The industrial weight of computational dreaming
In media art, the last decade also marked a shift from comparatively legible, hand-coded procedures towards the industrial machinery of contemporary diffusion pipelines. Earlier computational works often presented rules, heuristics, and constraints in a form that could be inspected, sometimes even by non-specialists. Contemporary generative production relocates agency into dataset design, model scale and optimization. It ties aesthetic possibility to GPU access and parallel throughput. This material pressure is not only metaphorical: data movement and memory limits increasingly shape the energy, latency, and feasibility of large-scale AI, motivating research into processing-in-memory. In this sense, the field moved from explicit rules towards optimized brute force, where archives and compute budgets strongly influence what can be generated, how quickly variants can be explored, and who can afford to participate. The present wave did not replace the underlying logic of search and selection that already operated in earlier procedural systems. It mainly scaled that logic, with more data, tighter pipelines, and more compute, while still remaining bounded by the constraints of the present moment. Stable Diffusion is emblematic of this industrialization because it operationalizes image synthesis as an efficient denoising procedure in a compressed latent space (Rombach et al., 2022). Generation begins from stochastic noise and proceeds through iterative refinement steps that progressively reduce entropy under guidance from learned text and image representations. Even when the algorithm is formally structured, the experience can feel like sustained computational pressure: many parallel operations and repeated acts of constraint satisfaction, until noise collapses into recognizable form. This is why the earlier installation remains useful. It showed a related pressure in a small, visible, and human-scale loop.
Within this paradigm, artistic agency often shifts from direct mark-making towards the orchestration of sampling, guidance and adaptation. Prompting and curation remain meaningful, but they often function as steering within distributions rather than composition through explicit, irreversible choices. Fine-tuning methods such as LoRA intensify this shift by making personalization and stylistic specialization inexpensive and fast: small, low-rank update matrices steer a fixed base model towards a style, subject, or local dataset with limited compute and storage (Hu et al., 2021). In daily use, this facilitates the rapid circulation of style packs and character likenesses, while raising questions about consent, authorship and attribution. For media art, these systems invite two competing trajectories. One treats diffusion as a rapid styling engine and risks collapsing practice into trend-following, where novelty is tracked through tool releases, presets and adapter marketplaces. The other treats the pipeline itself as material, investigating how latent space, guidance and adaptation encode priors about bodies, realism and value. The works most likely to endure, I think, are those that render this industrial weight legible as an aesthetic and political fact, rather than treating it merely as a convenient technique. LoRA makes brute force more personal and more portable, but not necessarily more grounded.
LeCun’s critique of large language models gives this argument a useful resonance from inside AI research itself. As he wrote, “They are useful. But they are not a path to human-level AI” (LeCun, 2025). The issue is not simply that such systems are weak; it is that fluency can still lack world models, persistent memory, reasoning and planning. Sofia Miguens’s recent work on human, animal, and artificial minds also helps to frame artificial intelligence as more than an engineering event: it becomes a philosophical disturbance around language, consciousness, and what counts as mindedness (Miguens, 2025). For me, this returns to a question that has accompanied my practice-based work on embodiment: human intelligence is not only made from representation, but from the encounter between a body and its limits. The body is a frontier where the world resists us. How Computers Imagine Humans? produced a human-like face without that frontier: “no soul, no history, and no memory” (Moura, 2024).
If everyone is using the same increasingly accessible models, a structural question follows. Does human creation drift towards a kind of zero point, where difference is compressed into shared latent conventions and novelty becomes a small perturbation within an industrial imagination? In such a regime, distinctiveness risks being misrecognized as mere parameter choice, while cultural value shifts towards the speed of iteration and distribution rather than sustained commitments of attention, time and situated experience. The problem is not that brute force cannot produce surprising forms. It can. The problem is that surprise produced at scale can quickly become another default unless the artist turns the machinery back into a question.
Conclusion
Revisiting How Computers Imagine Humans? shows how a modest computational loop can illuminate a broader cultural shift. The work demonstrated that likeness is not guaranteed by reference. A criterion could manufacture it, and that criterion could be inverted into an aesthetic. More precisely, brute force turned a recognition threshold into a generator of human-like form. This is the quiet lesson that becomes loud in the generative era: recognition systems do not only judge images; they can also manufacture the conditions under which images count. The artwork matters because it makes this operation perceptible. The viewer senses the loop, the delay, the rejection, the accumulation, and the strange moment when repeated computation begins to look like intention.
Contemporary generative systems amplify this logic through throughput: more operations per millisecond, more memory, larger corpora, tighter pipelines, and better engineering around sampling, attention and alignment (Vaswani et al., 2017; Kaplan et al., 2020; Hoffmann et al., 2022). Because training draws on archives, competence remains bound to the past. These systems recombine records; they do not access futures that have not happened or have not been documented. Brute force is therefore a material and cultural regime: computational capacity, but also a social way of producing plausibility. Wider archives and higher fidelity do not, by themselves, produce truth. The gap between what appears right and what is grounded becomes a problem for any domain that relies on records, testimony and evidence. At the level of civic life, plausibility at scale can corrode the trust systems of universities, law, and journalism by making production cheap and verification socially exhausting, a process described as institutional degradation rather than mere misinformation (Hartzog & Silbey, 2025).
Computational art remains central because it can make this machinery legible. It can expose decision boundaries, data dependencies, and interface politics, while proposing ways of working in which human agency resides in constraints, curation and accountable provenance. Future AI advances will likely involve new architectures and new forms of optimization, as well as renewed forms of scaling, even when hidden behind smoother interfaces (Pires et al., 2025; Simão et al., 2025). LeCun, embodiment, and artistic labour converge here as consequences of brute-force plausibility, not as separate concerns. If systems can generate fluent forms without grounded understanding and detach appearance from lived history, then art must insist on situated limits, bodily memory and responsibility. Technology can generate forms, variations, and surprises, but it cannot, by itself, make art. Art begins when the author turns a system into a situated proposition and makes its limits speak. What becomes interesting in AI art is not fidelity to a learned archive, but the moment when the system’s own rules are bent, exposed or revised. The artwork is strongest when the author bends, exposes, or revises the system’s rules, rather than merely satisfying them (Boden, 2010; Elgammal et al., 2017).
For media arts students and artists working with AI, three practical consequences follow from this condition. First, challenge AI constantly: invert its intended use, impose constraints, force failure, test its thresholds, and refuse the easiest output. This challenge can make AI less boring, less narrowly engineered, and less controversially opaque, because the artwork begins where the system is pushed away from its default distributions. Second, make the limit perceptible: disclose the loop, show the boundary, and expose the moments in which the system selects, fails, repeats, or smooths difference. This does not remove controversy, but it makes it more precise, because the audience can see what the system can and cannot do. Third, learn how the models are built. Ask whether you are using a foundation model, an open- source model, a closed commercial system, or a model fine-tuned on a particular dataset. When possible, train or fine-tune small or older models with your own data: your images, sounds, texts, gestures or archives. The result will not automatically be better, but it will be more yours, more situated, and less identical to what everyone else obtains from the same generic interface. Perhaps this is one serious way to continue now: not only using AI but taking responsibility for the material from which it learns.
Since YMYI (Moura, 2007), and through later works such as Nuve (2007), Câmara Neuronal (2011), Wide/Side (2015), Co:Lateral (2015), VV (2018), Una (2020), Out > there (2021), and EDNI (2024), I have kept the lived body as a central force in my digital practice. We all live in bodies, and this lived condition continues to mark the difference between synthetic appearance and embodied experience. I still look for that bodily pressure today, and try to transmit it to media arts students, even when contemporary tools keep pushing effortless production. In older systems, limits were unavoidable; in contemporary AI systems, they often have to be actively exposed.
The final pressure is therefore not nostalgic, but ethical: not to let brute force set the direction merely because it can generate more. As Edwards warned, “If it sets direction when you haven’t, you’re no longer the author” (Edwards, 2026, Chapter 8).
References
BARA, Iionela; RAMSEY, Richard; CROSS, Emily S. (2025). “Contextual information about AI shapes moral and aesthetic judgments of AI-generated visual art”. Cognition, vol. 257, 106063. DOI: https://doi.org/10.1016/j.cognition.2025.106063
BENDER, Emily M.; GEBRU, Timnit; McMILLAN-MAJOR, Angelina; SHMITCHELL, Shmargaret (2021). “On the dangers of stochastic parrots: Can language models be too big”. In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 610-623. DOI: https://doi.org/10.1145/3442188.3445922
BODEN, Margaret A. (2010). Creativity and Art: Three Roads to Surprise. Oxford University Press [online]. Available at: https://global.oup.com/academic/product/creativity-and-art-9780199590735
BRANDÃO, Daniel; MOURA, João Martinho; SOMMERER, Christa (2024). “Media arts: Convergência de esferas artísticas na era do digital”. Vista, no. 14, e024018. DOI: https://doi.org/10.21814/vista.6105
BROWN, Tom. B.; MANN, Benjamin; RYDER, Nick; SUBBIAH, Melanie; KAPLAN, Jared; DHARIWAL, Prafulla; NEELAKANTAN, Arvind; SHYAM, Pranav; SASTRY, Girish; ASKELL, Amanda; AGARWAL, Sandhini; HERBERT-VOSS, Ariel; KRUEGER, Gretchen; HENIGHAN, Tom; CHILD, Rewon; RAMESH, Aditya; ZIEGLER, Daniel, WU, Jeffrey; WINTER, Clemens; AMODEI, Dario (2020). “Language models are few-shot learners”. arXiv. DOI: https://doi.org/10.48550/arXiv.2005.14165
CAMILO, Eduardo; PIRES, Helena; GRIGORAŞ, Flarin (2025). “O Retrato Fotográfico Como Campo de Forças: Política, Identidade e Resistência na Cultura Visual Contemporânea”. Vista, no. 16, e025021. DOI: https://doi.org/10.21814/vista.7047
CARRINHO, Pedro; MOGHADASPOUR, Hamid; FERRAZ, Oscar; FERREIRA, João Dinis; FALEVOZ, Yann; SILVA, Vitor; FALCAO, Gabriel (2025). “An experimental exploration of in-memory computing for multi-layer perceptrons”. Journal of Signal Processing Systems, vol. 98, no. 1. DOI: https://doi.org/10.1007/s11265-025-01974-7
CARVALHAIS, Miguel; CARDOSO, Pedro (2015). “What then happens when interaction is not possible? The virtuosic interpretation of ergodic artefacts”. Journal of Science and Technology of the Arts, vol. 7, no. 1. DOI:
CHIODO, Simona (2024). “What AI ‘art’ can teach us about art”. Journal of Aesthetics & Culture, vol. 16, no. 1, 2395511. DOI: https://doi.org/10.1080/20004214.2024.2395511
COHEN, Harold (2002). “A self-defining game for one player: On the nature of creativity and the possibility of creative computer programs”. Leonardo, vol. 35, no. 1, pp. 59-64. DOI: https://doi.org/10.1162/002409402753689335
DATHATHRI, Sumanth; SEE, Abigail; GHAISAS, Sumedh et al. (2024). “Scalable watermarking for identifying large language model outputs”. Nature, vol. 634, pp. 818-823. DOI: https://doi.org/10.1038/s41586-024-08025-4
DE ROOIJ, Alwin (2025). “Bias against artificial intelligence in visual art: A meta-analysis”. Psychology of Aesthetics, Creativity, and the Arts. Publicació avançada. DOI: https://doi.org/10.1037/aca0000833
ELGAMMAL, Ahmed; LIU, Bingchen; ELHOSEINY, Mohamed; MAZZONE, Marian (2017). “CAN: Creative Adversarial Networks, generating “art” by learning about styles and deviating from style norms”. arXiv. DOI: https://doi.org/10.48550/arXiv.1706.07068
EDWARDS, Helen (2026, February 21). “3. The Three Ways You’re Changing. Stay Human: Authoring Your Mind in the AI Age”. Artificiality Journal [online]. Available at: https://journal.artificialityinstitute.org/chapter-3-the-three-ways-youre-changing/. [Data de consulta: 2 May 2026].
EDWARDS, Helen (2026, April 29). “8. Staying the author. Stay Human: Authoring Your Mind in the AI Age”. Artificiality Journal [online]. Available at: https://journal.artificialityinstitute.org/8-staying-the-author/. [Data de consulta: 2 May 2026].
FALLIS, Don (2021). “The epistemic threat of deepfakes”. Philosophy & Technology, vol. 34, pp. 623-643. DOI: https://doi.org/10.1007/s13347-020-00419-2
GEBRU, Timnit; MORGENSTERN, Jamie; VECCHIONE, Briana; VAUGHAN, Jennifer W.; WALLACH, Hannah; DAUMÉ III, Hal; CRAWFORD, Kate (2021). “Datasheets for datasets”. Communications of the ACM, vol. 64, no. 12, pp. 86-92. DOI: https://doi.org/10.1145/3458723
GOODFELLOW, Ian; POUGET-ABADIE, Jean; MIRZA, Mehdi; XU, Bing; WARDE-FARLEY, David; OZAIR, Sherjil; COURVILLE, Aaron; BENGIO, Yoshua (2014). “Generative adversarial nets”. arXiv. DOI: https://doi.org/10.48550/arXiv.1406.2661
GONZÁLEZ DÍAZ, Paloma (2021). “Digital creation: Possible (and achievable) futures”. A: P. González Díaz & A. García Méndez (eds.). In the limits of what is possible: Art, science and technology. Artnodes, no. 28, pp. 1-8. DOI: https://doi.org/10.7238/artnodes.v0i28.388278. [Data de consulta: 16 April 2026].
GRBA, Dejan (2022). “Deep Else: A Critical Framework for AI Art”. Digital, vol. 2, no. 1, pp. 1-32. DOI: https://doi.org/10.3390/digital2010001
HARTZOG, Woodrow; SILBEY, Jessica M. (2025). “How AI destroys institutions”. Boston Univ. School of Law Research, Paper no. 5870623 [online]. Available at: https://ssrn.com/abstract=5870623
HO, Jonathan; JAIN, Ajay; ABBEEL, Pieter (2020). “Denoising diffusion probabilistic models”. arXiv. DOI: https://doi.org/10.48550/arXiv.2006.11239
HOFFMANN, Jordan et al. (2022). “Training compute-optimal large language models”. arXiv. DOI: https://doi.org/10.48550/arXiv.2203.15556
HU, Edward J.; SHEN, Yelong; WALLIS, Phillip.; ALLEN-ZHU, Zeyuan; LI, Yuanzhi: WANG, Shean; WANG, Lu; CHEN, Weizhu (2021). “LoRA: Low- rank adaptation of large language models”. arXiv. DOI: https://doi.org/10.48550/arXiv.2106.09685
KAPLAN, Jared; McCANDLISH, Sam; HENIGHAN, Tom; BROWN, Tom B.; CHESS, Benjamin; CHILD, Rewon, GRAY, Scott; RADFORD, Alec; WU, Jeffrey; AMODEI, Dario (2020). “Scaling laws for neural language models”. arXiv. DOI: https://doi.org/10.48550/arXiv.2001.08361
KRUEGER, Myron W.; GIONFRIDDO, Thomas; HINRICHSEN, Katrin (1985). “Videoplace: An artificial reality”. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ‘85), pp. 35-40. Association for Computing Machinery. DOI: https://doi.org/10.1145/317456.317463
LeCUN, Yann (2025). “Why LLMs are not human-level AI”. [LinkedIn post]. LinkedIn [online]. Available at: https://www.linkedin.com/posts/yann-lecun_if-you-are-interested-in-applications-of-activity-7322617933149474817-mYTl. [Data de consulta: 3 de maig de 2026].
LEE, Rosemary (2019). “Uncertainties in the algorithmic image”. Journal of Science and Technology of the Arts, vol. 11, no. 2), pp. 36-40. DOI: https://doi.org/10.7559/citarj.v11i2.661
LIENHART, Rainer; MAYDT, Jochen (2002). “An extended set of Haar-like features for rapid object detection”. In: Proceedings of the 2002 International Conference on Image Processing (ICIP 2002), vol. 1, pp. I-900-I- 903. DOI: https://doi.org/10.1109/ICIP.2002.1038171
MESSER, Uwe (2024). “Co-creating art with generative artificial intelligence: Implications for artworks and artists”. Computers in Human Behavior: Artificial Humans, vol. 2, no. 1, 100056. DOI: https://doi.org/10.1016/j.chbah.2024.100056
MIRSKY, Yisroel; LEE, Wenke (2021). “The creation and detection of deepfakes: A survey”. ACM Computing Surveys, vol. 54, no. 1. DOI: https://doi.org/10.1145/3425780
MIGUENS, Sofia (2025). Mentes humanas, animais e artificiais: O trabalho das palavras a partir de Daniel Dennett. Edições 70.
MOURA, João Martinho (2007). “YMYI” [Instal·lació interactiva]. João Martinho Moura [online]. Available at: https://jmartinho.net/ymyi/
MOURA, João Martinho (2017). “How Computers Imagine Humans?” [Instal·lació interactiva]. João Martinho Moura [online]. Available at: https://jmartinho.net/how-computers-imagine-humans/
MOURA, João Martinho (2024). “Realidades virtuais incorporadas” [Tesi doctoral]. Universidade Católica Portuguesa [online]. Available at: http://hdl.handle.net/10400.14/45509
MOURA, João Martinho; FERREIRA LOPES, Paulo (2017). “Generative face from random data, on “How Computers Imagine Humans””. In: Proceedings of the 8th International Conference on Digital Arts (ARTECH 2017), pp. 85-91. DOI: https://doi.org/10.1145/3106548.3106605
PIRES, Helena; BARROS, Né; MOURA, João Martinho (2025). “EDNI: Rehearsing the embodied co-creation of emerging trans-spatio-temporalities”. In: Proceedings of the 12th International Conference on Digital and Interactive Arts: Media Art Cultures, Communities & Territories (ARTECH 2025). DOI: https://doi.org/10.1145/3773699.3773708
RABINOVICH, Misha; FOLEY, Caitlin (2025). “The work of art in the age of AI reproducibility”. AI & Society, vol. 40, pp. 1565-1567. DOI: https://doi.org/10.1007/s00146-024-01991-3
RAJI, Inioluwa Deborah; SMART, Andrew; WHITE, Rebecca N.; MITCHELL, Margaret; GEBRU, Timnit; HUTCHINSON, Ben; SMITH-LOUD, Jamila; THERON, Daniel; BARNES, Parker (2020). “Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing”. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pp. 33-44. DOI: https://doi.org/10.1145/3351095.3372873
ROMBACH, Robin; BLATTMANN, Andreas; LORENZ, Dominik; ESSER, Patrick; OMMER, Björn (2022). “High-resolution image synthesis with latent diffusion models”. arXiv. DOI: https://doi.org/10.48550/arXiv.2112.10752
ROY, Kaushik; KOSTA, Adarsh; SHARMA, Tanvi; NEGI, Shubham; SHARMA, Deepika; SAXENA, Utkarsh; ROY, Sourjya; RAGHUNATHAN, Anand; WAN, Zishen; SPETALNICK, Samuel; STRACHAN, John P.; LIU, Che-Kai; RAYCHOWDHURY, Arjit (2025). “Breaking the memory wall: next- generation artificial intelligence hardware”. Frontiers in Science, vol. 3. DOI: https://doi.org/10.3389/fsci.2025.1611658
SALAS ESPASA, David; CAMACHO, Mar (2025). “From aura to semi-aura: reframing authenticity in AI-generated art-a systematic literature review”. AI & Society, vol. 40, pp. 6727-6759. DOI: https://doi.org/10.1007/s00146-025-02361-3
SAMO, Andrew; HIGHHOUSE, Scott (2023). “Artificial intelligence and art: Identifying the aesthetic judgment factors that distinguish human and machine-generated artwork”. Psychology of Aesthetics, Creativity, and the Arts. Advance online publication. DOI: https://doi.org/10.1037/aca0000570
SCHOFIELD, Michael P. (2024). “Camera phantasma: reframing virtual photographies in the age of AI”. Convergence, vol. 30, no. 2, pp. 687-709. DOI: https://doi.org/10.1177/13548565231220314
SIMÃO, Emília; MOURA, João Martinho; BRANDÃO, Daniel (2025). “Liminal landscapes: Generative art as a technoshanic instrument in the representation of altered states of consciousness”. In: Proceedings of the 12th International Conference on Digital and Interactive Arts: Media Art Cultures, Communities & Territories (ARTECH 2025). DO: https://doi.org/10.1145/3773699.3774340
TOLOSANA, Ruben; VERA-RODRIGUEZ, Ruben; FIERREZ, Julian; MORALES, Aythami; ORTEGA-GARCIA, Javier (2020). “Deepfakes and beyond: A survey of face manipulation and fake detection”. Information Fusion, vol. 64, pp. 131-148. DOI: https://doi.org/10.1016/j.inffus.2020.06.014
VACCARI, Cristian; CHADWICK, Andrew (2020). “Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news”. Social Media + Society, vol. 6, no. 1. DOI: https://doi.org/10.1177/2056305120903408
VASWANI, Ashish; SHAZEER, Noam; PARMAR, Niki; USZKOREIT, Jakob; JONES, Llion; GOMEZ, Adrian N.; KAISER, Łukasz; POLOSUKHIN, Illia (2017). “Attention is all you need”. arXiv. DOI: https://doi.org/10.48550/arXiv.1706.03762
VIOLA, Paul; JONES, Michael (2001). “Rapid object detection using a boosted cascade of simple features”. In: Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2001), vol. 1, pp. I-511-I-518. DOI: https://doi.org/10.1109/CVPR.2001.990517
WANG, Yi-Qing (2014). “An analysis of the Viola-Jones face detection algorithm”. Image Processing On Line, vol. 4, pp. 128-148. DOI: https://doi.org/10.5201/ipol.2014.104
YANG, Ming-Hsuan; KRIEGMAN, David J.; AHUJA, Narendra (2002). “Detecting faces in images: A survey”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 1, pp. 34-58. DOI: https://doi.org/10.1109/34.982883
Recommended citation: MOURA, João Martinho. Brute force and the synthetic human: revisiting How Computers Imagine Humans? in the generative AI art era. Mosaic [online], June 2026, no. 208. ISSN: 1696-3296. DOI: https://doi.org/10.7238/m.n207.2605
Deja un comentario