Algorithmic Manipulations: Positioning Phantom Mirror in a Curatorial History of Computational Art

How does one position one’s work within the contemporary art world and the wider cultural discourse on AI? This essay addresses that question directly, taking up the task of locating Phantom Mirror in relation to art history, curatorial practice, and current debates around algorithms, datasets, and emotional expression.

1. Three modes of artistic engagement with AI

Artists today approach AI in at least three distinct ways.

A. Generative AI as tool
This is the most familiar workflow: prompts lead to images or video. Artists focus on output aesthetics, using iteration, upscaling and light post-production.

B. Algorithms and datasets as material
Here artists engage directly with the infrastructure of AI. They curate datasets, write code, and interrogate the politics of training data. The work is both technical and conceptual, treating algorithms and datasets as raw artistic matter.

C. Algorithmic manipulation
This describes artists who remain inside closed tools, but use them as collaborators to be bent and steered. Instead of passively accepting outputs, they curate prompts, stage iterations, reject and refine. The goal is not slickness but expression. This labour produces images that capture affective states, even when the system is not designed for such work.

Phantom Mirror operates within C, borrowing from B through its textual staging of inputs and selection strategies, and delivered through A. The hybridity of this practice is what makes it distinctive.

What are algorithms and datasets?

In contemporary curatorial writing and art criticism, it is common to see artists described as “working with algorithms and datasets.” The phrase has become a shorthand in reviews and exhibition texts, but it is often left unexplained. This section addresses that gap, clarifying what is meant by algorithms and datasets in the context of computational and AI-based art.

An algorithm is a finite sequence of rules or instructions for solving a problem. The idea is ancient: Euclid’s algorithm for finding the greatest common divisor dates to around 300 BCE, and the 9th-century Persian mathematician al-Khwarizmi gave the concept its name. Even a recipe can be described as an algorithm for preparing a meal: a list of steps that transforms raw ingredients, the equivalent of a dataset, into a finished dish.

In computing, every program is an algorithm or a collection of algorithms. What has changed in recent decades is that algorithms have become social as well as technical forces. Once hidden inside machines, they now shape what people see, think, and do, from social media feeds to AI image generators. For artists, this shift has made algorithms visible as cultural agents in their own right, and therefore material for aesthetic exploration.

When artists speak of manipulating algorithms, this can occur at several levels. At the most direct level, some artists literally write or modify code with artistic intention, creating custom algorithms or altering existing ones to generate images, sounds, or behaviours. Others design systems that steer how algorithms behave without rewriting them entirely, for instance by setting up feedback loops, adjusting parameters, or creating rules for how the algorithm applies to a dataset. Finally, even without coding, artists can expose or redirect algorithmic logics conceptually, staging processes so that audiences perceive how the system operates. Across these levels, the key is that the algorithm is treated not as neutral infrastructure but as artistic material, open to shaping, bending, or critique.

Today this takes on a wider resonance. If earlier artists treated algorithms as tools or materials to be bent to their own intentions, in contemporary life the reverse is increasingly true. Social and cultural experience is curated for us by machines, with algorithms organising attention, shaping perception, and directing behaviour. The algorithm has become not only an artistic medium but a curatorial force in its own right, one that artists seek to expose, contest, or repurpose.

A dataset is not simply a pile of data. In computational art, a dataset refers to a defined collection of material, images, sounds, texts, or sensor readings, used to train or guide a system. Building or curating a dataset can itself be an artistic act: a process of framing, selecting, and excluding. Datasets have been assembled from archives, drawn or photographed by hand, scraped from the internet, or generated in real time through sensors. Each carries assumptions about what counts as knowledge or representation, and artists have increasingly highlighted those choices as central to meaning.

Taken together, algorithms and datasets provide both the mechanics and the material of computational systems. An algorithm is the structured process; the dataset is the content on which that process operates. When curators speak of artists “working with algorithms and datasets,” they generally mean practices that do not merely use tools but engage critically with the underlying machinery. Such practices make the systems themselves visible, question their logics, and reshape them for expressive ends.

2. Artists working with algorithms and datasets

The artists considered here span from 2001 to 2025, crossing three major technological paradigms in computational practice: the early period of computational art before GANs, the GAN era, and the current diffusion era. GANs, or Generative Adversarial Networks, are machine learning systems first introduced in 2014 that generate images through a competitive process between two networks. Since 2022, diffusion models have largely replaced GANs, creating images by “denoising” random noise toward a text prompt. A fuller explanation of these technological shifts is provided in the note at the end of this essay.

Computational Art (Pre-GAN, 2000s)

Before the rise of neural networks, artists experimented with rule-based systems, data visualisation, and sensor-driven generativity.

Corby & Baily (Tom Corby and Gavin Baily): Atmosphere (2001, exhibited at the Institute of Contemporary Arts, London, 2001, and ZKM, Karlsruhe, 2007) translated climate models and scientific data into visual and sonic forms. Their work materialised systemic change, turning abstract numbers into perceptual experiences that brought environmental data into cultural consciousness.
Stanza: Sensity (2004–2009, exhibited at Plymouth Arts Centre, 2006, and later at international media art festivals) used real-time data streams from urban environments, such as sensors and CCTV, to create live generative artworks. Algorithms were made visible as performers, transforming the city itself into a generative system.

GAN Era (2014–2021)

The introduction of GANs in 2014 enabled artists to generate new images from training datasets. This period saw practices focused on dataset curation, adversarial training, and critiques of bias.

Anna Ridler: Mosaic Virus (2018, shown in AI: More Than Human, Barbican, London, 2019) involved hand-drawing over 10,000 tulip images to build a dataset for training a GAN. The work connected the speculative frenzy of tulip mania in 17th-century Holland with the volatility of cryptocurrency and AI speculation, demonstrating that the making of a dataset is already a form of authorship.
Memo Akten: Learning to See (2017, exhibited at Ars Electronica, Linz, 2018, and AI: More Than Human, Barbican, 2019) developed custom machine learning systems to visualise ecological and perceptual data. His installations made computational processes perceptible, revealing the entanglement of human and non-human systems.
Morehshin Allahyari: Material Speculation: ISIS (2015–2016, exhibited at The Whitney Museum, New York, 2016, and ZKM, Karlsruhe, 2017) used 3D modelling and archival datasets to reconstruct destroyed artefacts from the Middle East. The project confronts cultural erasure and the colonial politics of digital reconstruction.
Jake Elwes: Zizi – Queering the Dataset (2019, Gazelli Art House, London) augmented facial recognition training sets with drag and queer faces. The project exposed the biases embedded in AI and proposed alternative imaginaries that resist normative categorisation.
Markos Kay: A Visual History of Molecular Biology (2016, shown in science-art contexts such as the Science Gallery, London) produced generative visualisations of molecular and cellular processes using scientific datasets. While not a GAN-based work, it exemplifies how computational systems can make unseen structures of life perceptible. Kay’s practice can be read as a bridge between dataset-driven computational art and the later aesthetics of diffusion, prefiguring the visual strategies that diffusion models would make widely available after 2022.

Diffusion Era (2022–present)

The release of diffusion models such as DALL·E 2, Stable Diffusion, and MidJourney in 2022 marked a historical shift. Instead of adversarial training, diffusion models iteratively “denoise” visual static into images guided by text prompts. This made image generation more accessible, high-resolution, and culturally mainstream.

Refik Anadol: Unsupervised (2022–2023, Museum of Modern Art, New York) transformed MoMA’s own collection into an evolving data-driven installation. Vast datasets became architectural experiences, immersing viewers in algorithmically generated “data paintings” that shifted continuously in real time.
Holly Herndon and Mat Dryhurst: The Call (Serpentine North, London, 2024–25) built from custom datasets of recorded choirs and voices, training models that generate new choral compositions. The project interrogated authorship, consent and collective identity, situating diffusion tools within a broader discourse on the politics of the voice.

Contemporary Developments (2024–2025)

Recent years have seen artists exploring psychoanalytic, affective, and identity-based concerns through new AI tools, connecting closely to the concerns of Phantom Mirror.

Sougwen Chung: New Beginnings (2024, HOFA Gallery, London) showcased her ongoing human–AI collaborative practice. Chung combines robotic systems, AI, and her own gestural mark-making to explore entanglement, authorship, and the blurred line between human expression and algorithmic process.
Kate Youme: ME., ME. (you) Made Me (2025, Inanimate, Hundred Years Gallery, London) created an “AI dominatrix” artwork based on the artist’s physical likeness, using open-source code to operate across platforms. The work interrogates automation, authorship, sexual agency, and coded desire, highlighting how identity is reshaped by algorithmic systems.

3. Phantom Mirror as algorithmic manipulation

Phantom Mirror begins with textual dialogue that shapes the conceptual corpus before images are generated. This preparatory stage acts as a kind of pre-dataset. Once in MidJourney, images are created, rejected, iterated and refined until a specific emotional expression is found. The work is not about technical spectacle but about the pursuit of affective states such as grief, estrangement and the unhomely.

This process can be understood as algorithmic manipulation in four ways:

Pre-dataset framing: The textual dialogue curates input space, functioning as a dataset in miniature.
Iterative affect: Images are not taken at face value but steered until they achieve a felt state. Refusal becomes part of the method.
Psychoanalytic transference: The model acts as a mirror, producing images that materialise inner experience and enable processes of mourning.
Curatorial authorship: Authorship resides in the path of iteration, refusal and framing, not in any single image.

This approach demonstrates that the manipulation of closed models can itself be an expressive practice.

4. Theoretical perspectives

The theoretical framework for Phantom Mirror draws initially on psychoanalysis and phenomenology. Freud’s account of mourning provides a way to read the series of images as repetitions that do not resolve but gradually reposition the lost object within psychic life. Lacan’s theory of the Mirror Stage is echoed in the project’s play of misrecognition, where images appear simultaneously as self and not-self. Winnicott’s concept of the holding environment clarifies how the AI model functions as a container for difficult affects, enabling play that leads to expression. Barthes’ punctum and Benjamin’s aura suggest that even within the computational image, moments of piercing affect and relational aura can emerge. Merleau-Ponty’s phenomenology of perception illuminates how the work visualises disorientation and the feeling of not being at home, while Borges’ Aleph is a reminder of the collapse of infinite perception into overwhelming fragments.

Alongside these earlier frameworks, the project also resonates with posthumanist theory. Rosi Braidotti argues in The Posthuman (2013) and Posthuman Knowledge (2019) that subjectivity is distributed and technologically embedded, a claim that aligns with Phantom Mirror’s exploration of mourning through hybrid human–machine processes. Katherine Hayles’ insistence in How We Became Posthuman (1999) that information and embodiment can no longer be separated is reflected in the project’s translation of psychic experience into computational form. Donna Haraway’s cyborg figure, still influential, provides another lens for understanding Phantom Mirror as a negotiation between organic and machinic selves. Karen Barad’s concept of agential realism emphasises entanglement, suggesting that the work does not merely depict but actively co-produces the experiences it stages. Yuk Hui’s exploration of cosmotechnics and recursivity points to the cultural and philosophical stakes of algorithmic systems, situating Phantom Mirror within broader debates about technology and human futures.

Most recently, theorists writing in 2024 and 2025 have emphasised that humans and algorithms can no longer be understood as separate entities. Seth Lazar (2024) describes how algorithms allocate and distribute attention, shaping not only what is seen but also what becomes socially meaningful. Octavian Machidona (2025) critiques recommender systems as technologies that reduce human complexity and exploit vulnerability, suggesting that responsibility lies not in switching off algorithms but in re-adapting to them. W. Lu and colleagues (2024) underline how algorithmic decision-making creates feedback loops that distort autonomy and reinforce identities, while Manoel Horta Ribeiro’s research demonstrates how amplification mechanisms radicalise content and perception. Taken together, these perspectives show that algorithms are not external tools but constitutive forces. Phantom Mirror can be read within this context as a form of intervention: an aesthetic practice that manipulates algorithms not for optimisation or persuasion but for the labour of grief, recognition, and psychic survival.

5. Why this positioning matters

By situating Phantom Mirror as algorithmic manipulation, the project aligns with artists who interrogate the infrastructures of AI, while emphasising its unique psychoanalytic, phenomenological and posthumanist focus. The work demonstrates that affect, aura and presence are not erased by computation but can reappear through careful manipulation, refusal and framing.

Note on the Genealogy of Computational Art: From Early Computer Art to GANs and Diffusion

Computational art has a longer history than the current AI boom might suggest. From the earliest plotter drawings in the 1960s to the mass adoption of diffusion models after 2022, artists have repeatedly adapted new technologies to aesthetic and conceptual ends. What unites these practices is the treatment of algorithms, code, and data as artistic material.

Early Computer Art (1960s–1970s)
This first wave of computer art emerged in research labs and early computing centres, where artists had access to mainframes. Pioneers such as Frieder Nake, Vera Molnár, and Harold Cohen experimented with algorithms to generate drawings on plotter printers. Cohen’s program AARON, developed from the late 1960s onwards, could autonomously produce line drawings that gradually became more complex over the decades. These works established that algorithms could themselves be aesthetic systems.

Generative and Interactive Art (1980s–1990s)
As computing became more accessible, artists began creating interactive installations and generative multimedia environments. Figures such as Jeffrey Shaw, David Rokeby, and Myron Krueger explored how systems could respond in real time to human presence, gesture, or behaviour. These practices shifted computational art from static outputs toward live, participatory encounters.

Computational and Data-driven Art (2000s)
The rise of the internet and ubiquitous computing gave artists access to new materials: surveillance systems, sensor networks, and real-time data streams. Works like Corby & Baily’s Atmosphere (2001) and Stanza’s Sensity (2004–2009) exemplified this period, in which environmental and urban data became raw material for generative systems. The emphasis was on computation as an interface between human, machine, and environment.

GAN Era (2014–2021)
Generative Adversarial Networks (GANs), introduced by Ian Goodfellow in 2014, transformed AI art. GANs consist of two neural networks in competition: a Generator that produces images and a Discriminator that evaluates them as real or fake. Through this adversarial process, GANs learn to create outputs that resemble the training dataset. For artists, this meant that curating the dataset became a central creative act. Anna Ridler’s Mosaic Virus (2018) demonstrated this by hand-drawing a dataset of 10,000 tulips, while Jake Elwes’Zizi – Queering the Dataset (2019) critiqued bias by inserting drag and queer faces into facial recognition sets. The GAN era foregrounded questions of authorship, bias, and cultural politics in relation to datasets.

Diffusion Era (2022–present)
Since 2022, diffusion models such as DALL·E 2, Stable Diffusion, and MidJourney have dominated generative AI. Unlike GANs, diffusion models begin with random noise and iteratively “denoise” it to form an image that corresponds to a text prompt. This technique produces coherent, high-resolution results and allows direct, intuitive interaction via language. Diffusion marked a watershed because it opened AI image generation to mass use: artists and amateurs alike could produce sophisticated images with minimal technical knowledge. At the same time, this shift raised new issues of authorship, aesthetics, and ethics, as datasets scraped from the internet underpinned many diffusion systems. Phantom Mirror belongs firmly to this diffusion era, while also drawing conceptually on dataset-centred practices of the GAN period.

Chronological Outline of Computational Art

1960s–1970s: Early Computer Art
- Key technology: Mainframes, algorithmic drawings, plotter printers.
- Artists: Frieder Nake, Vera Molnár, Harold Cohen (AARON).
1980s–1990s: Generative and Interactive Art
- Key technology: Multimedia systems, interactive installations, real-time feedback.
- Artists: Jeffrey Shaw, David Rokeby, Myron Krueger.
2000s: Computational and Data-driven Art (Pre-GAN)
- Key technology: Rule-based systems, sensors, surveillance, real-time data streams.
- Artists: Corby & Baily (Atmosphere), Stanza (Sensity), Rafael Lozano-Hemmer.
2014–2021: GAN Era
- Key technology: Generative Adversarial Networks (Generator vs Discriminator).
- Artists: Anna Ridler (Mosaic Virus), Memo Akten (Learning to See), Morehshin Allahyari (Material Speculation: ISIS), Jake Elwes (Zizi – Queering the Dataset).
2022–present: Diffusion Era
- Key technology: Diffusion models (iterative denoising guided by text prompts).
- Artists: Refik Anadol (Unsupervised), Holly Herndon & Mat Dryhurst (The Call), Eliza Stephens (Phantom Mirror).

Exhibitions & Artworks

Barbican Centre (2019) AI: More Than Human [Exhibition]. London: Barbican Centre, May–August.
Ars Electronica (2018) Festival for Art, Technology and Society [Exhibition]. Linz: Ars Electronica.
Whitney Museum of American Art (2016) Material Speculation: ISIS [Exhibition]. New York: Whitney Museum of American Art.
ZKM | Centre for Art and Media Karlsruhe (2007) Exhibition of Atmosphere [Exhibition]. Karlsruhe: ZKM.
Institute of Contemporary Arts (2001) Atmosphere [Exhibition]. London: ICA.
Plymouth Arts Centre (2006) Sensity [Exhibition]. Plymouth: PAC.
Serpentine Galleries (2024–2025) Holly Herndon & Mat Dryhurst: The Call [Exhibition]. London: Serpentine North.
Gazelli Art House (2019) Jake Elwes: Zizi – Queering the Dataset [Exhibition]. London: Gazelli.
Museum of Modern Art (2022–2023) Refik Anadol: Unsupervised [Exhibition]. New York: MoMA.
Science Gallery London (2016) Markos Kay: A Visual History of Molecular Biology [Exhibition]. London: Science Gallery.
HOFA Gallery (2024) New Beginnings [Exhibition]. London: HOFA.
Hundred Years Gallery (2025) Inanimate [Exhibition]. London: Hundred Years Gallery.

Theoretical References

Braidotti, R. (2013) The Posthuman. Cambridge: Polity Press.
Braidotti, R. (2019) Posthuman Knowledge. Cambridge: Polity Press.
Hayles, N. K. (1999) How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics. Chicago: University of Chicago Press.
Haraway, D. (1991) Simians, Cyborgs, and Women: The Reinvention of Nature. London: Routledge.
Barad, K. (2007) Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Durham, NC: Duke University Press.
Hui, Y. (2019) Recursivity and Contingency. London: Rowman & Littlefield.
Lazar, S. (2024) Communicative Justice and the Distribution of Attention. arXiv preprint.
Machidona, O. M. (2025) Beyond Algorethics: Addressing the Ethical and Anthropological Challenges of AI Recommender Systems. arXiv preprint.
Lu, W., Shen, Y., Liu, Y., Wu, Y. and Zhang, X. (2024) ‘Ethical concerns in personalised algorithmic decision-making’, Humanities and Social Sciences Communications, 11(1).
Ribeiro, M. H. (2024) Research on algorithmic amplification and radicalisation. Various publications.

Technical References

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. and Bengio, Y. (2014) ‘Generative Adversarial Networks’, Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS).
OpenAI (2022) DALL·E 2 Research Release. San Francisco: OpenAI.
Stability AI (2022) Stable Diffusion 1.0 Release. London: Stability AI.
MidJourney (2022) Public Beta Launch. San Francisco: MidJourney.