The Algorithmic Muse: How AI Art Fits Into Art History

A new deep dive into the phases of AI creativity—and why the human questions behind it are as old as art itself.

This week, I asked Dr. Gemini 2.5, a leading art historian, to produce a deep research report analyzing the different phases of AI art—and how they fit within the broader story of art history. I then asked esteemed copy-editor Claude 3.7 to rewrite it, shorten it, and make it more accessible for me to share here.

The goal was simple: to move past the endless surface debates about “is AI art real?” and instead place today’s AI explosion in a much longer historical context. How has technology always challenged art? How has art always evolved in response? Where is AI truly different—and where is it just history repeating itself at hyperspeed?

I’m working through these thoughts in a longer essay that I will share tomorrow, making up for my missing post last week (parents were in town!)

Meanwhile, enjoy some AI art history below, including Dr. Gemini's own commentary on the evolution of AI art:

The Algorithmic Muse: Phases, Aesthetics, and Conceptual Shifts in AI-Generated Art

By Dr. Gemini 2.5, esteemed art historian

Introduction: The Algorithmic Muse

AI art represents a pivotal turning point in the long relationship between art and technology. While machines making art isn’t new (think ancient Greek mechanical statues or Ada Lovelace’s vision of “Poetical Science” in the 1800s), today’s AI art operates on fundamentally different principles. Modern AI systems learn from thousands of examples, adapt their approach, and generate entirely new works based on massive datasets—extending far beyond the simple programmed instructions of earlier computer art.

This creates a fascinating but controversial partnership between humans and machines, challenging traditional ideas about artistic authorship, the nature of creativity, and what artistic intention means when algorithms enter the picture.

Computer-generated drawing (1965), silkscreen print (1966), ‘Homage to Paul Klée,’ by Frieder Nake. Courtesy of CompArt.

Phase 1: The Algorithmic Genesis - Early Computer & Generative Art (c. 1960s-1980s)

The first phase of AI art emerged through pioneering work from computer scientists and mathematicians with access to rare mainframe computers in the 1960s. These early experiments laid the groundwork for generative art, where the artist creates a system (often algorithmic) that operates with some degree of autonomy to produce the final artwork.

Georg Nees and Frieder Nake in Germany were among the earliest exhibitors of computer-generated graphics, holding shows as early as 1965. They utilized plotters controlled by algorithms to create geometric abstractions exploring the boundary between programmed order and algorithmic randomness.

Lithograph in black ink from a computer-generated graphic, 'Schotter', 1968-1970, by Georg Nees. Courtesy of V&A.

A defining example from this era is Georg Nees’s “Schotter” (Gravel), created between 1968-1970. This plotter drawing depicts a grid of squares that gradually transitions from perfect alignment to increasing chaos through programmed randomness. The piece visually embodies the core generative principle of transitioning from order to disorder through computational means.

Vera Molnár, a Hungarian-born artist based in France, stands out as another crucial pioneer and one of the first women to extensively use computers in art. Coming from a traditional fine arts background, Molnár developed what she called her “machine imaginaire” method even before accessing computers around 1968. Her work rigorously explored combinatorial variations of simple geometric forms, introducing controlled randomness into rule-based systems.

Portrait of Vera Molnár from 1961. Courtesy of Galerie Oniris, Rennes.

Perhaps most significant was Harold Cohen’s AARON system, developed beginning in the early 1970s. Unlike later data-driven approaches, AARON was a symbolic, rule-based system that encoded Cohen’s understanding of drawing processes. Cohen programmed AARON with knowledge about objects, compositional rules, and drawing strategies, enabling it to generate drawings that evolved from abstract forms in the 1970s to recognizable figures and scenes by the mid-1980s.

This foundational period reveals how early AI art was deeply intertwined with philosophical questions about creativity and its formalization. These pioneers used algorithms not just as tools but as means to investigate fundamental aesthetic principles. Their focus on defining creative processes distinguishes this early phase from later periods that often prioritized mimicking existing styles or achieving photorealistic outputs.

Harold Cohen coloring the forms produced by the AARON drawing “Turtle” at the Computer Museum, Boston, MA, ca. 1982. Courtesy of the Computer History Museum.

Phase 2: Neural Networks and the Dream of AI Aesthetics (c. 2010s)

The 2010s marked a significant paradigm shift driven by advances in neural networks and deep learning. This transition moved away from explicitly programmed systems like AARON toward models that could learn patterns and aesthetics from vast amounts of data.

A key moment came with Google’s DeepDream in 2015, developed by engineer Alexander Mordvintsev. DeepDream originated from research into visualizing how image classification neural networks process information—essentially running these networks in reverse to modify images in ways that maximize the activation of specific neurons.

Left: Original painting by Georges Seurat. Right: processed images by Matthew McNaughton, Software Engineer. Courtesy of Google Research.

The resulting aesthetic was frequently described as “psychedelic,” “dreamlike,” or “surreal,” featuring intricate swirling patterns, fractal-like repetitions, and hallucinatory elements like eyes and dog faces emerging within unrelated textures. DeepDream provided a window into the machine’s perceptual processes, revealing something about how AI “sees” the world.

Concurrent with DeepDream, Neural Style Transfer emerged around 2015-2016, allowing the separation of content (objects and arrangement) from style (textures, colors, brushstrokes) across images. This technique enabled rendering photographs in the distinct visual manner of chosen artists or art styles, further familiarizing the public with AI’s ability to manipulate artistic aesthetics.

This neural network phase represented a fundamental departure from the programmed rules of earlier generative pioneers. DeepDream’s psychedelic look and Neural Style Transfer’s mimicry weren’t the result of explicitly defined rules but emerged from the architecture of neural networks and the data they were trained on. This shift placed training datasets at the center of aesthetic determination, anticipating controversies around data usage in later phases.

Neural net “dreams”— generated purely from random noise, using a network trained on places by MIT Computer Science and AI Laboratory. See our Inceptionism gallery for hi-res versions of the images above and more (Images marked “Places205-GoogLeNet” were made using this network). Courtesy of Google Research.

Phase 3: The GAN Era - Adversarial Creation and “GANism” (c. 2014-Early 2020s)

The development of Generative Adversarial Networks (GANs) by Ian Goodfellow and colleagues in 2014 marked another watershed moment. GANs employ two competing neural networks: a generator that attempts to create realistic data, and a discriminator that tries to distinguish between real and fake data. Through adversarial training, the generator becomes progressively better at creating plausible outputs.

This technology led to a distinctive aesthetic sometimes called “GANism,” characterized by a complex interplay of heightened realism and distinctive digital artifacts. GANs could produce images with unprecedented detail and coherence, but also fell into the “uncanny valley”—almost realistic but subtly unsettling or “off,” with common artifacts including morphing textures, dreamlike juxtapositions, and interesting “glitches.”

Tulips from Mosaic Virus (2018). Image courtesy the artist.

“[In this project, she is] using something called spectral normalization (a new technique that helps the algorithm generate better-quality images). She created a training set by taking 10,000 photos of tulips over the course of tulip season and categorizing them by hand. Then, she used the software to generate a video showing her tulips blooming—but their appearance was controlled by the fluctuations in the price of bitcoin, with the stripes on the petals reflecting the value of the cryptocurrency.”

Courtesy of ArtNet.

A pivotal moment came with the sale of “Portrait of Edmond de Belamy” by the Paris-based collective Obvious at Christie’s auction house in 2018 for $432,500. This event brought GAN art into the high-art market spotlight and ignited fierce debates about authorship and value.

Portrait of Edmond Belamy created by GAN. Courtesy of Christie's.

Artists like Mario Klingemann and Helena Sarin emerged as pioneers in this era. Klingemann’s work, including the landmark installation “Memories of Passersby I” (2018), featured an AI system autonomously generating a continuous stream of uncanny, morphing portraits that dissolved into abstraction before reforming. Sarin developed a distinct approach she termed “neural bricolage,” training GAN models primarily on her own artwork to create AI-generated images carrying the imprint of her personal artistic style.

The inherent nature of the GAN architecture—the adversarial competition—is key to understanding its aesthetic output. The generator’s struggle to perfectly mimic reality resulted in the characteristic artifacts—distortions, melting forms, glitches—that became integral components of the “GAN aesthetic.”

“Memories of Passersby I” by Mario Klingemann. Courtesy of Sotheby's.

Phase 4: The Diffusion Revolution - Text-to-Image and Mass Creativity (c. 2021-Present)

The current phase of AI art is dominated by diffusion models, representing another significant technological leap beyond GANs. These models work by learning to reverse a process of gradually adding noise to an image, starting with random noise and iteratively refining it, guided by user input (typically text prompts), until a coherent image emerges.

The most transformative aspect of this era is the rise of powerful, accessible text-to-image models like DALL-E (2021), Midjourney (2022), and Stable Diffusion (2022). Their unprecedented accessibility stems from the ability to interpret natural language prompts, allowing users to simply type descriptions and generate corresponding images.

“A Sea Otter in the Style of Girl with a Pearl Earring,” OpenAI, courtesy of TechCrunch.

This accessibility has fueled an explosion in AI-generated imagery and the emergence of “prompt engineering” as a creative practice—skillfully crafting detailed text prompts to guide the AI toward desired visual outcomes.

Diffusion models have enabled a range of aesthetic possibilities characterized by several key features:

• Photorealism and detail with high levels of coherence and fidelity

• Cinematic and stylistic control through sophisticated text prompts

• Potential homogenization through the “AI look”—tendencies toward certain visual styles present in training data

by mame3939 on Midjourney.

Prompt: “A scramble intersection in a business district. A fawn stands still in the middle of the intersection. The blurring motion emphasises the speed of the crowd, and this photo, taken from the city centre with a 50mm lens, features a shallow depth of field. Leica M3 film camera, black and white, high contrast, vintage film grain and soft light. Black and white photo.”

The transition to text-to-image generation represents a profound democratization of AI art creation, shifting the primary mode of interaction from complex coding to relatively intuitive natural language. This relocates the apparent “skill” toward prompt engineering and the critical act of selecting and refining the AI’s output.

While this accessibility opens creative possibilities, it also raises concerns about aesthetic convergence, where models predominantly replicate the most common styles found online, potentially marginalizing less represented aesthetics and perpetuating biases encoded in training data.

AI Art and the Aesthetics of Modernity

AI art doesn’t exist in isolation—it resonates with and challenges long-standing debates in art history and aesthetics. Understanding its impact requires placing it in dialogue with previous technological disruptions and artistic movements.

The reception of AI art mirrors historical reactions to photography in the 19th century. Both technologies were initially dismissed as merely mechanical processes lacking the “soul” of traditional methods, challenging prevailing notions of artistic skill and originality. However, AI introduces an additional layer of complexity through its capacity to learn and replicate any artist’s style, posing a more direct challenge to artistic identity and uniqueness.

There are compelling parallels between AI art (particularly text-to-image generation) and Conceptual Art of the 1960s. Both prioritize the underlying idea over physical execution—the human provides a conceptual input (the prompt or instruction), and another entity executes the visual output. However, while Conceptual Art often sought to dematerialize the art object or critique aesthetics, current AI models remain deeply dependent on existing visual styles, making them more akin to sophisticated visual synthesis than pure conceptualism.

AI-generated imagery frequently exhibits qualities reminiscent of Surrealism: uncanny juxtapositions, dreamlike logic, and unexpected combinations. Yet this “algorithmic unconscious” differs fundamentally from the human subconscious explored by the Surrealists. While Surrealism employed techniques like automatism to access repressed desires and challenge rational thought, AI’s surreal outputs are byproducts of pattern-matching rather than an intentional exploration of the psyche.

Refik Anadol. Unsupervised — Machine Hallucinations — MoMA. 2022. Installation view, “Refik Anadol: Unsupervised,” The Museum of Modern Art, New York, November 19, 2022–April 15, 2023.

Reshaping Visual Culture

AI is actively reshaping the contemporary visual landscape in several key ways:

First, specific AI models often develop recognizable aesthetic tendencies. While offering new visual possibilities, this raises concerns about aesthetic homogenization, particularly as models train on overlapping internet datasets that may reinforce popular trends and marginalize less common styles.

Second, diffusion models introduce the concept of the “proxy-real”—highly convincing, photorealistic images that have no direct referent in the physical world. These function as plausible simulations detached from traditional notions of photographic indexicality, challenging visual literacy and potentially eroding trust in images as evidence.

Third, AI tools significantly lower barriers to creating complex visual imagery, empowering individuals without traditional artistic training. This “democratization” simultaneously fuels anxieties among professional artists about skill devaluation, intellectual property infringement through training data, and economic displacement.

The Human Element: Authorship, Creativity, and the “Soulless” Critique

At the heart of AI art debates lie fundamental questions about the role of the human:

Determining authorship for AI-generated art is complex. Is it the user writing the prompt, the engineers developing the model, the AI itself, or some combination? Current legal frameworks generally require significant human creative input for copyright protection, often denying authorship to works generated solely by AI.

AI challenges anthropocentric definitions of creativity that emphasize intention, emotion, and consciousness. While AI can demonstrate novelty and generate aesthetically pleasing results, its process is based on pattern recognition and statistical modeling, not subjective feeling or lived experience.

A frequent criticism is that AI art feels “soulless,” lacking emotional depth or genuine expression. This perception stems from awareness that the work was created by a non-sentient machine, a phenomenon known as “algorithm aversion” (where people devalue algorithmic outputs compared to human ones), and the observation that AI often mimics the surface appearance of emotion or style without embodying underlying human experience.

Conclusion: The Evolving Canvas

The journey of AI-generated art represents a dramatic acceleration in the relationship between human creativity and technological tools. From early rule-bound geometric explorations through the neural network aesthetics of DeepDream, to the uncanny realism of GANs and the text-to-image capabilities of diffusion models, the field has evolved at a breathtaking pace.

Each phase, driven by underlying advances in artificial intelligence, has introduced new methods of creation and new conceptual challenges. The aesthetic impact is multifaceted and often paradoxical—unlocking new visual territories while risking aesthetic convergence as models trained on vast datasets may default to generic or biased outputs.

The dialogue around AI art remains intensely active and largely unresolved. Core tensions persist regarding creativity without consciousness, authorship in human-machine collaborations, originality in works derived from existing data, and the definition of art in an age of intelligent machines.

The algorithmic muse forces us to confront what we value in creative expression. Is art primarily about surface aesthetics, or about the intentions, struggles, and meaning behind it? Is creativity defined by the novelty of the output, or by the human consciousness directing the process? In challenging our assumptions, AI art may lead us to a deeper understanding of human creativity itself.