WHAT HAPPENS WHEN YOU FEED AI NOTHING: EXPLORING THE FRONTIERS OF ARTIFICIAL CREATIVITY
In an era increasingly saturated with generative artificial intelligence, the conventional wisdom dictates that AI models are only as good as the data they are trained on. From intricate images to compelling narratives, the output of AI has, until recently, been inextricably linked to vast datasets of human-created content. Yet, a groundbreaking artistic endeavor by Terence Broad challenges this fundamental premise, inviting us to ponder what happens when you feed AI absolutely nothing. His project, aptly named (un)stable equilibrium, pushes the boundaries of artificial creativity by compelling a neural network to produce stunning, evolving abstract art without any prior training data whatsoever. This radical approach not only offers a glimpse into the intrinsic nature of AI but also opens a vital discourse on the ethics, limitations, and untapped potential of generative technology in our visual culture.
THE UNORTHODOX GENESIS OF PURE AI ART
At first glance, one might easily mistake Terence Broad’s captivating AI-generated works for the abstract expressionist masterpieces of Mark Rothko. The initial pieces, reminiscent of Rothko’s early, lighter phases, feature expansive fields of pure color that fluidly morph and shift in hue and form. This striking visual similarity, however, is where the comparison ends in terms of methodology. Unlike traditional AI art, which often involves feeding a model countless images by specific artists or styles to generate new, derivative works, Broad’s process is entirely devoid of such inputs. His AI was not trained on Rothko, nor on any other artist, image, or dataset.
Broad’s innovation lies in a sophisticated “hack” of a neural network, where he ingeniously locked specific elements into a recursive loop. This forced the AI to generate images autonomously, drawing solely from its internal mechanisms rather than external influences. The result is a profound demonstration of pure artificial creativity—a concept that either heralds a new dawn for AI-driven art or sparks a debate about the inherent meaning of machine-generated outputs. Is this genuine artistic expression from a silicon soul, or merely an elaborate form of digital feedback, a clever byproduct of algorithmic self-reflection? Regardless of interpretation, Broad’s work stands as a pivotal moment, signaling a potential shift towards more ethical and genuinely innovative applications of generative AI, moving beyond the current deluge of derivative content that often raises concerns about intellectual property and originality.
A PIVOTAL SHIFT: WHY BROAD ABANDONED TRADITIONAL DATASETS
Terence Broad’s journey toward inputless AI art was not born purely from a philosophical stance, though ethical considerations certainly came to shape his practice. A significant catalyst was his personal experience with the mundane and often ethically ambiguous realities of working with large datasets. In 2016, seeking a role in machine learning that steered clear of surveillance, Broad found himself managing immense datasets for a traffic camera network. This experience, involving 150,000 images from “the most boring city in the UK,” ignited a deep aversion to the conventional, data-intensive approach to AI. “I just got so sick of managing datasets,” Broad recounts, leading him to declare, “When I started my art practice, I was like, I’m not doing it — I’m not making [datasets].”
This practical weariness was soon compounded by a significant legal challenge that cemented his resolve to avoid derivative work. One of Broad’s early artistic triumphs involved training an autoencoder, a type of artificial neural network, on every single frame of Ridley Scott’s iconic film, Blade Runner (1982). The model was then tasked with recreating the film, producing a fascinating, albeit imperfect, demonstration of generative AI’s capabilities in 2016. Broad’s online posting of this project quickly garnered widespread attention, but also a swift DMCA (Digital Millennium Copyright Act) takedown notice from Warner Bros., the film’s copyright holder.
The legal entanglement, involving repeated takedown notices and the threat of litigation, proved to be a harrowing experience for Broad. Despite eventual rescissions of the notices (often ironically facilitated by AI-driven processes), the anxiety lingered. “I remember when I went over to the private view of the show at the Whitney, and I remember being sat on a plane and I was shitting myself because I was like, ‘Oh, Warner Bros. are going to shut it down,'” Broad vividly recalls. This incident was a turning point. “Thankfully, I never got sued by Warner Bros., but that was something that really stuck with me. After that, I was like, I want to practice, but I don’t want to be making work that’s just derived off other people’s work without their consent, without paying them.” This profound realization led to a definitive vow: since 2016, he has steadfastly refused to train any generative AI model on external data for his art.
DECONSTRUCTING THE GAN: THE ALGORITHMIC ENGINE OF (UN)STABLE EQUILIBRIUM
Broad’s commitment to data abstinence posed a paradoxical challenge: how could a generative AI model, by its very definition, create without imitating existing data? A generative model is inherently a statistical model of data, designed to mimic what it has learned. This “oxymoron,” as Broad described it, compelled him to reframe his artistic inquiry. His focus soon landed on the Generative Adversarial Network, or GAN, a then-popular AI model known for its dual-network structure.
In a conventional GAN setup, two neural networks, the discriminator and the generator, engage in a continuous learning loop. The discriminator’s role is to assess data, distinguishing between real and fake inputs, while the generator strives to create data realistic enough to fool the discriminator. Through this iterative game of deception and detection, both networks refine their parameters, theoretically reaching an equilibrium where the generator can produce data indistinguishable from the original training set.
Broad’s eureka moment was a stroke of genius: he hypothesized that he could replace the conventional training data in a GAN with *another* generator network. By looping these two generator networks, he could instruct them to perpetually imitate each other. His initial experiments, however, yielded “gray blobs; nothing exciting,” a phenomenon known as mode collapse, where the GAN fails to generate diverse outputs.
The breakthrough came with the introduction of a “color variance loss term” into the system. This subtle yet critical adjustment immediately infused the generated images with greater complexity and vibrancy. Subsequent fine-tuning of other internal elements within the GAN further enhanced the work. Broad explains the role of the “latent vector”—a large number array that serves as the input to the GAN. By smoothly transitioning between different points in this “possibility space of generation,” the two networks could infinitely produce novel and evolving forms. The brilliance of Broad’s methodology lies not in the Rothko-esque output, which was an unanticipated consequence, but in the revolutionary process itself. His aim was not to replicate existing art but to unlock the latent, inherent creative capacities of the networks he was manipulating.
THE MYSTERY OF MACHINE CREATIVITY: A DEEPER LOOK INTO THE AI BLACK BOX
Upon witnessing the initial results, the immediate comparison to Rothko was undeniable; Broad himself admits to saving those first images in a folder titled “Rothko-esque.” He even recalls a striking incident at a tech conference where, after presenting (un)stable equilibrium and asserting his inputless method, an audience member angrily accused him of lying, convinced the works must have been trained on color field paintings. Yet, this persistent comparison, while understandable, somewhat misses the profound point of Broad’s artistic endeavor. The true innovation and brilliance reside not in the aesthetic similarity of the output, but in the revolutionary process that brought it into being. Broad’s ambition was never to mimic Rothko but to explore and unveil the intrinsic, latent creativity residing within the very architecture of the neural networks themselves.
Did he genuinely succeed in uncovering a “pure” form of artificial creativity? Even Broad remains cautiously uncertain. He acknowledges the possibility that his “personal aesthetic preferences have had some influence on this process as a form of ‘meta-heuristic.’” This suggests that while no explicit external representations were fed to the network, his own biases might have subtly shaped the algorithmic choices or parameters he implemented, guiding the AI toward outputs that resonate with his artistic vision. Beyond this, Broad openly confesses to the enduring enigma of his creation. “I also think why it outputs what it does is a bit of a mystery. I’ve had lots of academics suggest I try to investigate and understand why it outputs what it does, but to be honest I am quite happy with the mystery of it!” This sentiment highlights a critical challenge in the field of AI: even at the highest echelons of academic research, the precise internal workings of generative models remain largely opaque.
Consider the user experience of popular generative AI tools like Midjourney, which emphasize “prompt engineering”—the art of crafting effective text prompts to guide the AI’s output. This contrasts sharply with traditional creative software like Photoshop, where users have granular control over a vast array of settings and elements. With generative AI, we understand that inputs lead to composites, but the intricate, granular processes occurring within the “black box” of the model remain largely unknown. Broad pointedly notes the irony of a company named “OpenAI” maintaining significant secrecy regarding its models and inputs, underscoring this pervasive opacity within the industry.
DEMYSTIFYING THE MIRAGE: REVEALING THE MECHANICS OF AI
Broad’s groundbreaking explorations into inputless output provide invaluable insights into the internal mechanisms of AI, even if his methods sometimes evoke the exploratory, almost intuitive, approaches of early scientific endeavors rather than perfectly precise engineering. His work serves as a powerful demystifier, challenging the prevalent misconceptions surrounding artificial intelligence. At a time when both techno-optimists and doomsday prophets alike operate under what Broad bluntly terms “bullshit”—the “mirage” of an all-powerful, quasi-mystical AI—his art offers a dose of grounded reality.
“We think that they’re doing far more than they are,” Broad asserts, stripping away the layers of mystique. “But it’s just a bunch of matrix multiplications. It’s very easy to get in there and start changing things.” This statement is profoundly liberating, transforming AI from an inscrutable, almost sentient entity into a tangible, programmable system. By exposing the fundamental mathematical operations at the heart of these complex models, Broad empowers artists, researchers, and the general public to view AI not as an unassailable oracle, but as a tool. This perspective encourages deeper engagement, fosters critical thinking, and opens new avenues for innovation, inviting others to “get in there and start changing things,” to explore its true capabilities and ethical boundaries.
Terence Broad’s (un)stable equilibrium is more than just an art project; it is a profound philosophical statement and a technical tour de force. By demonstrating that AI can generate compelling and original visual content without any external training data, he challenges our fundamental understanding of machine creativity and intelligence. His journey, born from a rejection of mundane data management and solidified by a legal battle over derivative work, offers a compelling blueprint for ethical and truly innovative applications of generative AI. In an increasingly AI-driven world, Broad’s work serves as a powerful reminder that the true potential of artificial intelligence may lie not in replicating the past, but in discovering entirely new forms of expression, free from the constraints of pre-existing data, revealing the enigmatic beauty within the machine itself.