Those who encounter Terence Broad’s AI-generated piece, (un)stable equilibrium, on YouTube might be led to believe it draws inspiration from the earlier works of acclaimed painter Mark Rothko. His images, characterized by vivid fields of color, evoke the essence of Rothko’s lighter pieces. However, unlike Rothko’s journey toward darker themes, Broad’s creations display a constantly evolving nature, altering in both form and hue.
Contrary to assumptions, Broad did not train his AI model using Rothko’s work—or any data, for that matter. By manipulating a neural network to lock certain components into a recursive loop, he achieved a system that generates images independently, without prior input or influence. This innovative approach can be interpreted as either a groundbreaking illustration of artificial creativity or simply a sophisticated electronic by-product akin to unstructured noise. Regardless, Broad’s work hints at potential avenues for the ethical application of generative AI, steering clear of the derivative content that currently permeates the visual arts landscape.
Broad has expressed significant concerns regarding the ethics surrounding the training of generative AI on existing artworks. His journey toward creating (un)stable equilibrium was propelled less by philosophical considerations and more by a frustrating job experience. In 2016, while searching for a machine learning position that neglected surveillance tasks, Broad ended up working for a company managing traffic surveillance cameras in Milton Keynes, prioritizing data privacy. He recalls, “My job involved training models and handling massive datasets, sometimes up to 150,000 images, from a rather mundane city. I grew weary of dataset management, and upon launching my art practice, I was adamant about avoiding it.”
Further discouragement arose following legal threats from a major corporation. One of his early projects involved training an autoencoder on every frame of the 1982 film Blade Runner to generate a copy of the film. This project, fragments of which remain accessible online, proved to be both an illustration of generative AI’s limits at that time and a cheeky commentary on human-generated intelligence. Broad gained significant attention after posting the video online, only to be met with a DMCA takedown notice from Warner Bros. Upon receiving such a notice, Broad discovered that contesting it could lead to lawsuits, a risk he, as a recent graduate carrying considerable debt, was unwilling to undertake.
The situation escalated when a journalist from Vox reached out to Warner Bros. for a comment, prompting the company to rescind the notice briefly before reissuing it rapidly. Broad mentioned that the video had been reposted multiple times, each time resulting in a fresh takedown notice, often generated by AI software. As curators began to approach him, he received invitations to exhibit his work at prestigious venues like the Whitney Museum, the Barbican, and Ars Electronica. Despite these opportunities, the legal uncertainties surrounding his work caused significant anxiety. Broad recounts feeling apprehensive during a gathering at the Whitney, nervously boarding a plane with the fear that Warner Bros. might shut down his exhibit. “Fortunately, I was never sued, but the experience marked me. I decided I wanted to create art without deriving from others’ work without their consent or without compensating them,” he said. Since that time, he has refrained from using any external data to train his generative AI models.
In pursuit of this goal, Broad began a PhD in computer science at Goldsmiths, University of London, in 2018. There, he delved into the profound implications of avoiding training data. “How could a generative AI model be trained without imitating any data? It took time to realize that was inherently contradictory. A generative model is inherently a statistical model that mimics the data it’s exposed to,” he explains. He subsequently shifted his focus to generative adversarial networks (GANs), which were prevalent at the time. In a typical GAN setup, two networks, the generator and the discriminator, work together, with the generator trying to create convincing fake data based on a dataset while the discriminator distinguishes between genuine and fabricated inputs. Over the course of this process, the interaction hopes to yield a model capable of generating outputs comparable to the original dataset.
Broad had a breakthrough when he conceived the idea of using a separate generator network in place of traditional training data, allowing for a feedback loop between the two. His initial attempts produced unsatisfactory “gray blobs,” lacking in visual interest. Upon adding a color variance loss term to the system, the resulting images became increasingly intricate and vibrant. Further innovations with the GAN’s components led to even richer results. “The input to a GAN is a latent vector—a compilation of numerical data. You can smoothly navigate through various points in the generation space, influencing the output. The capability for infinite generation of new imagery is truly fascinating,” he elaborates.
Upon reviewing his early outputs, Broad noted parallels to Rothko’s style and labeled these initial images in a folder aptly titled “Rothko-esque.” Although audience reactions varied, including accusations of dishonesty regarding his lack of input data, he emphasizes the core of his work lies in the generative process itself, not merely in the resulting images. His aim was to explore the latent creative potential of the networks he employed more than to replicate any specific artistic style.
Is Broad convinced he has successfully tapped into “pure” artificial creativity? He responds with some uncertainty, stating, “No external forms or features are necessarily imposed on the outputs of the networks, yet I speculate my personal aesthetic might influence the process subtly. The reasons behind the outputs remain somewhat enigmatic, and while I have received advice to investigate these results further, I’m content to embrace the mystery.”
Engaging with Broad about his creative process, along with insights from his PhD research, highlights a crucial takeaway: the complexities of understanding generative AI, even at an academic level, remain a significant challenge. Unlike tools such as Midjourney, which prioritize “prompt engineering,” Photoshop offers users the ability to manipulate an extensive array of options. While generative AI processes can yield outcomes based on provided data, the intricate mechanics at play within the algorithms often remain obscure. Broad points out the contradiction present in an organization like OpenAI exhibiting a degree of secrecy about its models and training data.
Broad’s investigation into output-generated processes provides insights into AI functionality, although at times it resembles the early, crude explorations of the brain rather than refined psychoanalytical techniques. By demystifying these models, Broad advocates for a clearer understanding of their capabilities at a time when exaggerated perceptions of AI’s power dominate discussions. “There is a tendency to think that AI performs beyond its actual scope,” he suggests. “In reality, it involves merely a series of matrix calculations, and engaging with these systems is far more accessible than many believe.”
