An AI-generated artwork just won an art prize. Yup, you read that right. In a groundbreaking moment at the Colorado State Fair's Annual Art Competition last year, a novel entrant named Jason M. Allen swept first place in the digital category, but not with the traditional scrape of a brush or glide of a stylus. He created the winning piece “Théâtre D’opéra Spatial” purely with Midjourney, an artificial intelligence program that turns lines of text into hyper-realistic graphics. This marks one of the first times an AI-generated artwork has been laureled with an art award, at such a scale.
And of course, artists were not happy.
A strong backlash emerged from the artist community, accusing Allen of cheating, although his submission was made under the name “Jason M. Allen via Midjourney”, clearly disclosing from the onset the AI's role in the artwork’s creation. Some artists defended him, arguing that the use of Generative AI tools like Stable Diffusion and Midjourney is akin to utilizing Photoshop - human creativity is still pivotal to devising the right prompts for generating an award-winning piece. Yet, the majority voiced concerns over the noticeable disparity in effort required, contrasting the raw skill, time, and diligence traditionally needed to create art.
And what did Allen, the creator of “Théâtre D’opéra Spatial”, have to say about all of this? Well, it was simple, really.
“This isn’t going to stop. Art is dead, dude. It’s over. AI won. Humans lost”
Simple, but pretty terrifying. This victory not only stirred the waters of traditional artistry but hinted at a subtle undercurrent that would only emerge a year later; data poisoning, a tool that could potentially ripple through the calm surface of machine learning and Generative AI in the realm of art.
Generative AI Controversy in the Art Industry
These applications have left numerous human artists apprehensive about their own futures, and understandably so. Why would anyone pay for art, they ponder, when it can be autonomously generated? This has also sparked robust discussions regarding the ethics of AI-generated art, accompanied by opposition from individuals who argue that these applications essentially represent a high-tech variant of plagiarism.
It draws a historical parallel, reminding us of the disdain many painters had for the camera and digital tools like Photoshop and Procreate during their early days, viewed as threats to traditional artistry. Generative AI, however, is seen as a more widespread force, not confined to the art industry but extending its influence across all creative sectors. Whether it's scriptwriting, video editing, or beating writer's block, Generative AI presents a ready solution, thereby escalating fears of human obsolescence in creative pursuits.
The sentiment is underscored by Jeffrey Katzenberg, the co-founder of DreamWorks Animation, who remarked “I don’t think there is any industry that Generative AI will disrupt more than the creative industry.” And looking at how “Théâtre D’opéra Spatial” swept up its competition, it isn’t hard to see why lots of people share that sentiment.
Well, what now? Do we embrace our new AI overlords and accept the future as it is? Well, it wouldn’t be a dystopian AI movie if the humans didn’t fight back, would it?
So, fight back they did, and this time, the weapon of choice is poison.
Data Poisoning: Artists’ Counterstroke Against Generative AI
At the core of deep learning lies data points—a boundless ocean of information that burgeons with every ticking second, serving as a rich reservoir of knowledge for AI programs to tap into and better serve their mortal masters.
The idea is to use AI’s incredibly efficient way of machine learning against itself. The data poisoning tool, aptly called Nightshade, lets artists add invisible changes to the pixels in their art before they upload it online; if said art is scraped into an AI training set, it can cause the resulting model to break in chaotic and unpredictable ways. Poisoning the machine learning data reservoir, or data well, if you will.
Nightshade exploits this security vulnerability in Generative AI, which uses artists’ work to train their models without the creator’s permission. Ben Zhao, a professor at the University of Chicago, led his team to create Nightshade and another data poisoning tool called Glaze; Glaze works similarly by changing pixels of images in subtle ways that are invisible to the human AI, but machine-learning models would process something drastically different.
Poisoned data samples can manipulate models that are the basis for machine learning, for example, identifying pictures of dogs as cats, furniture as trees, and cars as cows. Researchers tested the attack on Stable Diffusion’s latest models and on an AI model they trained from scratch, and the results were incredibly effective. When they introduced a mere 50 poisoned images of dogs to Stable Diffusion, the subsequent attempts to create dog images began to exhibit oddities. By the time they fed 300 poisoned samples into the system, the platform started generating images of dogs that resembled cats instead.
To add salt to the envenomed wound, Generative AI models are excellent at making connections between words, which helps the poison spread. In the dog-to-cat example, Nightshade would also infect words of similar concepts, such as “puppy”, “husky” and “wolf”. An example below depicts how a poisoned model of “fantasy art” affected related prompts and outputs, such as turning a dragon and a castle in the Lord of the Rings into something completely unrelated.
Beyond the Canvas: The Larger Implications of a Daunting War
The story of Jason M. Allen's "Théâtre D’opéra Spatial" and the artists' fight back with tools like Nightshade and Glaze highlights a growing tension in the world of creativity. But looking closer, the attempt to retaliate against Generative AI through data poisoning seems like a mere ripple against a formidable tide of advancing AI technology. This brings up a crucial question: Is this fight enough?
The situation between the continuous growth of AI and artists' efforts to keep their place feels like an ongoing battle, much like an arms race. As artists come up with new ways like data poisoning to defend their work, AI, with its never-ending ability to learn, might be close to adapting and overcoming these defenses, making them useless over time.
The Nightshade saga is indeed a bold statement, a reflection of human ingenuity and the indomitable spirit to uphold the sanctity of creativity. Yet, it may merely be a precursor to a relentless, iterative battle, where each side ups the ante, but the scales may be tipping in favor of the machine and the big corporations that back them.
Independent efforts may offer a glimmer of hope in preserving the essence of artistic quality, yet given the current trajectory, the vast amount of data at Generative AI's disposal could overshadow these endeavors, leaving artists in a precarious position. Unlike past advancements that faced resistance, Generative AI represents a paradigm shift: no prior skills are needed. Essentially, it ushers in a reality where anyone can don the hat of an artist. But here lies the paradox: when everyone is an artist, nobody is an artist. If more robust measures aren't explored, the very vocation of an artist might slowly fade away into oblivion. At the rate things are going, it looks to be the case, and perhaps we will be resonating with Allen’s stark pronouncement.
“Art is dead, dude.”
Very captivating read Ryan, the notion of artists going to war with AI is especially compelling. This seems like an interesting evolution to cybersecurity, where humans now have to battle against the ever growing intelligence of AI. With how fast generative AI is progressing, could AI one day learn how to bypass this data poisoning solution?
Thanks for the interesting perspective on the evolving relationship between art and AI. The introduction of data poisoning as a countermeasure to Generative AI in the creative field is indeed a noteworthy development. It got me thinking about the potential for artists to creatively incorporate AI into their work, rather than resisting its influence. Overall, a thought-provoking piece that sheds light on the ongoing dynamic between human creativity and AI advancements.
Interesting read! I enjoyed how this article introduced data poisoning, an unexpected development in the field of AI and Machine Learning (ML), which further prompts thoughts into the upcoming regulations of data consent and the possible outlawing of or countermeasures against data poisoning.
In terms of art, this also makes one wonder if artists could find creative ways to utilize and integrate generative AI or ML into artistic messages, displays, or exhibits, rather than go against the grain of this natural human progression.
I’m enlightened!
Great read