When Linear Algebra Makes You Cry
The biggest shock of the AI revolution has been that it came from behind. We were sure that vision, robotics, and logic were easy, but creativity was so mysterious that it was almost unprogrammable. We thought that the first casualties of the revolution would be the low-skill, repetitive workers, and right at the back after everyone else had fallen - there would be the artists. But that isn’t how it went. The AI revolution came from behind and routed the artists first. A real Hannibal move.
The illustrators, photographers, painters, and concept artists were the very first to confront AI when generators like Midjourney showed how quickly a prompt could go from idea, through iterations to finished work. The writers were next when the monumental chatGPT was released and showed that written work could be rendered in perfect English just as quickly as an image generator could paint a mountain range.
AI generated work which won first prize in the Colorado State Fair, 2022, Paul Allen
Fear, anger, and disgust abound in the art community.
Fear that commercial art roles would close up through corporate cost-optimization, and that commissioned or even hobbyist artists would be swarmed into insignificance by a torrent of cheap, easy, abundant AI art. Also, though it’s not often said out loud, there was/is a fear that the machines would do it better than us.
Anger that the generators were trained on human art without the consent of the artists and that the humanities were being stolen.
Disgust that the humanities were being infringed on by faceless machines built by soulless technocrats.
Though the dust has far from settled, the art world has had ample time to digest and get familiar with generative AI tools. It’s notable that through familiarity fear seems to have waned, but anger, disgust, and even appreciation have grown. In this article, I want to discuss AI’s limitations, and where humanity still retains a place in a world dominated by creative machines.
Amy Winehouse and the AI masterpiece
In 2007 Amy Winehouse working with Mark Ronson released a cover of The Zutons’ Valerie. It was so definitive and resonant that it swept the original into obscurity. Most people who know the song have no idea that Amy didn’t know a Valerie, that she didn’t look across the water, or wonder about that fine that she was dodging all the time. The art was penned by someone else. Amy animated the song, but the emotion, energy, and calculations that wove the words were someone else's.
A candid photo of Amy Winehouse, 2003, Photo taken by Charles Moriarty
Think for a second how your feelings towards Amy’s version would change if you found out that The Zutons didn’t write it, but a massive piece of linear algebra called chatGPT did. Personally, I’d feel mostly the same, because the emotion came from Amy. When she cries “stop making a fool out of me”, barking the ‘stop’ and ‘fool’ and then modulating ‘me’ in a drawn-out cry - that is Amy's contribution in a nutshell. When Amy sang Valerie, The Zutons’ contribution was practically irrelevant.
Amy didn’t know the Valerie in the song, but she knew a Valerie from her own life, and that was enough. The Zutons didn’t need to know a Valerie either. Why couldn’t AI have written it?
Amy was an artist, and like any artist she was loved for her unique perspective on life. Her challenges and triumphs in this world shaped her over time. Her art formed a continuum showing her evolving perspective, and ultimately her tragic deterioration.
The woman herself and all of her associated stories, gossip, history, and art were brought into every piece she worked on. This is something that AI cannot ever truly equal. AI can make art, but can never be an artist.
It’s easy to imagine that AI might strike upon a super resonant metaphor, a catchy hook, or even sum up a complex feeling in a tight verse. It’s also possible that AI will craft a compelling narrative, or write a stunningly true piece of dialogue. But if the art is delivered through human performance, that becomes central to the enjoyment of the piece, and the supporting contributions like writing become less relevant.
The same is true for art that is not performed but where the artist him or herself is a powerful draw. Think Margaret Atwood, Gerhard Richter, Paul Thomas Anderson. The personality, aura, and continuum of work is the draw; we’re along for the ride. Never underestimate the power of cult over the human mind.
Yes, Paul Thomas Anderson is great, but he’s in his 50s. What about the director-to-be who is still in middle school, watching AI get more creative by the day? How will she build her reputation?
Breaking the mould
Humans will always break the mould.
Most of the time we can tell when some text or image was rendered by a generative model. This is for two main reasons:
AIs are “tuned”, which is to say they’re moulded and shaped into a suitable creative profile. ChatGPT writes the way it does because of its tuning. The tuning is what gives generative models their “personality”, “voice”, or “creative biases.”
They’re increasingly trained on their own AI-generated work, creating an incestuous feedback cycle which reinforces its own creative bias.
This is very important to understand. When tuning a generative model they are usually guided by a consensus over what “quality” work looks like. Do you see the issue? They may pick multiple examples of quality, for example, quality prose, quality poetry, and quality article writing, but these are still biases, frozen in time, that will not update until it's tuned again.
AI generated image, 2023, Courtesy of MemoryMod
To see why this is an issue, let’s explore the following scenario:
Let’s imagine a world where AI is dominating the Advertising Poster industry. What we’re imagining is a world where everyone is effectively soliciting art from the same few artists who each have their own ideas about what quality looks like. Posters from New Delhi to Miami would have a similar look and feel, and everyone would notice it.
The first people to try to tackle this would be the generative model engineers who would hastily adjust the tuning or maybe even undo some tuning altogether to broaden its creative ability, but this wouldn't suffice.
Changing the tuning just superficially modifies its creative biases, it doesn’t get rid of it. And removing tuning altogether lessens its creative bias, but just makes it tougher to control. If the model isn’t biased towards quality it will start producing the opposite of quality. For example, it might use haiku to create a eulogy.
The next set of people who would want to tackle this imagined issue would be the advertisers who would bring on a human artist to take the AI’s results and modify them in a new creative direction. They would do this because of a person’s inherent ability to be boundlessly creative while retaining the appeal of their work. A person can write a solemn eulogy a million different ways and will have the wisdom to not attempt a haiku.
AI generated image, 2023, Courtesy of MemoryMod
In order for a new poster design paradigm to breakthrough, a human is needed. This is true at the bleeding edge of all art.
The AI’s creative scope is limited to its medium, whereas a human can draw from an almost infinitely faceted experience for inspiration. We also inherently understand our own culture, so we can create within the bounds of taste and broad appeal - we have a sense for these things. We can be incredibly creative and out-of-the-box in our output but still appeal to humanity. This is something that AI will struggle with, perhaps forever.
