One day, sometime in the future, I may no longer have any reason to be a writer. One day, sometime in the future, machine learning may become so advanced that creative endeavors like writing and art are no longer a human domain. One day, sometime in the future, my ability to write may be considered nothing more than a niche commodity. But I’m okay with that. Because for all the fuss about how AI is the doomsday of the human creative mind, they’re also the next frontier in progress. And progress should not be stopped - so neither should AI.
AI, or more accurately generative machine learning, has been an area of intense study for several decades. The statistical methods we use today were discovered a century ago, and simple algorithms were devised as early as the 1950s. Bayesian methods, which many now use to create simple image generation algorithms, were first used for machine learning in the 1960s, and though the 1970s brought about a lapse in development, a new method called “back propagation” put new life into the field during the 1980s. As computers grew more powerful at the turn of the millennium, AI reached near-perceived sentience, and the last two decades introduced deep learning, a widespread buzzword among startups these days, and neural networks that could learn by themselves. All this to say, the train that is AI is not something that can be stopped. It may experience a frigid winter, but it will never lose all momentum.
But to slow progress down, to rob it of its inertia, is not just to prevent innovation. It’s to cause greater future harm to the environment and to those that machine learning is hurting now. Because right now, we have the speed to overcome the great obstacles we face. Urgency in passing regulation around AI has already begun pushing lawmakers across the globe to address artists’ concerns about the technology, and ethical and environmental concerns, along with their associated costs, have driven those behind the latest machine learning algorithms to create better methods, more powerful computers, and move away from using unpaid or coerced labor. If we slow down now, we’ll have to pay the price of more time when we speed up again - is it not better to rip off the bandage than to slowly peel it back for fear of a minute of pain?
Another of the greatest concerns about the technology is copyright. Many of the models that populate the internet today are trained on unlicensed works created by artists and authors who are unaware of, nor did they consent to, the use of their pieces. But this is a problem easily solved with proper, ethical licensing and legal regulation. Laws that would limit the data machine learning could be trained on, and licensing agreements that would keep creators both aware and compensated for the use of their work. There’s already such a litany of laws regulating data - though perhaps lacking in the US, the EU and other nations have already proven that regulating the management of information is very much possible - that adding a few bills about disclosure and compensation in generative machine learning would be near trivial.
Significantly, some of the greatest concerns about generative algorithms, particularly visual algorithms like Midjourney, Stable Diffusion, and DALL-E, stem from the common misconception that these models simply stitch together parts of existing art to create their outputs. This is an easy mistake to make - without knowing the intricacies of machine learning, the simplest explanation for how a model can mimic a style or produce normal-looking art is that they’re not; that they’re instead simply spitting out what they got it with a little bit of photo editing in the middle.
However, one must note that the size of these models is nowhere near the size of their collective databases. A few gigabytes is not enough to store the millions of images a neural network is trained on. In fact, because Stable Diffusion and other models are open source, or based on open source technology - that is, their code and algorithms are public facing - we can easily examine how they work. Diffusion models are trained to take distorted images, and restore them. These images become more and more distorted the more these models are trained, so they become better at creating details that they aren’t given. After training, to generate a new image, these models are simply provided with a random array of pixels, then told to restore the image. This random array is truly random; it came from nothing. So, a model then “restores” an image that never existed in the first place, and so we end up with a new piece of art. Other models like DALL-E are proprietary, but their core algorithms are still the result of decades of peer-reviewed, scientific pursuits.
And there’s one more misconception to be addressed: artificial intelligence is not, primarily, artistic. Though the most public-facing models are image and text generators, the most powerful, next-gen machine learning models are used in academic research to advance all sorts of virtuous aims, including those who are trying to cure cancer, research rare diseases, combat world hunger, solve poverty, and fix the housing crisis. Advanced deep learning can help scientists and their analytical teams look at and generalize previously unconquerable datasets, and derive conclusions and solutions that represent the next step in making our world a better place. That’s innovation that saves lives, changes futures. There’s no excuse or reason more important than advancing society, and to argue for the end of such important research is entitled and selfish.
I am, first and foremost, not a writer. I am a developer who writes code as a hobby. Coding is truly a passion, a way to wile away the empty hours of the day. I’m also an artist. I model 3d characters and objects to gaze at and to populate the worlds I create while making games. I do not believe that my passions will go extinct. I do not believe that human artisans will lose their niche in society. When factories replaced blacksmiths, hand-crafted tools didn’t go out of fashion. They became generational heirlooms, special treasures, beautiful pieces of art. When watches became available to all, those made in the traditional, time-consuming style became symbols of luxury and opulence. When 3d modeling made animation accessible and available to everyone, stop motion and hand-drawn animations became artistic styles that allowed imperfections to be seen as charming.
So I do not believe that the progress of machine learning should be stopped. I do not believe it can be stopped. But even if one day, being a writer, or an artist, or a programmer means being outdone by a computer, I don’t mind. I watch as machine learning becomes more advanced and closer and closer to becoming indistinguishable from a human, and I feel as if I’m watching the lunar landing. As if I’ve somehow lucked into watching humanity tackle the next big thing. So one day, sometime in the future, I may go extinct. But I look forward to that day - and everyone else should too.
Denoising Diffusion Probabilistic Models. (n.d.). arXiv.org. https://arxiv.org/abs/2006.11239
Generative AI Has an Intellectual Property Problem. (n.d.). Harvard Business Review. https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem