Is This Art?

The recent explosion in AI-generated original work points to some larger questions—about art, originality and what it means to be human.

From the spring 2023 issue of the Sacred Heart University Magazine

by Cliff Clive

"We’re five years away from watching movies and TV series written entirely by artificial intelligence.”

That’s not my prediction. This claim came, very confidently, from a former co-worker of mine, a fellow data scientist at Microsoft, eight years ago.

He was off by a bit.

For decades we’ve seen industries taken over by technology and automation, rendering some workers redundant while empowering others to do more with less effort and in safer conditions. In either case, it’s typically the menial tasks that we hand over to robots and computers. Automating creativity—writing, painting, music—has been mostly dismissed as the domain of science fiction.

Now multiple new technologies are forcing us to reconsider that assumption. ChatGPT is generating cocktail recipes and scoring very well on the GMAT. Midjourney AI is printing out stills from a Lovecraftian horror movie starring a cast of Muppets—a movie that was never made, of course.

Artificial intelligence (AI) has racked up an astonishing number of big wins in the past year. Whether or not ChatGPT will ever win a Pulitzer is up for debate, but the point is: it’s up for debate—a debate that many people now take seriously. And while it often seems that the only way to spot AI-generated images is by counting fingers and teeth, that’s likely to be sorted out in another iteration or two of the software—possibly even before this article goes to print. As the tweaks get tighter and the mistakes are fewer, are we approaching an AI-generated Renaissance?

Let’s stop and consider why that’s a terrible idea.

There are some obvious red flags. Generated essays and images are creating headaches for university professors and hiring directors around the world, and plenty of concern has been expressed for the artists whose work was scraped from the internet to build the technology that Silicon Valley will soon use to compete with them. But while the ethics in both instances are questionable at best, it’s only a matter of time before legal and technological solutions are found for these legal and technological problems. Nor should we lose sleep over the standard sci-fi fear that the rise of AI is the dawn of a robot apocalypse.

That said, in good science fiction, the robot apocalypse is never really about the robots. It’s about us. Their rise is an allegory for our fall. And in that light, if we allow ourselves to think of these AI-generated works as art, we’re already losing touch with what it means to be human.

A Peek Inside the Black Box

For present purposes, though these arguments apply to any artistic medium, let’s focus on the visual arts and ask the question: If AI-generated imagery looks like art, surely isn’t that what it is?

To answer that, we need to begin with how the art is generated.

Higher math aside, the grand design of AI-generated work is surprisingly straightforward. If the algorithms seem harder to understand than the human creative process, it’s only a familiarity bias at play. Sure, the human mind feels like it’s more in your comfort zone—you are one, after all—but remember that AI machines are ones we built and programmed. They aren’t doing anything we haven’t taught them to do. They follow our instructions.

And those instructions, simply put, are these: study pixel arrays (which the humans call “pictures”) to find patterns; associate those patterns with established labels (which the humans call “words”); get good enough at this to make successful word/image associations with samples presented from outside the training set.

A model is shown millions of images. It guesses labels that have been assigned to them. The learning algorithm sharpens the model’s parameters after each correct guess and blurs them after each error. Eventually it collects a library of patterns it can identify and reliably label. For example, these two round shapes often appear above this triangular shape, positioned above a line, all within a larger oval shape. These are identified and labeled “eyes,” “nose” and “mouth,” and the full construct is labeled “face.” As the algorithm reads data from more and more images, the machine learns that the face pattern possesses slightly different qualities in images labeled “fantasy” from those labeled “renaissance.”

Flipping the calculations around, the program becomes a generator model that assembles a portmanteau of patterns to build images for new text prompts.

This generator is pitted against a discriminator, a separate program trained to detect machine-generated images in a lineup of photos or art from human artists, forcing the generator to improve its work until the discriminator can no longer tell the difference. Paired together, the generator and discriminator form a generative adversarial network (GAN).

In short, machines look at millions of labeled images to learn how to draw new ones in a similar style—which is how human artists learn too, isn’t it?

Well, up to a point.

Mastering the Craft

What all this means is that the images coming out of a GAN must look statistically similar to the ones the model was trained on; otherwise they won’t pass the discriminator’s test. In other words, AI is structurally incapable of innovation. All it can do—by the rules of its own design—is create pastiche.

It’s exceptionally good pastiche—AI creates pastiche virtually impossible to identify as different from the original work from which it learned—but it’s pastiche nonetheless.

And that’s a really important distinction.

Because, yes, art students take classes on such topics as form, composition, color and shading. They learn to work in a variety of mediums and study the work of a pantheon of great artists. They fill portfolios with practice pieces, training their muscle memory to match what they’ve seen with what they hope to create. This is analogous
to the patterns that AI neural nets learn and combine into new images.

But as they hone their craft, artists live and engage in the world around them in a way that the computers do not and cannot. Because while computers process data, which is objective, quantitative and structured, the human mind runs on meaning, which is subjective, qualitative, impressionistic. And living makes us subject to limitless impressions that we can plumb for meaning.

What this means in practice is that a piece of music that doesn’t even have words can somehow articulate the rapture of a first kiss for humans, but will never be fully appreciated by a computer. Because art is an abstract language. It doesn’t make literal sense. It doesn’t need to. That’s the point. So, while AI taps out at representing the correct number of fingers on a hand in the style of Rembrandt, an artist carries on, experimenting with ways to turn heartbreak and triumph and the full spectrum of the human experience into the physical reality of a painting or a sculpture or a script or a dance.

The skills to be mastered may be the same between the artist and the machine, but AI will only ever evaluate the data from an image it is creating through probability distributions it has drawn from similar images. It will only ever create an abstract painting by replicating the work of abstract artists, which by definition means it is not abstract. And neither is it art since the ultimate goal of the artist isn’t to replicate; it’s to create. Those skills are not the ends; they are the means—to make what hasn’t been made before, to say what hasn’t been said before, to show what hasn’t been seen.

And that’s only the beginning.