This morning I had an interesting 90-minute Zoom discussion about the use of artificial intelligence (AI) in producing works of art. I expect it to be uploaded soon to the Synth & Software website.
Music has always been a collaborative art. Even in the 18th century, few musicians built their own instruments. Orchestral and chamber musicians seldom play their own compositions; they’re collaborating with composers. In the pop music world today, you can buy and use beats (drum loops, with or without other instruments) and craft your own songs by collaborating with musicians you’ve never met and whose names you don’t even know.
In that sense, the use of software to generate musical ideas is not a radical departure. I could dissect some of the details, but I want to look at a different field.
Can AI software generate reasonable approximations of fiction, or even ideas that you can use as springboards for your own fiction? No, it can’t. Whether it will ever be able to is, I would say, very doubtful.
An AI can be fed ten thousand works of fiction, and can analyze the input searching for patterns. That’s what AI does: It searches for patterns. In the case of prose paragraphs, the patterns consist of groups of words. But an AI does not know what any of the words mean.
I did a search for AI services that generate prose, and shortly found myself on a site called sasswords. This site invites you to enter a “prompt” of 15 to 30 words, after which, at the touch of a button, the AI will generate a paragraph or two. Calling the output a story would be stretching the term until it twangs like a rubber band. It’s just word hash.
Let’s look at what sasswords wrote. The prompt I dreamed up was this: “The sun rose over the steaming jungle. The shrill cries of brilliantly colored birds assaulted our ears as we lay trembling in the mud.” I had no particular idea in mind for this, it just popped up out of my unconscious.
Here’s what the AI produced:
The sun rose over the steaming jungle. The shrill cries of brilliantly colored birds assaulted our ears as we lay trembling in the mud. The first creature that we saw stood on the banks of a stream, looking towards the sea as the sun set behind it, the birds flapping in its wake, singing to it in response to the rising sun. At the sight of its reflection, I began to laugh. In the first few seconds, its head turned towards us. Its eyes were two large, empty black pits. We were its eyes, we thought, then turned back to our task. Our eyes had been turned by an enormous creature who stood a kilometer away, a hundred times bigger than the gargantuan fish we were.
We had seen the creature before. A few days earlier, when it had taken over a planet, our very first encounter with it occurred when the planet had already been transformed. When it took over, an underground city had fallen to its claws. Before the day was out, all life disappeared. A single moment of human life had disappeared in that city. Now, in this hour, every living creature on our planet was gone. No, not gone, just gone.
We had come too late, he said. He smiled. My father had told him about my father, my grandfather, as they called him. What had my Grandfather done to him? My grandfather had not done anything. Not anything at all. None of them had done any of it to me! I realized that my laughter was not directed at him, or the others.
Wow! Here’s another AI-generated story opener, starting from the same prompt:
The sun rose over the steaming jungle. The shrill cries of brilliantly colored birds assaulted our ears as we lay trembling in the mud. It was late morning, the sun was going down. We had heard the sounds of crickets earlier, when we had first awoken, from the woods. So it seemed that we were not entirely alone and that the jungle was not so far away. Yet I knew that it still hadn’t dawned. And that, in fact, we would not see the rest of our time here, not for a very long time. Or until we died.
I was the first to wake. There was no sign of anyone else. In fact I don’t think anyone had ever moved in. But the morning dawns over some of your fellow humans here and then they are gone for good. No one here has a pulse. Some of them are still alive, by the looks of it, while others are just skin and bones, a little slower than most. One of my fellow humans, she was an African, had been dead for just about an hour. She had a pretty good head on her, which I suspected she would want to keep. With the one we called the man. He was about 50, I guessed, tall and broad, like me a bit, a short and stout figure in his own right. His eyes were closed, one hand held the other, his eyes closed as he slept, eyes open, hands on his chest. When he opened them and saw us he was smiling. Our faces were covered in mud, just like the others. Though I hadn’t realized how muddy they were, or how thick. By the look of him I assumed he’d been drinking at least. That’s how I came to find him. Another was just a slight figure. Perhaps a girl. Her face was still covered. A man was slumped over. Again, our faces covered, as well as the woman. Where we stood was the tip of a tree, looking down at a clearing in the middle of this. Behind us were an old man, an elderly man with a cane, holding an infant in each arm, who was asleep.
I claim that these raw outputs make the principle clear. The AI cannot produce meaningful fiction, because it does not know what words mean. The sun setting as it’s rising, a man who is tall but also short, a man with his eyes closed and also open. Beyond that, no story is being told. There is nothing even faintly resembling action or a coherent scene.
Perhaps the underlying question is this: How could a team of software engineers be so utterly clueless that they would think this tripe is worth being offered to the public? I can understand how the software is failing; what I don’t understand is how the people are failing.
In real fiction, or for that matter in almost any other type of writing, the meanings of sentences and paragraphs are not on the page at all; they’re assembled in our brains. And the wetware we use to assemble the meanings has access to enormous amounts of subconscious or semi-conscious knowledge about the real world. Nobody has to tell a human writer that a tall man can’t be short, or that people who don’t have a pulse cannot still be alive.
Music is entirely abstract, so in some sense an AI can produce an acceptable analysis of musical patterns. Music software can quite easily generate ideas that we might never think of ourselves, and a skilled musician can then take those ideas, perhaps massage them, and build on them to produce a satisfying piece of music. But prose? No. I assert confidently that software cannot write a short story, and will never be able to.
Artificial intelligence is a one-trick pony. And if you look at the sentence I just wrote, look at it carefully, you’ll notice not only that software does not know what a pony is, nor a trick, nor what it means for there to be only one of something; beyond that, the software is not likely to have access to any of the cultural references implicit in the phrase “one-trick pony.” It might have the phrase available in a database, but it cannot know what the phrase means.
Some things, such as writing fiction, will remain the domain of humans for the foreseeable future. And that’s a good thing. Other observations could be made about AI-assisted creativity; I may return to this topic later, but a nice splash of cold water is a good place to start.
Jim Aikin's Oblong Blob 