Art is how minds hold hands

Why I'm Not Worried About Chat-GPT

I attended two writer’s conferences last year. Generative AI was a frequent topic of conversation. Opinion seemed to fall into two camps. One was a sort of defiant assertion that machines can’t make art, generally delivered with a kind of haunted look in the eyes such as one might see on the face of a child trying to persuade themselves that there are no monsters under the bed. The other was the blithely confident assertion that all the problems with AI art, such as “hallucinations” where the machine makes up its own facts like a mechanical Donald Trump, would all be solved, probably within weeks.

I’ve written about this before. For me, the issue is not whether generative AI can ever be as good as a human, though it probably can’t. The issue is that however good it gets, it can’t be human. Patrick R of Banana Peel Pirouette sees things differently. It is a case worth answering. He begins with the familiar assumption that AI’s dominion over the arts is just a matter of time:

Today, AI-generated tunes are on their way towards obsolescing the human composer who tells a program which notes to strike and specifies the texture and timbre of the synthetic sounds, and living singers are no longer indispensable to recordings with vocal and lyrical components. Music is just music, however it’s composed.

The first thing we should note here is the phrase, “on their way.” It is a feature of AI that its real abilities are always “on their way.” They are never actually quite here yet. This has been the case for decades. We had the promise of spaceships and talking robots before we had color TV. Today we have color TV and spaceships, but we still don’t have the talking robots. Somehow we keep believing that AI is finally just one more year away. After all, we all know that technology is advancing faster and faster all the time, right?

Actually, no. If one looks at the pace of technological change over long periods, the pace has definitely accelerated over the last couple of centuries. But look closer and the picture is not one of uniformly accelerating growth in all areas. Indeed, we have been in something of a lull of late. Think back to the last truly game-changing technology to come along. That was the smartphone, a technology now 25 years old that has seen only incremental improvements of ever-diminishing significance over the last decade. If your phone is five years old today, you are not missing much, except battery life and OS updates.

This is how technology works. The pace of development slows over time. The car, for instance, has seen only incremental refinement over my entire lifetime. Its form and function are identical to what they were in 1957. Cars may be quieter, safer, more comfortable, and more efficient, but they still get you around in more or less the same amount of time they did in 1957.

Every technology has a fundamental underlying model. The development of a new model sparks a leap forward, usually followed by a rapid period of development which gradually slows over time as the model is more fully understood and exploited. There were all kinds of wild and varied car designs in the early years. But things quickly settled down to the basic form and function we have had since the 40s. Function and form converge around the refined model and products begin to be very similar in form and function, as they quickly became in smartphones, for instance.

Every model has its limits. The expectation of continuing rapid development falls apart once the model becomes mature and approaches those limits. The question with generative AI generally is, what are the limits of the model, and how close are we to reaching them? For instance, is its tendency to hallucinate an early glitch, soon to be smoothed out, or is it a baked-in limitation of the model? Well, according to the president of Open AI, the company that created Chat-GPT, the model is already played out. So don’t assume that there is more rapid advancement to come. Worse than that, though, there is increasing concern that the model has not only reached its limits but is actually collapsing, getting worse, not better.

Nor are the models living up to all the hype. A recent experiment pitted Chat-GPT in a Turing test against ELIZA, a chatbot from the 60s to see which could fool the most people into thinking it was human. ELIZA won.

And as Eric Hoel has recently pointed out, the industries that AI is supposedly poised to disrupt are collectively worth considerably less than the current investment in AI. Infinite development under those economic conditions is hardly guaranteed, even if there is still scope left in the model.

AI is essentially a programming technique. It is a way of attacking problems that are too complex for a human being to write an algorithm to solve. An AI is trained rather than programmed, which means that it creates its own code and organizes its own data, in ways that are practically impossible for human programmers to understand or tweak due to their size, complexity, and often inscrutable structure. So we should not think of problems like AI hallucination as something that programmers can quickly fix. Programmers did not write the code that produced the hallucination, and probably wouldn’t understand it if they could find it. It is entirely possible that hallucinations are inherent in the model.

I’m no computer scientist, and I don’t pretend to be able to tell you what is a hard problem and what is not, or whether the next model will crack the current hard problems or not. But I do know that developers turn to AI precisely to address the problems they can’t fix themselves, which means that fixing the problems with AIs is apt to be no small task.

But let’s grant for a moment that the model may not be exhausted and that generative AI may actually advance to the point where it is not painfully obvious that something is AI-generated. The issue then becomes, does it matter? Is music just music however it’s composed? Or is music essentially human, no matter how it’s composed? Banana Peel Republic again:

The skeptic’s complaint about the lack of a “human touch” (or something of that sort) in AI-generated culture product is a fundamentally confused statement. While we’ll always feel at least a tinge of disappointment upon learning that a poem or digital illustration that gives us pleasure was in fact produced by a machine instead of a person, what if we never found out? Our naïve enjoyment would remain inviolate; we’d paste the poem into Facebook, save the .png in our images folder, and blithely get on with our life.

Well, no, I don’t think we would. I think that if it ever did become difficult to know, we would insist ever more fervently on knowing.

What we’re actually bristling at is the idea that a machine can encroach onto what we feel is (or should be) our exclusive domain. It is unnerving to discover that a nonliving assembly of integrated circuits with an internet connection can be trained to turn out surreal fantasy imagery at a pace some orders of magnitude faster than a human artist, and at a level of quality very few people can ever hope to attain—but when the blinds obscure the methods by which a .png takes form, it doesn’t matter, does it?

That’s the question that matters. Not, does it matter to the artist, but does it matter to the consumer? And the answer is that it does matter. In all sorts of things, it matters to us to know that a thing was done by a human being. Take sports for example. Do you remember Robot Wars? It was a TV show a few years back about robots fighting other robots. No, not Transformers, actual robots. Its viewership and revenues now exceed all of boxing, MMA, and pro wrestling combined. Oh, wait, no they don’t. The show went off the air because, once the novelty wears off, no one cares about robots fighting robots.

No one wants to watch robots play other robots at soccer or golf or tennis either. I’m sure that robots would be able to beat humans at all those sports, and if they can’t do it now, I am sure Boston Dynamics is working on it. Robots can already beat humans at chess, but we are still more interested in human chess players than in watching robots play. Our fascination with golf (if we are fascinated with golf) is not with seeing a small ball placed in a small hole with weapons singularly ill-designed for the purpose,¹ it is in seeing a human being overcome these difficulties.

Indeed, all sports are designed to be difficult for humans precisely so we can watch humans overcome the difficulties. That they can be overcome by robots is not a matter of any interest to us. The difficulties were specifically designed for humans. And even if the robots could be made so human-like that we could not easily tell them apart from humans, we still would not want to watch them play sports. “Fembot wins Wimbledon” is not a headline anyone wants to read, at least, not more than once.

The genius of games, the thing that makes the few that stand out from the millions that are designed and promoted unsuccessfully every year, is to create a game or a sport that is easy enough for the average person to understand, but hard enough that no human being can entirely master it. Thus chess is the king of board games. You can write the rules on one piece of paper, yet people devote their lives to it and never entirely master it.

If this is so in sport, why would we imagine it was not also the case in art?

The human provenance of any given cultural artifact has until very recently been a sure assumption. Human-ness is not, and has never been, a quality inhering in a printed book, a vinyl record, an oil painting, etc.

Now I would dispute most vehemently that human-ness is not a quality inhering in a printed book, a vinyl record, or an oil painting. But that’s not really the argument here. The argument is that art is an artifact. One does not watch art-making like one watches sport. Being present in the Sistine Chapel while Michelangelo was painting the ceiling would not have been a great artistic experience. There would have been scaffolding and dropcloths everywere and the danger of getting paint in your eye if you looked upward. No, the art is in the artifact. And if the art is in the artifact, this argument goes, then how can it matter if the artifact was produced by a machine or a person?

We don’t get sentimental about the traffic cop obsolesced by the traffic light, the elevator operator put out of a job by push-button controls, the replacement of drugstore cashiers by automatic check-out lanes, or the 411 operators made obsolete by smartphones and Google. Why should we feel any different about creative workers outpaced and underpriced by machines? As long as it doesn’t perceptibly detract from either the quantity or the quality of the delicious, delicious content we crave, is there any reason that we should care?

This assumes that the value of the artifact consists entirely in how it affects our senses in the moment. But that has never been how we value art. The Mona Lisa and a cheap reproduction of the Mona Lisa both reflect photons in the same way. Yet if you were offered the possession of the original or the reproduction, you would choose the original every time. Perhaps you would do so only because you know you could sell it for more money. But why can you sell it for more money? Because human beings don’t value art merely for the photons it reflects. They value it for its history. Every piece of art is, in some sense, a relic, and the artist is the saint whose holiness imbues the relic with value beyond its mere function.

Even the cheap reproduction of the Mona Lisa is a kind of second-class relic. The reason for making the reproduction is the value we place on the original, which is in turn the value we place on the artist.

But why does the artist matter so much to us? I have long thought of art (including both literary art and art art) as one mind reaching out across the void and asking one poignant question: do you see it too? The appreciation of art is of the same character; it is the mind of the viewer seeking an expression that validates or enhances what they see in the world. Lovers can hold hands. Minds cannot touch each other in the same way. Art is how minds hold hands. By which analogy, AI art becomes a kind of sex doll. However lifelike you make it, there is no soul there, no acceptance, no affection, no loneliness reaching out to meet your loneliness.

Consider what art and literature actually are and how they work. Language is a bridge across the void between minds. Human existence is fundamentally lonely. We can never enter the mind of another, know what they know, think what they think, feel what they feel. And the loneliness of this is so terrible, so unbearable, that we use language to build a bridge between our experience and the experience of others.

By the same token, language only works because it is a bridge between the experiences of the people talking. I can understand what you are saying only because elements of my experience correspond to elements of your experience and we have developed words and stories that we can recognize and agree as things that point to the same experience.

AI advocates seem to imagine that the whole of consciousness, the whole of experience, the whole of intelligence, is captured in language. But this is obviously false. One does not have to look at humans alone to realize it. Look at the animal kingdom, where language is either elementary or entirely lacking, and ask if experience and consciousness are entirely absent. Clearly, they are not. Language is not experience. Language is not consciousness. Language is not intelligence. Those are properties of minds. Language is simply a bridge between one mind and another. Our ability to build bridges between minds is enormously important to our civilization, but, important as it is, language is still just a bridge.

AI is trained on language, not experience. It is trained on bridges, not on solid ground. No wonder it hallucinates. AI treats intelligence as something disembodied. But human intelligence is not disembodied. All our desires, all our pleasures, all our experiences, come from the body. Our brains exist to enable our bodies to function. Increasing evidence shows that our brains depend on our guts to think. We do sometimes speak as if the body were merely a possession of the mind, but the truth is that we do not “have” a body. We are a body. This is something that the most sincere atheist biologist and the most ardent Christian theologian would agree on: We are flesh. Human beings are creatures of flesh; our experiences are the experiences of fleshed creatures, and AI does not have flesh. Art is the recording and sharing of human experience, and AI does not have human experience.

AI does not have wants or fears, because it does not have a body. AI does not have motives because it does not have the needs that come from the body. Human intelligence is ineluctably tied to the body. A mind deprived of all access to the senses would be the most pitiable state of loneliness and deprivation. If AI actually were sentient, it would go mad, and it would be monstrously cruel to turn it on.

Human language, and human stories, make sense because they refer back to experiences that come from the flesh. An hallucination is an experience that does not come from the flesh. In this sense, AI hallucinations are aptly named since Large Language Models (LLMs), the form of AI model in question when we are talking about generative AI, have no physical experiences, only billions of words. In a very real sense, then, everything that Chat GPT produces is an hallucination. It is just that sometimes what it hallucinates happens to match the words used to describe a real thing. But it does not know real from unreal because it has no touchstone, no fleshed experience against which to test its hallucinations. The notion that it can be trained to tell the difference between the real and hallucination, therefore, seems dubious to me. It lacks experience, the very thing that we humans appeal to to tell hallucination from reality.

Chat-GPT4, the latest iteration of the system was trained using trillions of words. You and I were not. We could not read a trillion words in ten lifetimes, and we certainly could not remember them all. And yet we are better at telling truth from falsehood than Chat-GPT because we have experience, we have flesh. A four-year-old child, who may not have read a single book for themselves, nevertheless, through experience, has a far more sophisticated appreciation of the world than Chat-GPT and is becoming increasingly savvy at telling truth from falsehood. Chat-GPT does not know when it is being teased. A four-year-old does. A four-year-old understands body language, facial expression, and tone of voice. Chat-GPT hasn’t a clue.

Art is the expression of human experience. Chat-GPT can’t do art because it has no human experience to express. What it can do, up to a point, is pastiche. In fact, it would be a fair technical description to call an LLM a pastiche generator.

Being a pastiche generator is not inconsequential. There is actually quite a lot of market demand for pastiche. There is a huge literary demand for stories in the vein of some famous author or another who cannot write as fast as their readers can read or has other impediments, such as being dead. There are hundreds and perhaps thousands of authors who make a living writing pastiches of other authors to fill this void. Agents practically demand that new authors tell them up front which authors they are pastiching as this makes placing their books in the market easier. This kind of writing is not strictly art, though there may be an element of artfulness in it. But it is a valuable commercial commodity and a perfectly legitimate way to make a living.

Does generative AI pose a threat to such authors? Not yet it doesn’t. It isn’t nearly good enough yet. Will it get there? I’m inclined to doubt it. Commercial pastiche may not be art in the full sense, but there is still a substantial element of humanity and human experience and artfulness in its composition, at least in the better examples.

The point is not that if AI can consistently pass the most sophisticated Turing test then it will cease to matter whether you are speaking to a person or an AI. Quite the reverse, the harder it becomes to tell if you are talking to an AI or a person, the more we will demand proof that we are talking to a person. Because that is what we want. We want a bridge across the void of loneliness to another human being.

We don't want content, we want communication. Human existence is fundamentally lonely. We have access only to our own minds. Everyone else is a stranger to us. Except when we communicate. Knowing that there is a human being with human experience on the other side of the bridge is the only reason I am interested in the bridge at all.

The notion that AI can disrupt art by building shinier bridges faster, therefore, misses the point that we never wanted bridges in the first place. What we wanted was a way to cross to the other side. A bridge is of value only if there is something I want on the other shore. And the thing I want is a human being.

Something that pretends to be a human being but isn’t is an amusement when it first appears, a temporary distraction, as long as you can tell the difference. It would not satisfy your need to hold hands with another mind, but it may amuse you for a moment. But if it becomes something that you can’t tell from human, something that will hold your hand as if it were human, it becomes a horror. This is, after all, the plot of a thousand sci-fi stories: the horror of the golem, the imposter. Such an encounter will only leave you more anxious for the comforting touch of a genuinely human hand.

AI art is a bridge to nowhere, or to somewhere worse. A bridge to nowhere, no matter how bright and shiny it may be, and no matter how swiftly it may be built, is still a bridge to nowhere, and I'm not interested.

---

Speaking of building bridges, can an enchanted princess and an orphaned farm girl build a bridge that ends a reign of blood and redeems a lost soul? Give Lady Isabel and the Elf Knight a read, if you will. And if you enjoy it, please leave a review. That too builds bridges between readers and writers and helps make the robots our servants rather than our masters.

1. ‘Golf is a game whose aim is to hit a very small ball into an even smaller hole, with weapons singularly ill-designed for the purpose.’ – Winston Churchill