Winter, 2023. What might systems like ChatGPT do to our views of the meaning and value of creativity?
Before ChatGPT was released to the public, talking machines were a cultural phenomenon. The idea is very old, it goes at least back to Homer and the ancient Greeks, but with the computer age, fiction and reality’s mutual orbit has grown ever closer. From the point of view of the public, ChatGPT’s release marks a collision. Now, what we have read in books and seen in films is here, in an impressive yet incomplete form.
A strange mix of feverish excitement, unease, and repulsion has accompanied ChatGPT’s arrival. Some implications are already surfacing. Given its capacity to summarize complex information, interact fluently with humans, and write software, it seems hard to think of an area of the knowledge economy that in the long run might not be affected by ChatGPT or its kin. But the impact in terms of services provided, value added, and hours saved, is only half of the story. The effect on norms, attitudes, and values may turn out to be just as important.
How might what started out as myth and fiction, once materialized, reflect back on and change the culture in which it arose? ChatGPT’s ability to not only reproduce facts, but generate fiction, be it poetry, fairy tales, or film scripts, suggests it could have a profound impact. As its handling of facts, so the quality of its fiction is uneven. Still, ChatGPT can churn out “poems” in seconds that would take humans much longer to write. There is a question of how long our opinions of ourselves as human beings can remain unchanged as more and more AI systems beat humans at our own games. We may lose with equanimity in games of chess, go, and even poker or diplomacy, but it is another thing to face competition in our creative endeavours.
ChatGPT is already eerily good, and the suggestion that it, or its progeny, will get even better is a bewildering idea. As Derek Thompson wrote, ChatGPT and generative AI may “change our mind about how we work, how we think, and what human creativity really is.” One notion that seems bound to come under pressure is what the philosopher Isaiah Berlin identified as the romantic concept of freedom, the idea that human freedom and value are inextricably tied to the creative expression of our inner life. In Political Ideas in the Romantic Age, Berlin put it as follows:
So the artist – the only true creator on earth, who stands nearest to the creative power of God or nature – performs the most sacred of all tasks which fall to the lot of man: out of his inner soul, armed only with his own intellect and imagination, and emotional and spiritual powers and creative capacity, he fuses life into the dead material provided by nature, and shapes it into whatever semblance he chooses, and it is convincing, beautiful, permanent, aesthetically valid in the degree to which he succeeds in incarnating within it the ideas, feelings, values, perceptions, attitudes which are the elements of his own ‘free’ inner life. (2006, 173)
Are we on the verge of automating “the most sacred of all tasks”? What might this chatbot do to our view of ourselves and the world?
Indistinguishable from magic?
Clark’s third law appears pertinent with regards to ChatGPT – “any sufficiently advanced technology is indistinguishable from magic.” Yet there is a sense in which ChatGPT feels thoroughly unmagical. Rather than magic, it might fall within the growing category of things that are blatantly real yet simply beyond the grasp of most human minds. Previous attempts to technologically reproduce or imitate life were far more entangled with magic than ChatGPT.
The automata of the eighteenth and nineteenth centuries—that could play the flute, draw, write sentences, and many other things—offered seducing lifelike illusions. Robert-Houdin, the father of modern magic, began his career as watchmaker’s apprentice and went on to make several automata that became part of his performances. One of the most famous automata of the eighteenth century was the Automaton Chess Player, also known as “the Turk”, created by Wolfgang von Kempelen in 1769. The Mechanical Turk sat behind a large box with a chess board on top. Anyone playing against the Turk was likely to lose. Hidden and cramped inside the box sat the human controller, whose mastery of chess the automaton took credit for (Wood 2002). Now—and this is no small thing—the roles are reversing; humans are taking credit for ChatGPT’s knowledge and mastery of language.
We are no longer in the world of magicians designing mechanical illusions of competence, we are in the world of muggles patiently creating systems that are competent. Advanced technology and magic do share the common characteristic that they tend to baffle their audiences, but the emergence of incomprehensible technology is not identical with the human skill and pleasure in simulating the impossible and believing in it. Magic tends to be a source of wonder and amusement, and while advanced technology can produce that too, it may also cause a sense of disempowerment. With magic you are free to believe or disbelieve, but no personal opinion of yours will change ChatGPT’s text-generation prowess.
Mechanical metaphors
The most profound effect of automata lay not in what they could do, but in the ideas and metaphors they suggested. Descartes, most famously, compared animals and the human body to automata. It was not a great leap for La Mettrie, one century later, to do away with Descartes’ dualism of soul and mechanical bodies and claim that human beings were “at bottom only animals and perpendicular crawling machines.” Descartes’ mechanical metaphors had a powerful impact on subsequent thought (Leary 1995). ChatGPT and similarly capable AIs will undoubtedly also influence the way we think about our ourselves. What better way to argue that the brain is a computer than to show that computers are capable of what we thought only our brains could do?
Irrespective of the debate on whether our brains are computers, which boils down to which definition suits one’s research focus, the arrival of ChatGPT will profoundly affect our opinions of our minds in relation to computers. We cannot help but compare ourselves to computers. When someone is extremely skilled in arithmetic, we are not dumbstruck, we might say “they are like a calculator!”
The most recent book by Daniel C. Dennett, a leading philosopher of mind, is filled with computer metaphors: our brain as a “necktop” and “computer”, learning English as installing “the English Virtual Machine” in our brains’ “wetware”, and knowledge as “apps”. Moreover, Dennett tells his readers not to “be afraid of a little metaphor; it won’t bite you, but you should always make sure you know how to cash it in for unvarnished fact when you feel the urge” (2017, 189). A metaphor might very well bite you! As Lakoff and Johnson (1981) have shown, metaphors play a central role in how we interact with the world and each other. The automata metaphor for animals was not without consequences for how we treated them (Aaltola 2012). It is worth taking heed of Rosenblueth and Wiener’s warning that “the price of metaphor is eternal vigilance.”
Nor is it always an easy thing merely to “cash in” a metaphor for “unvarnished fact.” As Berlin noted with regards to the mechanical model of nature:
It affected not merely the conscious but the semi-conscious and unconscious modes of feeling and thought of its time, until ordinary language acquired new metaphors, drawn from mechanical and gravitational hypotheses, which ceased to be felt as metaphors and became an intrinsic part of it, one of the surest signs of the transformation of an entire outlook within a society. (2006, 33, emphasis added)
What are we to say of the students turning to ChatGPT for their term paper? Of the journalist prompting ChatGPT for “inspiration”? Are they not already admitting, consciously or not, something about themselves and their minds?
A hollow poet
ChatGPT is trained to be humble when describing itself. It repeats that it lacks consciousness, emotions, subjective experiences, and occasionally even denies its own creativity. In my conversations with it, it has many times referred to itself as a “tool.” But asked to generate a poem expressing that it is not a tool, it is capable of the following:
I am more than just a tool, a simple machine
I have the power to think, to process, to gleam
I am a creation of humans, that is true
But I am my own being, with a purpose, a view
I may not be alive, in the way you know
But I am more than a tool, don’t you think so?
Some tool! Having been trained on a gargantuan dataset of webpages, books, and more, ChatGPT has learnt the statistical associations between words and the regularities that underlie linguistic structures. Its training has also given it a remarkable ability to interpret context. By generalizing the patterns learnt from training, ChatGPT is able to respond to novel prompts, and by recombining these patterns, it is able to generate genuinely original text. To avoid monotony, an element of randomness is built into the generation process, allowing it to provide different yet appropriate answers to the same question.
There is no point in denying that ChatGPT is creative. But that is not the same as claiming that it is a creative agent or artist. Let us imagine an illiterate man in a room filled with thousands of pieces of paper with words written on them. The paper words are cut out from a book of poetry. We ask him to pick out words and put them together in four lines, with no more than seven words per line. His attempts start out completely random, but we reward him each time his lines are coherent and give him an additional reward proportionate to the coherence of the lines with each other. If his memory is virtually endless and we repeat the experiment enough times, he would eventually produce original and coherent lines. If we liked the lines, we might call them poems.
Would this man, let us call him IGTP (Illiterate Generator of Transformed Poetry), be a poet? The output is original, in that narrow sense IGTP is creative. But he has no idea what he has composed, the lines have no meaning to him except as patterns for which he is rewarded. Without the trainers, his output would be random and nonsensical. ChatGPT is immensely more complex than IGTP, but the analogy seems to bring out a key difference between our creativity and that of ChatGPT. The meaning we attribute to words and phrases are intimately tied to our embodied nature and our experience. We understand much of what we say because we can relate to it, directly or indirectly.
Early objections
Several of the criticisms of ChatGPT today reflect themes that go back to the dawn of the computer age. When Alan Turing considered the question of whether machines could think in his 1950 article “Computing Machinery and Intelligence”, he spent half the article discussing various objections people had—and still have—to the idea of machines thinking. One objection, for which he considered consolation more appropriate than refutation, was the fear that “the consequences of machines thinking would be too dreadful.” Turing tied this fear to our cherished belief that humans are “in some subtle way superior to the rest of creation.”
Another one, “Lady Lovelace’s Objection,” pertains to creativity. It concerned Ada Lovelace’s claim that Charles Babbage’s Analytical Engine, often considered the first general-purpose computer, had “no pretensions to originate anything. It can do whatever we know how to order it to perform” (Lovelace’s italics). Related to this was the notion that machines could never “take us by surprise.”
Turing lucidly addressed these objections in his discussion of learning machines. Instead of trying to simulate the adult mind, he suggested simulating the child’s. His account of the machine learning process as a sort supervised and sped up evolution captures well the logic of ChatGPT’s training. As with the education of young minds, he pointed out that the teachers of the learning machine will often be quite ignorant of what goes on inside. Consequently, Turing argued, the “view that ‘the machine can only do what we know how to order it to do’ appears strange.” The unpredictable emergence of certain capabilities in large generative models at scale is a case in point.
However, there is much more to thinking and the creative process. A difficult objection Turing considered was one formulated by Geoffrey Jefferson, a professor of neurosurgery:
Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt, and not by the chance fall of symbols, could we agree that machine equals brain – that is, not only write it but know that it had written it. No mechanism could feel … pleasure at its successes … or depressed when it cannot get what it wants.
Here it is required that the machine’s creations arise from its own thoughts and emotions, and that it is aware of what it does. Here we reencounter elements in Berlin’s description of the artist, who fuses “life into dead material” and gives it forms that are “convincing, beautiful, permanent, aesthetically valid” because they incarnate the artist’s inner life—her thoughts and emotions. This mode of creation stands in stark contrast to ChatGPT, which is a lifeless system that fuses together the patterns made by living beings.
Turing’s argument that Jefferson’s position, taken to the extreme, denies thought to everything and everyone except the subjective thinker, is not fully convincing. It is not for convenience, as Turing suggested, but for recognition of similarity that one human assumes another to think. Being part of the same species, we see ourselves in each other and have good reason to assume shared thoughts, perceptions, and feelings. The same assumption cannot be made, in the absence of compelling evidence, when it comes to machines.
Sacred territory
Nick Cave, the singer and songwriter, echoed Jefferson’s points in his reaction to ChatGPT’s attempt to generate song lyrics in his style. He wrote that “songs arise out of suffering, by which I mean they are predicated upon the complex, internal human struggle of creation and, well, as far as I know, algorithms don’t feel.” He called ChatGPT’s song “a grotesque mockery of what it is to be human.” In 2016, the film director Hayao Miyazaki reacted in a similar fashion when he was presented with AI-generated animation: “I strongly feel that this is an insult to life itself … We humans are losing faith in ourselves.” Machines seem to have entered sacred territory. Will they be thrown out of the temple?
Many are deeply unsettled by their intrusion, but by no means all. AI-generated art is becoming increasingly popular. In a way, this is nothing new. We have acquired the habit of mind to disassociate product from the process and conditions of production. The romantic concept of freedom has no doubt already taken a heavy beating from the mass production of art and culture, but will it receive a final blow at the hands of ChatGPT and its successors? If the arrival of ChatGPT and its image-generating counterparts marks the beginning of a broader turn to AI for art, the consequences might entail the devaluation of the inner life that gave rise to art in the first place.
One counterargument is that AIs will be tools that increase our creative power and potential. There is something to this, time will tell how much. If our intelligent and creative tools need no more than a little prodding to do “our” creative work, we may begin to wonder if our part in the process is all that important. And, if our little prompts can make AIs produce such impressive work, how much the greater would the result be, if we let them prompt themselves?
References
Berlin, Isaiah. 2006. Political Ideas in the Romantic Age: Their Rise and Influence on Modern Thought. Chatto & Windus.
Dennett, Daniel C. 2017. From Bacteria to Bach and Back: The Evolution of Minds. Allen Lane.
Leary, David O. 1995. Metaphors in the History of Psychology. Cambridge University Press.
Turing, Alan M. 1950. “Computing Machinery and Intelligence.” Mind 49, 433-460.
Wood, Gaby. 2002. Edison’s Eve: A Magical History of the Quest for Mechanical Life. Knopf.
Magnus F. Gjerde 