AI may be able to perform some routine tasks traditionally done by creative professionals, but could AI possess genuine creativity? Two of the great thinkers of AI disagreed on this very question. Ada Lovelace was an English mathematician and writer of the nineteenth century who was one of the first to recognize the importance of Charles Babbage’s machine, the Analytical Engine, often considered the first computer. She believed the machine capable of a great many things, but ultimately incapable of independent learning.
Article continues after advertisement
Because computers merely executed commands, she considered that any creativity should be attributed to programmers: “The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform.” For Lovelace, the real limitation of machines was their inability to come up with something new outside of their programming. They merely executed tasks designed by their programmers.
Alan Turing, who many consider the father of modern artificial intelligence, considered that computers could in fact surprise us, particularly where the consequences of certain states of affairs were not immediately recognizable to humans. Turing believed that a machine could be said to exhibit intelligent behavior if it could pass what he called “the imitation game,” which is now better known as the Turing test. In a paper published in 1950, Turing proposed that if a human evaluator had a text-only conversation with a machine and could not reliably distinguish it from a real human then the machine would pass the test. The purpose of the Turing test was to develop an alternative to the more ambiguous and potentially unanswerable question of “can computers think?”
The problem with the alternative formulation is that passing the test turns on whether a programmer can trick a test subject. The emphasis is on the process of deception rather than the capacity of the machine or the value of its outputs. For example, a chatbot called Eugene Goostman developed in Saint Petersburg in 2001 by three programmers is often considered the first to have passed the Turing test. Goostman was introduced to participants of the study as a thirteen-year-old Ukrainian boy, a background intended to excuse his poor language skills and lack of general knowledge. So when the chatbot responded to questions with strange non sequitur and nonsense answers, participants may have been encouraged to overlook these, potentially undermining the validity of the results. The imitation game simply encourages tricksters.
Creativity may appear to be a spontaneous activity, but the actual practice of creation often requires long stretches of artists honing their skills.
In the early 2000s, Professor of Computer Science Selmer Bringsjord and his team saw this problem and attempted to develop an alternative test. They wanted to verify whether AI could exercise human-like creativity, developing a new formulation which they named the Lovelace test. The researchers proposed that AI could be said to be creative if its programmers could not account for how it produced an output. The point was to insist on a certain epistemic relationship between a human programmer and a system that produced outputs. In other words, did the programmer know how the machine did it?
And yet this alternative also proves unable to test whether a computer exercises genuine creativity. There are two main issues. The first is that computers may be able to produce outputs that their programmers cannot account for but which might still be entirely lacking in worth or value. To focus exclusively on the epistemic relationship is to ignore the quality of what is being produced. For a computer to be counted as creative or intelligent its creative outputs must have artistic value for people and add something new to the world. Programmers of the current generation of chatbots already cannot account for certain outputs, particularly when they are led down hallucinatory paths.
But we wouldn’t count unintelligible outputs as instances of creative flair. The second problem cuts in the other direction: the test is too restrictive. Given enough time, developers of a program should be able to offer an account of the underlying reasons why it has been able to perform certain actions. The real question is not whether we can understand it, but whether it has produced something new that could be described as a valuable and creative work.
A better test, and one we hope still captures the original spirit of the Lovelace test, is what we call the Creativity Test. An artificial agent or system could be said to exercise genuine creativity if it can produce an output that originates something new that is judged as valuable by human observers. There are two elements to this test. The first is Lovelace’s originating principle. Even if elements of a computer’s output are present in its training data, its product must contain an original element that can be considered a novel contribution provided by the process undertaken by the computer.
Secondly, this creative act must have value as an artistic creation. This is not to say that it must fetch a high price on the art market or that it must command universal adoration from critics, but there must be some sense in which its outputs are considered worthwhile by a human community. On both elements, there is a degree of subjectivity involved in any assessment, but we see this as unavoidable for any creativity test. Ultimately, the subjectivity of the test makes it difficult to use as a measurement tool for scoring specific creative systems. It is rather intended as a thought experiment to help us understand the important elements involved in whether an AI system could be said to exercise creativity.
To determine how close AI currently comes to passing the Creativity Test, let us first consider how AI outputs could be compared to human creativity. The field of computational creativity is focused on modeling and understanding creativity using computers. One of its objectives is to see if a computer could be capable of achieving human-level creativity. It is widely considered that humans possess a spontaneous capacity for innovation. We can come up with new ideas, insights and creative ways of understanding and representing our world. This creativity relies on a combination of bursts of insight and long periods of hard work developing our skills, often in conversation with others.
On the one hand, artists commonly describe some of their best work coming to them as a flash of brilliance. Many ancient philosophers used to attribute artistic greatness to divine inspiration. In Plato’s early dialogue Ion, for example, he presents poetry as a result of divine madness, a form of truth revealed through the poet as a prophet of the gods. One version or another of this view still holds sway in modern times. Paul McCartney described waking up one day with a song playing in his head which he assumed must have been written by somebody else. After asking others if they had ever heard it before he stated, “eventually it became like handing something in to the police. I thought if no one claimed it after a few weeks then I could have it.” German polymath Johann Wolfgang von Goethe contemplated how to write The Sorrows of Young Werther for two years until suddenly it came to him: “at that instant, the plan of Werther was found; the whole shot together from all directions, and became a solid mass, as the water in a vase, which is just at the freezing point, is changed by the slightest concussion into ice.” Many of us could recount such experiences where ideas simply appeared to us.
On the other hand, such bursts of creativity could not be translated into great works of art without a long process of training and development. It is unlikely that had McCartney not received music lessons, played multiple instruments from an early age, and been among other creative minds, he would have written such transformative music. Many artistic breakthroughs that fundamentally change how a discipline operates are based on a thorough understanding and familiarity with a corpus of existing works. Creativity may appear to be a spontaneous activity, but the actual practice of creation often requires long stretches of artists honing their skills and messy periods of trial and error.
It is also important to note that there are many aspects of human thinking and creativity that could be described as algorithmic. The process through which we learn new skills is often rule-following and based on repetition and reinforcement. Many ideas that one might want to describe as innovation are often forms of imitation with subtle differences that distinguish them from that which has been copied.
While it is debatable whether AI has truly originated a new idea, it is clear that it can do more than that which Ada Lovelace thought possible. Deep artificial neural networks enable computers to generate far more surprising outputs from their input data than simple processes of command and execution. Contemporary approaches to machine learning allow computers to mimic the process of a child learning new patterns. One type of computer program that has surprised its creators is the modern chess engine, particularly those that use neural networks and reinforcement learning to master the game. An engine called AlphaZero, developed by Google DeepMind, became one of the strongest programs through a novel technique: rather than learning from previous games of grandmasters, AlphaZero developers only taught the program the rules of the game and nothing else, allowing it to play itself millions of times with the goal of achieving a better position until it attained superhuman performance.
As a result of this style of reinforcement learning, the engine discovered new moves that had never been played by humans, appearing completely counterintuitive based on how a human would play the game. While primarily based on logical calculations, chess has an aesthetic quality in which discovering certain moves relies on a strong imaginative capacity. Russian grandmaster Mikhail Botvinnik thought “chess is the art which expresses the science of logic.” In this game and others, AI has proved itself capable of generating beautiful moves that surprised its creators and masters of the game. Some would argue that on our Creativity Test, a chess engine could be said to have passed. These engines have produced new ways of understanding the game considered valuable contributions at the highest level of play. Yet the objection could still be made: an innovative move in a rulebound game is one thing, but what about a genuinely spontaneous and creative idea?
We see glimpses of creativity in certain AI outputs and understand the reasoning of people who want to see these as examples of genuine creative expression.
One thing is certain: AI does not create ex nihilo. AI’s creations are based on its training data, from which it discovers patterns and produces outputs that resemble the data. In 2018, Christie’s auction house announced its intention to sell the first piece of AI-generated art at auction. Edmond de Belamy is a blurry portrait of a man based on a training set of 15,000 portraits from the fourteenth to the nineteenth centuries. It was produced by French art collective Obvious, based on a type of image generation called generative adversarial networks (GANs), invented by Ian Goodfellow in 2014. While initially estimated at less than $10,000, the painting sold for $432,500 and achieved worldwide media attention for AI’s ability to generate unique artworks practically indistinguishable from human art. GANs work by using two neural networks to generate and then judge the authenticity of an image until the AI can create convincing copies of its dataset. Although everything is ultimately based on its training data, these image generators can combine elements in new ways to produce strikingly original pictures.
The sceptic might still see copying and creating images—even if there is a degree of novelty—as fundamentally a reproductive act and not quite sufficient to pass the Creativity Test. Yes, the painting is new, but ultimately it is simply derivative of the other works in the dataset. But what about a novel? The creator of the Lovelace test, Selmer Bringsjord, stated that if an AI could write a novel that captured his attention and that he found compelling, this would satisfy his criteria. In this field, AI still has some way to go. In 2016, The Day a Computer Wrote a Novel was celebrated as the first AI-generated novel and even passed the first stage of a literary award. However, the team of developers behind the novel imputed a large amount of the guidelines for the program themselves, including the plot line, characters and even key sentences and phrases, amounting to roughly 80 per cent of the novel, according to one of the developers.
Left to its own devices, AI struggles to produce a coherent piece of work. Another AI-generated project, the novel 1 The Road (2018) sought to emulate Jack Kerouac’s On the Road through an American cross-country road trip. Writer and engineer Ross Goodwin drove from New York to New Orleans in a car equipped with a camera, microphone, GPS and a portable AI writing machine in an attempt to reproduce the experience of a road trip with the AI writing as they travelled. Here is the opening line of the novel: “It was nine seventeen in the morning, and the house was heavy.” The results of this exercise were lackluster, with most of the prose appearing nonsensical. However, these obvious limitations of AI for writing novels have not prevented people from using it to produce large quantities of AI-generated books to sell on Amazon, including books written under the names of famous authors without their consent.
Whether or not the Creativity Test could be said to have been passed depends on a domain-specific assessment. There is a strong case, for example, that chess engines have created truly original and valuable moves—if one considers these creative acts. Image generators have also produced outputs with striking quality and originality. AI writing tools, on the other hand, are still much further from passing the test due to the low quality of their outputs and the amount of work that human programmers must inject into the process.
Ultimately, however, many studios in the creative industries are not concerned about whether AI art is genuinely creative or not. If it makes money and passes for a marketable product then it will likely be widely used. Art, too, is a commodity in a capitalist marketplace, and must be viewed from this perspective to understand how studios are likely to respond to new technologies. For the authors’ part, we see glimpses of creativity in certain AI outputs and understand the reasoning of people who want to see these as examples of genuine creative expression. At the same time, we argue there are hard limits to what AI can produce when it comes to genuine works of art.
__________________________________
Excerpted from Feeding the Machine: The Hidden Human Labor Powering A.I. by Mark Graham, Callum Cant and James Muldoon. Copyright © 2024. Available from Bloomsbury Publishing.