Essay

Can computers be creative? A review on artificial creativity arguments Eric Kok – [email protected]

Introduction Often when thinking about artificial intelligence, we tend to directly think of the higher level notions of intelligence that humans seem to have. Concepts like consciousness, emotions, free will and reasoning frequently arise in the discussion on human-level intelligence. And thinking of this, I’d like to think of one other concept as being the most important of them to create intelligence – creativity. Creativity seems not only one of the most important aspects of human intelligence, it is also often cited as perhaps the hardest problem to solve of them all – if even possible. It has many connections to related problems (1), but even on its own creativity is a concept that is frequently. If we like to build artificial creativity some day, it is important to see if this is possible and how. In this essay I will review existing literature and explain the problems involved with creativity in computers. Based on this, I will form a conclusion that will answer our question ‘Can computers be creative?’. In this study into creativity, I will first look at the problem of defining what creativity is and what kind of creativity we are willing to create in computers. This will result in several questions concerning whether it can be done in principle. These are divided in technical and non-technical considerations. From all this information, I will draw a conclusion and explain how I come to this. Creativity relates to several other artificial intelligence topics. Questions like ‘Can computers think?’ or ‘Can computers have free will?’ share a lot of the arguments made in this essay. However, we will focus on the creativity of computers, it being hard enough to cover fully already.

The definition of creativity Before we can start our research into artificial creativity, we need know what it is that we want to build into our computers. Because we are looking to build this creativity and test whether we succeeded we need to have a definition of what this concept of creativity is. Ask two people what creativity is and they are very likely to give a whole different answer. Let us go though several concepts that we may come across when looking into artificial creativity. The origin of the artificial intelligence question is usually credited to Turing (3) in his 1950-paper introducing the Turing Test. Already here he talks about creativity and whether it can be done. From his point of view, creativity has to do with ‘surprise’ and ‘originality’. If a computer surprises us with something we didn’t expect it to do, we may say it is being creative. Since even random number generators are than creative, he extents his idea to originality, stating it should produce original work. He claims computers can come up with new facts.

Thinking about the words surprise and originality, it seems these are important for creativity indeed. Creativity certainly has to do with coming up with something that wasn’t there – something that wasn’t an existing fact. We could see a new research theory as such a creative outcome, being something that wasn’t there before. I will later look into whether computers can do this. For now, we first further analyse if these concepts ‘surprise’ and ‘originality’ are enough, and I don’t think so. Surprise can also be achieved by a random number generator or a failing computer. An original outcome in the sense that something has never been done before, is not what creativity is about. We could take a chess piece move as an example. We see making such a move, when being made by a chess grandmaster in particular, as a great creative process. However, the actual outcome (the action taking a piece from 1 place to another) is not something that has never been done before. If we take this to the general problem, we see that creativity can be specified in two ways: the process or the outcome. In the chess move, the creativity was in the thought process of the chess grandmaster. However, in an art painting for example, the outcome is what we think of as being creative. You will have to come up with the idea of making this painting. The process and creative outcome are often intermixed as we already see. This makes it important to each time view at both ideas of creativity. Creativity as an outcome is interesting because this is what we are actually being able to sense. It’s the painting that we see, not the thought process. Does the painter really have an underlying idea on why he made the painting this way? Does it have meaning and does it not when it was made without thinking about it? It will be really difficult to test this. Testing the outcome on creativity however is easier. The Turing Test (3) for intelligence can be extended here to test for human-like creativity. We could state that if the creative outcome (like a music composition or a research thesis) can just as well be from a human as a computer, it passes the Turing Test for creativity. So if a viewer cannot tell whether it was a computer who made it or a human, the computer possesses human-level creativity. Of course, it is debated whether passing the Turing Test implies being truly creative, since it doesn’t test the actual thought process. However, for now this is no problem as long as we state that at least the outcome is creative. Apart from the concept surprise and originality, we may look at more complex definitions of creativity. For example, the Merriam-Webster dictionary (4) states that creativity is ‘Having the quality of something created rather than imitated’. This is an interesting explanation because this implies it should not be just an imitation. You could say that by making artificial creativity, you implicitly imitate human-like creativity and thus humans. However, this is only true when you look at this from the implementation view, not from the behavioural view. Looking at behaviour, or the outcome, it is needed for the computer to create something that is not the result of something he did already know. Now, it is interesting to see if a computer can actually come up with such new ideas. Since computers only reasons on its symbols which are put in by the programmer, can it suffice to only derive new facts based on this knowledge? Just before we see if artificial creativity can be done in principle, let’s take a look at the degrees of creativity we can distinguish. Firstly, it seems that human-level creativity is the most elaborate form of creativity we know. Therefore, it is the level which we want to compare our computer to. We will establish if we have come so far, but keep in mind there are more levels of intelligence which are of interest. The other extreme is the random number generator, which may take us by surprise.

However, nobody considers this as being creative. So currently we are somewhere along this line between the random number generator and humans.

Can it be done in principle? Talking about creativity, with both people in the field of artificial intelligence and outside, can become pretty heated. Concepts like emotion or consciousness may arise and from there the discussion seems endless. Also, existing programming concepts are considered and refuted. Both these points of view deal with the suggestion that creativity should be possible. We would like to see if we, in principle, can build artificial creativity. Because if we find any shortcoming, we might leave the idea of creating creativity altogether. First we will look at the programming paradigms considered for building artificial intelligence. The most widely used approach is that of classical pre-programmed or deterministic systems. These programs are build based on predefined algorithms that deterministically operate on knowledge that the system has. From this knowledge the algorithm derives new facts which are previously nonexisting outcomes of the system. These new facts might be considered creative, if it suffices our definition of creativity. Do such systems have any shortcomings because determinism cannot create creativity in principle? If we look at humans, it seems we use this kind of reasoning ourselves a lot in certain areas. For example, in science we search for new unexplored facts in our research projects. Both the research process and the outcomes are considered to be creative. Therefore, this kind of reasoning, the process of combining existing knowledge to create new facts, may very well be sufficient for building creativity. The same holds for chess moves: we reason on all the possibilities we know of and choose one that seems best. A research project that uses existing knowledge is Cope’s Experiments in Musical Intelligence (14). It generates musical compositions based on pieces of music by Bach, Vivaldi and Cope himself. The software analyses these pieces and uses them, as well as generated variations, to create new work. These compositions, especially to the non-experienced listener, sound interesting and amazingly human-like. Often the argument is raised that deterministic systems can only create something that already exists. The computer can never take us by surprise, because it just knows what we put into it (3). Creativity involves new ideas which cannot emerge by simply following existing rules (5). However, this totally in the eye of the beholder. Surely, we could make all calculations that the computer has to make and predict the outcome. But then, if we would fully understand the human brain, it’s no longer creative anymore too. In contrast, it even seems reasonable to create something based on existing knowledge, since we can only think about something if we (or the computer) has a representation of it. Computers take us by surprise all the time, either wanted or unwanted (computer failure). Based on this argumentation, I believe that creative outcomes are just combinations of existing knowledge, without the need for randomness. Note that we might use randomness in deterministic software. Although it not being a necessity for creativity, it might help to improve on creative

outcome. We could even claim that this randomness is also just a computer algorithm though and so the same line of argumentation holds as for determinism. What we do need is variation. Like in evolution, variation is needed to make enough new useful combinations. If we the knowledge base is too small, we would get stuck because of this limitations. If we have enough variety, the computer will be perfectly capable of being creative. It can create something that is surprising, novel and before non-existing. Next to classical deterministic reasoning systems, we might use learning systems or connectionistlike models to build artificial creativity. Various techniques can be used including evolutionary algorithms (like genetic algorithms), reinforcement learning, genetic algorithms and function approximation (like neural networks). Such systems do not reason on their knowledge to logically come to new facts, but rather come to the optimal behaviour by trying things out and learning from this experience. Such systems are therefore black boxes (although the algorithms themselves can be explained of course). If some output of the system is given, you can’t explain why this was the optimal solution. There is no argumentation for the outcome, it is there because this turned out to be most beneficial. Learning can be used to build artificial creativity. It can for example be used to analyse knowledge of musical compositions and combine this to create a new composition. The outcome is likely to be something new – non-existing. If you use neural networks, you could even compare this with how the human brain works. The neurons in the net are connected by links with certain weights. These patterns can be compared to the synapses in the human brain. The neurons on itself are not creative and do not have any knowledge, but the systems as a whole can display creative behaviour. Neural networks and other learning and connectionist approaches therefore seem suited to program artificial creativity. However, I personally believe they are merely a very handy tool and nothing more than that. Like randomness, it is a very nice technique that can help us build better programs, but it is not compulsory to be able to build artificially creative programs. As we will see later, a property of the software might be that it can adjust itself. It changes its own functions or constraints it operates on. Several projects used this approach, such as Lenat’s Automatic Mathematician and EURISKO (12). Apart from the available programming paradigms, there are often arguments raised that go deeper than just the programming techniques. These include claims that artificial creativity cannot be build because computers lack emotions, consciousness, intentions or even a soul. At first, these arguments may seem valid because people associate creativity directly with such concepts. In effect, we need to address these objections and try to refute them if we want to build artificial creativity. Especially the process of making something creative is being contested. The most commonly raised argument is that computers cannot be creative because they lack emotions. The idea behind this is that without emotions, a certain creative outcome just has no meaning. Obviously this implies that meaning is needed to get a truly creative outcome. If a computers makes a musical composition by just analysing existing music it has no meaning. I do not agree with this. If the composition made is just as original and interesting as a human composition, why should the listener believe this is not a truly creative piece of work? It is the listener that decides

not only if the outcome is creative, but also the process of making it. This is because we can only judge the outcome on what we can sense. We can hear the music (or read the notes), but we give meaning to it ourselves. Maybe we think we know what the author meant, but this is still our interpretation. The same holds for other kinds of creativity. If we view a painting, we can only criticise the meaning of it based on what we see. The painter may explain the meaning, but we, or even the painter, cannot be sure this is what it truly means. It is always in the eye of the beholder to judge the genuine meaning. I see no reason why computers cannot use this line of argumentation. Emotion is not needed to build artificial creativity. Cohen’s AARON project is the best-know research project generating full colour paintings with its own style (15). Although Cohen himself does not claim the software to be creative, it is if we as a viewer of its paintings judge him to be. A different view is that consciousness is needed. Consciousness is the idea that the system should be aware of itself. Humans are considered to be conscious and arguably some animals are as well. The point is often raised that it is the consciousness that sets us apart from computers since computers have no subjective experiences, they are not aware of their place in the world. A computer cannot like red better than blue for example, or like anything at all. This difference between humans and computers makes that computers cannot be creative (16). Although this seems reasonable, it has a problem. It assumes consciousness is something a computer cannot be. It is a certain property that humans have and that cannot be build. However, if a computer simulates the conscious behaviour of a human, why should we judge the computer as not being conscious? This reflects the Turing test for creativity. If the computer exposes the conscious choices and actions resulting in a creative outcome such as a painting, the computer is itself conscious and creative. In the colour example, why can’t the computer prefer red over blue? As long as it’s consistent and honest about its feelings, it has the same consciousness about it as humans have (6). It is hard to see why this argument should not hold. Why is a human or chimpanzee conscious and not an ant? It is because we, as humans, judge the ant’s behaviour as not intelligent enough to be conscious. However, a super intelligent alien race may not even consider humans as conscious. It is not a property you have or not have, but an objective interpretation of the beholder. Apart from emotion or consciousness, intention is being addressed as the missing link to build artificial creativity. What is meant by intention here is the goal the author has for which it needs creativity. What does the creator intent to achieve? One of the main defenders of this theory is John Searle and it resembles his Chinese Room argument. (7) He designed an experiment in which a man sits in a room with books to translate English to Chinese. Through a hole in the closed room, he receives an English text, translates it using the books and returns it through the hole. This, Searle states, is what a computer does when you ask him to do a certain task. Now, he claims that the man in the room, although he translated these texts, can’t speak Chinese because he doesn’t truly understand it. This argument is now called the Chinese Room argument and can be extended to intentionality for creativity. Searle claims the whole point of the experiment is to show that the man

in the room, or a computer, does not have the intentionality it needs because it only operates on symbols and not on semantics. To see if this is a problem for building artificial creativity indeed, the fact must hold that genuine intentionality is needed in the process of creating. Now, who determines this? If a computer makes a piece of art, you might say it has no real argumentation or intention based for it, because it does not possess intentionality. That’s where I think the problem lies. For one thing, as with feelings, who says this is the true intention the creator based his work on? Does it make a difference if the work had been done by someone who didn’t care? Or by an elephant (8)? No, I believe that creativity, even the creative process itself, is in the beholder, not the creator. It is the person that is viewing the painting or listening to the music who determines the degree of creativity, if at all. Therefore, the fact that computers have no biological intentionality is no setback in making artificial creativity. Intentions can however be an interesting way to view intelligent or creative systems. As described by Dennett (9), there are different ways of describing systems such as by the physical or functional stance. He proposes the intentional stance in which the goals, or intentions, are used to explain the systems workings. This view is already used as a basis for the BDI software agent architecture (10) and can also be of use in building creativity. This approach might make it not only easier to oversee the program’s inner workings but might also make artificial creativity easier to accept because it doesn’t look like the classical GOFAI approach.

Conclusion Having considered several technical and fundamental arguments on why artificial creativity is impossible I will try to answer the question ‘Can computer be creative?’. Yes, I believe computers can be creative, even as creative as humans. We have not come that far yet, but there are no fundamental issues on why artificial human-level creativity should not be possible. This is because creativity is a non-random combination of existing knowledge. Since computers can reason on their beliefs and create new, previously non-existing facts, computers can be creative too. There are three counterarguments against this statement. The first being that biological consciousness or any other animal-specific property exists that is needed that computers can’t have. This is for example advocated by Searl (7), and is called ‘brain stuff’. It is the biological substance that can contain intelligence and thus creativity. This material, neuroprotein, makes animals, humans in particular, so special. It allows for the consciousness and intentionality to exist upon. The problem is that it is never been proven that such a material is needed to build intelligence on. In fact, any material will do as long as it is possible to create the functional characteristics upon it. Therefore, artificial creativity may very well be build on silicon-based computers. Secondly, some refute that all possible (creative) outcomes already exist in the exploration space. A creative process creates new facts based on its existing knowledge, thereby creating novel and surprising new combinations. These outcomes already existed, but were just not made until then. Sometimes it is argued that only humans can be creative because these possibilities just didn’t exist before. The facts were not yet there to be explored (and selected). However, this is false. Take for example a painter. All the options to make brush strokes in all the different colours are there, available to the computer. Boden (11) calls all these possible combinations the conceptual space. It is

in this conceptual space that we can explore to come up with new combinations of the existing known facts. All the options are already there. Sometimes however, to come up with something genuinely creative, never been done by any human (or computer), it is not enough to just explore this space. What it needs is to transform the conceptual space itself. For computers to do this, they need to change the way it works itself, like human introspection. This can be done to edit their own programs, or at least there constraints that defined the conceptual space. Although very few of such programs exist today, it is very well possible for the computer to do so. Now if we accept that artificial human-level creativity is possible, how do we make it and when will it be available? This, of course, is still impossible to answer. If we would know exactly how to build creativity into computers, we would have already done it. However, we might not be as far off as we might think. Several research projects have already shown very interesting results on a low and domain-specific level. I believe the main ingredients are already there: conceptual space exploration, randomness, learning and connectionist algorithms and self-adapting software. Near-future research will show increasingly higher degrees of creativity. If we adopt ideas like Kurzweil’s law of accelerating returns (13), it could be in just a few years that we see human-level creativity. He believes the increase in available technology, as well as computation power, will increase exponentially and we are in the ‘knee’ of the explosion. In other words, within years the increase in knowledge and computer power will explode. He calls this the Singularity and we will have all the available resources and knowledge to create intelligence. From that point, computers will be viewed at as being just as creative as we are. One final point to notice is that I don’t claim that computers, if they poses this human-level creativity, are biologically conscious or have genuine emotions or feelings. The point is that this isn’t needed for our computers to accept them as being like us. If we pass this creative Turing test, artificial creativity will be just as creative as human creativity. From that moment, we can expect computers to not only produce novel and interesting new music or paintings, but create new musical genres, artistic movements and scientific theories. And humans will have all this to marvel at and use for the future extension of human intelligence.

References (1) MacroVU Press (Bainbridge Island, Wash.), & Horn, R. E. (1998). Mapping great debates: can computers think? Issue mapping series. Bainbridge Island, WA: MacroVU Press. (2) Brooks, R., Kurzweil, R., & Gelernter, D. (2006). Creativity: The Mind, Machines, and Mathematics. http://mitworld.mit.edu/video/422/. (3) Turing, A. M. (1950). Computing machinery and intelligence. Mind : a Quarterly Review of Psychology and Philosophy. LIX(236), 433. (4) “creative.” Merriam-Webster Online Dictionary. 2007. http://www.merriam-webster.com. (5) Boden, M. A. (1998). Creativity and artificial intelligence. Artificial Intelligence. 103, 1-2 (Aug. 1998), 347-356. (6) Danto, A. C., & Morgenbesser, S. (1966). Philosophy of science; readings. Cleveland: Meridian Books. (7) Searle, J. R. (1980) Minds, brains, and programs. Behavioral and Brain Sciences. 3 (3): 417457.

(8) Novica. Elephant Art. http://www.novica.com/region/elephantartists.cfm. (9) Dennett, D. C. (1987). The intentional stance. Cambridge, Mass: MIT Press. (10)Wooldridge, M. J., & Jennings, N. R. (1995). Intelligent agents: Theory and practice. Knowledge Engineering Review 10(2). (11)Boden, M. A. (1991). The creative mind: myths & mechanisms. New York, N.Y.: Basic Books. (12)Lenat, D. B., & Brown, J. S. (1984). Why AM an EUISKO appear to work. Artificial Intelligence 23, 3 (Jul. 1984), 269-294. (13)Kurzweil, R. (2005). The singularity is near: when humans transcend biology. New York: Viking. (14)Cope, D. (2006). Computer Models of Musical Creativity. Cambridge, MA: MIT Press. (15)Cohen, H., Cohen, B., & Nii, P. (1984). The first artificial intelligence coloring book: art and computers. Los Altos, Calif: W. Kaufmann. (16)Nagel, T. (1974). What is it like to be a bat? Philosophical Review 83: 435 - 451.