Our Spring magazine is finally here! Click here to view and read our new articles!

Can Robots Be Creative?

The limitations of early A.I.

A problem that long bedevilled Artificial Intelligence (A.I.) Developers was how to program an A.I. which can learn something of its own accord. Whilst absent of this, they had no option but to laboriously code everything into an A.I. which it may conceivably need to know upon implementation into its environment. For example, the code for a chatbot could not have been developed by telling it to imbibe the vernacular and grammatical structure of a piece of online literature. Instead, it was down to the tedious efforts of a human developer, in enumerating the explicit rules that are assumed to underlie general conversation, such as, if the user says ‘Hello’, then reply with ‘Hello, how are you?’. As you can imagine, this approach entailed a piece of code thousands of lines long reading, if the user asks this, then reply with this. Whilst this was sufficient for programming simplistic, predictable chatbots, it could never have been applied to an issue containing such complexity as facial recognition or autonomous driving; there are too many possible scenarios these may encounter, a near infinite number, that it renders futile any attempt on behalf of the developer to explicitly encode them all.

The ‘game-changing’ potential of neural networks

Much to the delight of developers, algorithms inspired by the human brain, known as neural networks, later came to fruition, and obviated A.I. incapable of learning as a result. A chatbot could hence learn how to talk merely by reading online literature, and an A.I. could hence recognise a face merely by being fed a huge dataset of them. The potential that lurks within neural networks can thus hardly be overstated, a lot of which we have already realised through self-driving cars, voice recognition technology, and medical X-ray software to help diagnose cancers as either malignant or benign. But there exists to them a darker, more ominous side, that we have perhaps overlooked due to the giddiness that tends to stem from the promise of new, exciting technology. Namely, neural networks are able to produce jokes, poetry, screenplays, music, art and literature that is uncannily resemblant of human creativity. Do we, therefore, need to reconsider which abilities we can justly regard as exclusively ours and those we cannot? The ability to communicate, for example, is obviously not just ours: other animals can do it too. But since seldom do we see a cow compose a compelling sonata, or a chimp write a poem of lachrymose profundity, it appears as though we may regard creativity – the manifestation of what is often reportedly felt to be an unconscious, higher impetus to produce something unique, valuable and surprising – to be one of our highest active faculties and something which only we can enjoy the privilege of being endowed with. But I believe that we may no longer be alone in enjoying that privilege; I believe that A.I. could be just as if not more creative than humans! And to illustrate what I deem the most compelling example of this, it is necessary to provide an account of the world’s oldest game still played today, Go.

The computer that defeated a world champion at his own game

Go is played by two people on a 19×19 square board. Each player assumes the colour of either black or white, and takes it in turn to place a piece on the board, with the overall intention being to build as much territory as they can – territory being the sum total of the area inside the shapes produced by counters on the board. Simple as these rules may appear, they in fact lend themselves to something that is more often regarded as a form of art, rather than a game, not least because there are more possible board positions, 10170, than there are atoms in the universe! Furthermore, in contrast to a more systematic game like Chess, Go does not lend itself to being encapsulated by an explicit rubric; it is an intuitive game, where the best players, after making a certain brilliant move, will often remark that the rationale for doing so was that ‘it just felt right’. It was thus supposed to be intractable, if not impossible, to ever build an A.I. that could play Go, with enough formidability as to compete with even the average player. In spite of this, Deepmind, a software development company based in London, resolutely embarked upon a mission to do just that. In 2017, having utilised the newfound power of neural networks, they created an A.I. called AlphaGo, and invited Lee Sedol, the then best player of the last decade, to compete in a best-of-five tournament to determine who, or what, was better: human or human creation. Over 200 million people tuned in live to watch as AlphaGo, contrary to all expectations save those of Deepmind, won in dominant and portentous 4-1 fashion! Although the A.I. researchers were rightly rejoiced, AlphaGo’s victory evoked what was a rather solemn and sobering epiphany from others: mastering the game of Go requires profound creativity and intuition, which were hitherto supposed to be exclusively human capabilities. It should have been impossible therefore for these to be displayed by something that wasn’t human – by a computer. So the fact that it now was, meant that maybe we aren’t so special after all!

It wasn’t just the fact of AlphaGo’s win that was monumental, however, it was also the sheer brilliance with which it played. Namely, on move 37 of game two, it played a move that was so unconventional it was dismissed as a mistake by professional commentators at the time. It transpired, however, no sooner than 100 moves after, that AlphaGo had not only played a move that proved pivotal in winning the game, but had also insouciantly flouted one Go’s cardinal maxims: do not place a piece past the third or fourth line of the board inside the opening of the game. AlphaGo played on the fifth. Therefore, if AlphaGo hadn’t won in such flamboyant fashion, and had instead executed some prosaic sequence of moves over and over to secure its win, then it would have been much easier for people to impugn, or dispense with, the significance of its win. However, not only did this not occur, but AlphaGo has since influenced how the game is played by professionals all around the world: it is now considered completely viable to play past the third or fourth line inside the opening of the game. AlphaGo has thus reminded us that there must always exist kernels of truth to which we are oblivious, and that our conventions, established over centuries and providing with them the appearance of reliability, are indeed the very things which paradoxically impede us from their discovery.

I believe then that if Lee Sedol had instead played such a move, a move even half as seminal, we would deservedly recognise him as being not just deeply creative, but a genius. Why then do the goalposts shift so zealously when it comes to deliberating whether or not AlphaGo can be creative, and therefore, whether or not A.I. itself can be creative? Are we so hubristic as to claim that there is something so inimitable about creativity, that it is only us humans who can be in its possession? Is not the only germane difference between AlphaGo and Lee Sedol the fact that one is human and the other is not? The fact of our reluctance is quite trivial, however, when considered in light of the following two propositions. The first being that such technology is still inchoate, meaning that we are just beginning to realise the low hanging fruit of A.I., yet it can already beat us in Go, diagnose breast cancer as either malignant or benign, recognise voices and faces and is rapidly learning how to drive cars. Secondly, A.I. will certainly not cease to improve any time soon: for barely do we get the chance to enjoy the convenience and advantage afforded to us by one product, before it soon demands swift replacement by a new and even better one! Thus, even if you believe that the victory of AlphaGo is trivial, or that the art hitherto produced by A.I. in no way competes with that produced by humans, you must surely hold some trepidation for how much better A.I. will likely get, and what the implications of this may be – the most expensive piece of art generated by an A.I., for example, sold for a staggering $432,500 in 2019.

What makes us human?

I have endeavoured to illustrate why we should no longer assume ourselves to be the only beings capable of creativity, and how the fact of this should logically cast doubt on whether the rest of our highest active faculties – consciousness, reason, rationality and empathy – are as insuperable to the advancement of technology as we might like to flatter ourselves in thinking. To leave on a rather ominous, but nevertheless relevant, question: will there be any room left for us in the future as it becomes increasingly filled with A.I.? Only time will tell…

By Kyle Mace